Datasets:

rntc

/

pubmed-preprocessed-quality

like 0

text

predicted_class

text

predicted_class

text

predicted_class

The original disease list of neurodegenerative disease from MeSH database contains 55 subdivisions. Several diseases are counted more than once since they are in different classifications, such as, ‘Gerstmann-Straussler-Scheinker Disease’ existed in both ‘Heredodegenerative Disorders, Nervous System’ and ‘Prion Diseases’ groups; ‘Shy-Drager Syndrome’ were classified to ‘Multiple System Atrophy’ and ‘Shy-Drager Syndrome’ groups, etc. In addition, no variation record reported for nine diseases in PubMed yet, the nice diseases are Lambert-Eaton myasthenic syndrome, limbic encephalitis, myelitis, transverse, opsoclonus-myoclonus syndrome, paraneoplastic cerebellar degeneration, paraneoplastic polyneuropathy, postpoliomyelitis syndrome, subacute combined degeneration and diffuse neurofibrillary tangles with calcification. There are some data associated with Huntington disease, myotonic dystrophy and Olivopontocerebellar Atrophies, but these data are not suitable for LOVD 3.0, we therefore have a final list of 37 diseases (Table 2). Up to now 1942 PubMed citations were manually screened, checked and 6374 related DNA variations for 289 genes were extracted and stored in our database. The GO analysis of these genes was done and the result is shown in Supplementary Material. Table 2.Neurodegenerative disease associated genes and variation collected in LOVD 3.0No.Disease nameAssociated genesNo. of variationsNo. of references1Alzheimer diseaseABCA7, ABCB1, ADRA1A, AGBL3, ANKS1B, APOE, APP, ATP8B3, BCL3, BIN1, C16orf96, C1orf112, C3orf20, CASS4, CD2AP, CD33, CELF1, CELF2, CENPJ, CFAP70, CHGB, CHMP2B, CHRNB2, CLU, CR1, CSMD1, CST3, CTSF, DSG2, EBLN1, EPHA1-AS1, EXOC3L2, FAM47E, FANCD2, FERMT2, FPR1, FRAS1, FRMD4A, GAL3ST4, GPR45, GRIN2B, HERC6, HFE, HMGCR, IL1B, INPP5D, IP6K3, IPMK, IQCK, KCNQ3, KIF13B, KLHDC4, LRAT, MAGI3, MAPT, MEF2C-AS1, MS4A1, MS4A13, MS4A14, MS4A2, MS4A3, MS4A4A, MS4A4E, MS4A6A, MS4A6E, MS4A7, MSRB3, MYCBPAP, NECTIN2, NFATC1, NFIC, NLGN1, NT5C3A, OPRD1, OPRM1, OR52E4, PDE6B, PEBP4, PICALM, PRNP, PSAP, PSEN1, PSEN2, PTK2B, PVR, QRFPR, RGS11, SIRT1, SLC22A14, SLC24A4, SORCS1, SORL1, SPI1, SUN2, SYNPR, TFAM, TM2D3, TNK1, TOMM40, TP53INP1, TREM2, TREML1, TREML2, TREML4, TTBK2, TTR, UNC5C, WDR46, ZCWPW1, ZNF6468242192Alexander diseaseGFAP108583Amyotrophic lateral sclerosisALS2, ANG, APEX1, ARHGEF28, C9orf72, CCNF, CHCHD10, CHGB, CHMP2B, DAO, DCTN1, FUS, GLE1, GRN, HFE, HNRNPA1, KIF5A, LIF, LRSAM1, MATR3, MOB3B, OGG1, OPTN, PARK7, PFN1, PON1, PON2, PRPH, SETX, SIGMAR1, SOD1, SPAST, SQSTM1, SS18L1, TARDBP, TBK1, TUBA4A, UBQLN2, UNC13A, VAPB, VCP7622834Canavan diseaseASPA86285Cockayne syndromeERCC5, ERCC6, ERCC8124216Creutzfeldt-Jakob diseasePRNP, SPRN70547Dystonia musculorum deformansADCY5, ATM, ATP1A3, GCH1, GNAL, PNKD, PRKRA, SGCE, SLC2A1, THAP1, TOR1A173908Familial amyloid neuropathiesAPOA1, GSN, TTR1471109Fatal familial insomniaPRNP8810Frontotemporal lobar degenerationCCNF, CHCHD10, CHMP2B, DAPK1, FUS, GFAP, GRN, GSK3B, LRRK2, MAPT, MOB3B, OPTN, PRNP, PSEN1, SOD1, SQSTM1, TARDBP, TBK1, TMEM106B, TREM2, UBQLN2, VCP37623011Gerstmann-Straussler-Scheinker diseasePRNP362912Hepatolenticular degenerationATP7B1665113Hereditary sensory and autonomic neuropathyFAM134B, IKBKAP, NGF, NTRK1, PRNP, RAB7A, SPTLC1, SPTLC2, WNK1883314Hereditary sensory and motor neuropathyDCAF8, DYNC1H1, EGR2, FGD4, FIG4, GDAP1, GJB1, HSPB2, HSPB8, KIF1B, LITAF, LMNA, MFN2, MPZ, MTMR2, NDRG1, NEFL, PEX7, PHYH, PMP22, PRX, RAB7A, SBF2, SH3TC2, SLC12A6, TFG24814015KuruPRNP1116Lewy body dementiaCYP2D6, DNAJC13, GBA, LRRK2, PRNP, PSEN1, PSEN2, SNCA, SNCB321217Lafora diseaseEPM2A, NHLRC11192518Lambert-Eaton myasthenic syndromeSYT22119Lesch-Nyhan syndromeHPRT11735520Myotonia congenitaCLCN1, SCN4A1175521Menkes Kinky hair syndromeATP7A1632622Multiple system atrophyCOQ2, POLG26523Neuronal ceroid-lipofuscinosesCLCN6, CLN3, CLN5, CLN6, CLN8, CTSD, MFSD8, POLG, PPT1, SGSH, TPP13931124NeuroFibromatosesNF2109625Optic atrophyAFG3L2, MFN2, OPA1, OPA3, SLC25A462986126Parkinson diseaseABCA7, ADORA1, APOE, BST1, BTNL2, CD2AP, CLU, CR1, DGKQ, DNAJC13, FBXO7, GAK, GALNT3, GBA, GCH1, HLA-DRA, LRRK2, MAPT, MS4A6A, NUCKS1, PARK2, PARK7, PCGF3, PICALM, PINK1, PM20D1, PODXL, PRDM2, PRNP, PTRHD1, RIC3, RIT2, SEMA5A, SLC2A13, SLC41A1, SLC45A3, SLC50A1, SNCA, SPPL2C, SREBF1, SYNJ1, TMEM175, VPS356167827Pantothenate kinase-associated neurodegenerationPANK2, RAB39B1323128Pelizaeus–Merzbacher diseasePLP1964929Progressive Bulbar palsySOD1, TTR3330Progressive supranuclear palsyDCTN1, MAPT, PARK2161031Rett syndromeCDKL5, FOXG1, MECP239410032Spinocerebellar degenerationsAFG3L2, C10orf2, CACNA1A, CACNA1G, ELOVL4, ELOVL5, ITPR1, KCNC3, KCND3, SPTBN2, TGM6, TMEM240, TTBK2533333Spinal muscular atrophy of adultsHEXA, LMNA, SMN1, VAPB411734Spinal muscular atrophies of childhoodHEXA, IGHMBP2, SMN1492235Tourette syndromeHDC, SLITRK16336Tuberous sclerosisTSC1, TSC25752237Unverricht-Lundborg syndromeCSTB, PRICKLE1, SCARB21914

study

Totally, 5680 amino acid variations were collected and 2839 of them are substitutions without duplicates. Among these variations, the arginine (R) residue is the most common one (both in mutated residues and mutants), which is in agreement with the previous study (12). Top row of Figure 1 shows the amino acid distributions for the wild (left), mutated (middle) and mutant (right) residues of the studied proteins. For the wild amino acids in the studied NDD associated proteins, Tryptophan (W) happened with the lowest observed frequency and was chosen as reference (marked ‘1’). We further calculated the mutability of amino acids as both mutated (middle) and mutant (right) residues. The result shows that R, G, L are the most common mutated residues while R, V, S are highly mutant ones in NDD associated proteins. The variation profiles for all NDDs are similar as that of a larger dataset illustrated in the previous study which contains over 2000 variations related to multiple diseases, as well as some variations related to some NDDs, like amyotrophic lateral sclerosis, etc. (13).

other

Neurodegenerative diseases (NDDs), caused by the progressive dysfunction of neurons, are very common worldwide, affecting people of all ages but especially the aged ones. Since the first patient was diagnosed with Alzheimer’s disease a century ago (2), millions have been found suffering from the neurodegenerative disorders such as Alzheimer’s, Parkinson’s and Amyotrophic lateral sclerosis. According to the MeSH database (http://www.ncbi.nlm.nih.gov/mesh), there are totally 55 subdivisions under the term ‘neurodegenerative disease.’ Although the pathogenesis of NDDs can be quite different, a growing number of studies implied that many NDDs are associated with genetic variations, most of which locate in completely unrelated genes. In spite of the separate symptoms, the intracellular mechanisms of different NDDs share a lot in common. Mitochondrial dysfunction and oxidative stress were reported to play a causal role in pathogenesis for many NDDs (3). The genetic deficiencies in different disorders are also associated, for instance, the poly glutamine mutant which is induced by repeat of CAG nucleotide triplet, was regarded as the dominant pathogenesis for many inherited NDDs such as Huntington’s disease and spinocerebellar ataxias (4).

other

An integrated literature-based variation database for a series of related diseases can serve as a research platform for further discovering of relationships between diseases and genetic variations. The variation data of NDDs can be compared with that of other diseases, e.g. immunodeficiencies (5), to find their similarities and differences. It can provide data for identification of common variation spectrum within NDDs and for developing universal biomarkers or drugs for NDDs as well. At present, 4 Locus-Specific Databases (LSDBs) have been created for 5 individual NDDs (Table 1). However, information of more other NDDs is not collected yet. In this study, we collected variation data manually for all NDDs and stored them in the latest LOVD system (6). We intended to provide a complete and cross-diseases platform including up-to-date genetic variation information related to all subdivisions of NDDs. The integrated database will serve as a valuable tool for quickly querying of the NDD-related variations and systematically analysing of the relationships between diseases and variations. Table 1.Existing LSDBs for NDDsNo.Disease nameDatabase nameWebsite address1Alzheimer disease/Frontotemporal lobar degenerationAlzheimer disease and frontotemporal dementia mutation database (24)http://www.molgen.ua.ac.be/ADMutations/2Alzheimer disease/frontotemporal lobar degenerationAlzforum mutation database (25)http://www.alzforum.org/mutations3Amyotrophic lateral sclerosis (ALS)ALS mutation database (26)https://gwas.biosciencedbc.jp/cgi-bin/hvdb/hv_pos.cgi? gid=EG1044Amyotrophic lateral sclerosis (ALS)ALSOD (27)http://alsod.iop.kcl.ac.uk/Als/Overview/gene.aspx? gene_id=SOD15Parkinson diseaseParkinson disease mutation database (24)http://www.molgen.ua.ac.be/PDmutDB/6Parkinson diseaseParkinson disease mutation databasehttps://gwas.biosciencedbc.jp/cgi-bin/hvdb/hv_disease.cgi? did=27Parkinson diseaseParkinson’s disease mutation databasehttp://grenada.lumc.nl/LOVD2/TPI/home.php8Rett syndromeRettBASE (28)http://mecp2.chw.edu.au/#mutations9Tuberous sclerosisTuberous sclerosis databasehttp://chromium.lovd.nl/LOVD2/TSC/home.php10Tuberous sclerosisBIPMed–variants in tuberous sclerosis patients from Brazil (29)http://bipmed.iqm.unicamp.br/tuberous-sclerosis/genes/TSC2

study

Most of diseases are associated with genetic variations including point substitutions, copy number alterations, insertions and deletions. The genetic variations in DNA sequences may lead to abnormal messenger RNA splicing or coding and produce pathogenic proteins. It is well-known that the relationship between genes and diseases are often multiple to multiple mode i.e. one disease is often associated with many variations in different genes; the variations in the same gene may be responsible for several different diseases (1). These disease-related genetic variations were uncovered and kept in many individual studies. The literature-based genetic variation repository, therefore, exerts considerable significance in systematically pathology study.

other

In LOVD3.0 view, the information is shown in eight different tags: genes, transcripts, variants, individuals, diseases, screenings, submit and documentation. It is a LOVD standard structure, which provides series useful information through a user-friendly interface by clicking each hyperlink. Users can also register as submitters to search the database or submit new genes and variations. The LOVD also supplies functions for importing and exporting data between different resources and the NDDVD is available at http://bioinf.suda.edu.cn/NDDvarbase/LOVDv.3.0.

other

The disease list of NDDs was obtained from the category of ‘neurodegenerative disease’ in MeSH database (http://www.ncbi.nlm.nih.gov/mesh). Each disease associated gene and variation was collected by the following method, i.e. the disease name or aliases plus ‘variation’ or ‘mutation’ was used as the keyword for querying in PubMed. The disease aliases are listed in Supplementary Table S1. The query results were further checked and the disease-related genes, SNPs or amino acid variations were then extracted manually. The variations collected encompass both those associated with autosomal dominant disease as well as those identified through association studies which affect disease risk. Variants in both translated and un-translated regions are collected. The general gene information such as gene synonym and genome location was fetched from NCBI. For variants with SNP IDs, the associated genes are described as the same as that used in dbSNP. If a variant does not located exactly within a gene region in dbSNP, we describe it based on their genomic information in NDDVD. The personalized information of patients, especially their ethnicities, demographic and epidemiological data if have, was collected from the reference papers to provide data for the future stratified medicine study, since these data will be useful to the future classification of patients to different subgroups (Supplementary Table S2). We followed the guidelines and standard for establishing Locus Specific Databases (7) when building NDDVD. That is to say, HGNC gene names (8) and HGVS variation nomenclature were applied when we built our database (9). For each gene, the reference mRNA and amino acid sequences were searched and recorded from locus reference genomic (LRG) database (10). If the sequence was not available, we used RefSeq and UniProt instead. The biological effects of DNA, RNA and protein variations were annotated using Variation Ontology (VariO) (11). The steps of this manually screening process are shown in Supplementary Figure S1.

other

LOVD (Leiden Open Variation Database) platform, supplied by Leiden University Medical Center, provides a flexible, freely available tool for gene-centered collection and display of DNA variations. The current version LOVD 3.0 extends this idea to provide storage for patient-centered data, NGS data and even variants outside of genes. The NDDVD database is established based on this new LOVD version which supports the storing of genes and variations of different diseases in one database.

other

To study the biological functions of the variations, especially the amino acid substitutions, a number of models and tools were developed to characterize their effects to protein’s sequence conservation, structural stability, aggregation, disorder, etc.(14–17). We developed a SVM classifier for predicting the effects of variations on protein stability based mainly on structural information especially the change of contact energy (17, 18). PON-P2 is a machine learning-based classifier and groups the variants into pathogenic, neutral and unknown classes, on the basis of random forest probability score (19). SIFT predicts the effects of all possible substitutions at each position in the protein sequence by using sequence homology (20).

other

We still chose the previous three diseases, MC, DMD and FTLD for the analysing. For MC there are 83 amino acid substitutions collected from 2 genes, SCN4 and CLCN1, corresponding to protein sodium channel protein type 4 subunit alpha and chloride channel protein 1, respectively. The results are shown in Supplementary Table S3. Both proteins have such a long sequence, and the results indicated that majority variations are on the residues with high conservation. 25 variations are predicted to be pathogenic by PON-P2 and they are all in high conservation positions. Some variations in very low conservation sites, like p.Gln831Arg, p.Ala659Val and p.Phe167Leu, are all considered to be neutral by PON-P2. There is no stability prediction results since the structure of these two proteins are not available in PDB.

other

Amino Acid distribution and variation profiles. Top row, amino acid distribution (left), overall mutability of mutated (middle) and mutant residues (right) for all the NDDs related proteins. The same information for Myotonia Congenita (MC), Dystonia Musculorum Deformans (DMD) and Hereditary Sensory and Frontotemporal Lobar Degeneration (FTLD) are presented, respectively, in the lower three rows.

other

Variation distribution according to physiochemical properties. Left top: mutability distribution for all NDDs in 36 variation situations: number 1–6 denotes to 6 groups of amino acids according to their physiochemical properties: (1) hydrophobic (V, I, L, F, M, W, Y, C), (2) negatively charged (D, E), (3) positively charged (R, K, H), (4) conformational (G, P), (5) polar (N, Q, S) and (6) Alanine and Threonine (A, T) group. The same information for Myotonia Congenita (MC), Dystonia Musculorum Deformans (DMD) and Hereditary Sensory and Frontotemporal Lobar Degeneration (FTLD) are displayed in the right top, left and right bottoms, respectively.

other

We chose three diseases abundant with variations, i.e. Myotonia Congenita (MC), Dystonia Musculorum Deformans (DMD) as well as Frontotemporal Lobar Degeneration (FTLD) to study their mutation profiles (lower three rows in Figure 1). Arginine (R) is the most common variant residues for all the NDD diseases, the same as analysed in previous study. The previous research reported that random variations at W and C are the most pathogenic (12). The mutant C in MC and the mutant W in DMD are one of the most common mutants in the disease, although this is not observed in MC (W), DMD(C) and FTLD (W and C). This could be caused by the low occurrence rate of C and W residue itself in the disease associated proteins. This difference needs to be further investigated considering their specific pathogenic mechanisms.

other

To investigate the functional effects of these variations, we grouped the 20 amino acids into six groups based on their physicochemical properties as, hydrophobic (V, I, L, F, M, W, Y, C), negatively charged (D and E), positively charged (R, K, H), conformational (G and P), polar (N, Q, S) and (A and T) (14). Therefore, the variations (substitutions) can be divided into 36 types based on the changes between 6 types of mutated residues and 6 types of mutants. For all the NDD diseases studied here, the most common mutations are physicochemical property changes from hydrophobic to itself (1 to 1), from positive charged, conformational, A and T to hydrophobic (3 to 1, 4 to 1 and 6 to 1), respectively. The variation profiles are partially similar for MC, DMD and FTLD, e.g. the variation ratios from hydrophobic to hydrophobic residues are very high in all three diseases, and it is easy to be understood since the hydrophobic group is the biggest one among the six groups. There are some obvious different variation profiles between these diseases, e.g. variation from positively charged to hydrophobic (3 to 1) is one of the lowest types for MC, while for DMD and FTLD it is one of the highest types; for FTLD conformational to polar (4 to 5) is very common, as shown in Figure 2.

other

In total 101 variations from 10 different proteins are found for DMD, the analysis result is shown in Supplementary Table S4. Most residues are high conserved predicted by SIFT. Three proteins, solute carrier family 2 and facilitated glucose transporter member 1 (gene: SLC2A1), interferon-inducible double-stranded RNA-dependent protein kinase activator A isoform 1 (gene: PRKRA) and serine-protein kinase ATM isoform a (gene: ATM) have their structures reported in PDB, all the 19 variations found on them are predicted to decrease the protein stability by our method. Variations of 61 are predicted to be pathogenic by PON-P2 and all of them are in high conservation positions with extremely low SIFT scores.

other

In total 142 variations from 20 different proteins related to FTLD are analysed (Supplementary Table S5). Some variations are in low conservation position predicted by SIFT and most of these variations are considered not pathogenic by PON-P2 (unknown or neutral). Variations from 15 proteins can be analysed by PPSC since the structures are available. Only two variations, p.Lys238Glu in protein sequestosome-1 isoform 1 (gene: SQSTM1) and p.Lys263Glu in protein TAR DNA-binding protein 43 (gene: TARDBP) can increase protein stability using PPSC prediction while others are all predicted to decrease protein stability. Usually the variations that are considered as pathogenic by PON-P2 are in high conservation positions. But for FTLD related cases, there are a few exceptions: e.g. p.Leu424Val in protein presenilin-1 isoform I-467 (gene: PSEN1), p.Pro348Leu in protein sequestosome-1 isoform 1 (gene: SQSTM1) and 8 variation in protein transitional endoplasmic reticulum ATPase (VCP).

study

Variations found in more than one disease were also collected and analysed. There are 67 such variations found in 25 proteins (Supplementary Table S6). Majority of variations are happened at the conserved sites, only a few exceptions, e.g. p.Ile723Val, p.Val380Leu and p.Ala53Thr are predicted very low conserved (Supplementary Table S7). Of 39 variations with structure information, 33 are predicted decreasing the protein stability. About one third of the variations (24 out of 67) are predicted to be pathogenic by PON-P2. But there is no direct relationship found with conservation or stability prediction result. Since the dataset is not big enough, further studies are required in future.

other

The primary objective of this work is to design a disease-centric resource for further data analysis and clinical research. We will make the database open to data submission and expert checking, and try to develop data mining tools to collect and update data from existing database automatically. In addition, more tools will also be developed for the analyses and applications of the variations.

other

With the paradigm shifting toward personalized medicine and precision medicine, the personal phenotyping data will be collected for the precision mapping to the genotyping information (21, 22). The NDDVD database will be updated with more personalized and paired genotyping-phenotyping data for systems or network level modeling (23), which will be helpful to the future screening of high risk NDD population and personalized diagnosis and treatment of NDD patients.

other

This work is supported by the National Key Research and Development Program of China (No. 2016YFC1306605), the National Nature Science Foundation of China (Grant No. 31670851, 31470821, 91530320, 61602332, 31600671) and the University Science Research Project of Jiangsu Province (No.14KJB520035).

clinical case

Respiratory syncytial virus (RSV) is a leading viral cause of acute respiratory illnesses (ARI) worldwide (Haynes et al. 2013), with the virus infecting 5–10% of the world population annually (Falsey et al. 2005) resulting in an estimated 3 million hospitalizations of children aged under 5 years (Nair et al. 2010) and more than 160,000 deaths across all age groups each year (Nair et al. 2010). An important epidemiological feature of RSV disease is its highly seasonal patterns in communities (Stensballe et al. 2003). Globally, RSV disease occurs as recurrent annual epidemics that peak during the winter in temperate climatic regions but shows less consistent timing in the tropical or subtropical climatic regions (Stensballe et al. 2003; Haynes et al. 2013). No licensed RSV vaccine exists but several candidates are in development with some in phase three trials (Higgins et al. 2016). Infection prevention and treatment are currently limited to passive immunoprophylaxis, case isolation, and supportive care (Drysdale et al. 2016).

study

RSV belongs to family Paramyxoviridae and its genome is a non-segmented single-stranded negative-sense RNA molecule (∼15,200 nucleotides long) that encodes eleven viral proteins (in the order NS1-NS2-N-P-M-SH-G-F-M2 (1 and 2)-L). Two genetically and antigenically distinct RSV groups are recognized (A and B) whose local predominance alternates over successive epidemics (Mufson et al. 1985; Cane 2001, 2007). Based on phylogenetic analysis of the immunogenic and variable attachment (G) gene (Johnson et al. 1987), at least eight genotypes (and several variants within these genotypes) have been identified within each of the two groups (Peret et al. 1998, 2000; Agoti et al. 2015a). Analysis of RSV strains detected in several parts of the world found that RSV epidemics frequently comprise multiple genotypes (and variants) but locally a single genotype normally predominates an epidemic with periodic replacement in successive epidemics (Cane et al. 1992; Peret et al. 1998, 2000; Agoti et al. 2015a; Otieno et al. 2016).

other

Improved understanding of RSV epidemiological patterns, transmission chains, and mechanism of persistence in host populations can help with infection control (Munywoki et al. 2014; Agoti et al. 2015a). Information on the origins of RSV seed strains for local epidemics, hubs of virus transmission, and spread patterns during outbreaks is limited (Nokes and Cane 2008; Munywoki et al. 2014; Agoti et al. 2015a). Detailed molecular analyses of RSV strains sampled during epidemics have the potential to elucidate these patterns (Agoti et al. 2015a and 2015b). However, such studies to date have primarily used samples collected from hospitalized individuals, representing a small and biased proportion (<1%) of all RSV infections during epidemics (Cane 2007). Community-based studies of RSV are rare (Munywoki et al. 2014). As a result, many aspects of RSV transmission, spread, and survival in the settings where majority of the infections occur remain unknown.

other

RSV surveillance in Kilifi County, located in coastal Kenya, has been ongoing since 2002 with a continuous hospital-based arm and intermittent community-based arm (Nokes et al. 2004, 2008, 2009; Munywoki et al. 2014; Agoti et al. 2015a]. Recently, we reported the RSV infection epidemiological findings from a cohort of forty-seven households followed over one epidemic season (Munywoki et al. 2014). Consistent with previous findings in developed countries (Hall et al. 1976) school-going children were found to be frequent introducers of the virus into households (Munywoki et al. 2014). Infection spread in the households was confirmed by group matching (typing into RSV A and B) and nucleotide comparison of the G gene (Munywoki et al. 2014). However, efforts to map transmission chains by combining the date of sampling and G sequence results showed limited success due to low phylogenetic signal from this short fragment (Munywoki 2013, 2014).

other

The intensive sampling regime during the household study provides an opportunity to uncover RSV transmission and evolution patterns in community epidemics. We recently showed that analysis of the relatedness of G gene sequences identified within and between epidemics can distinguish virus strains newly introduced into the community from those locally persisting (Agoti et al. 2015a). We also pointed out that a large fraction of RSV strains collected from local epidemics possess identical or highly similar G sequences (Agoti et al. 2015a; Zlateva et al. 2004; 2005). This illustrated the challenge of low phylogenetic resolution in undertaking detailed tracking of RSV transmission in a community by analyzing G gene sequences alone (Munywoki et al. 2014). However, when we compared full genomes of G identical strains, nucleotide differences were found occurring outside the G region (Agoti et al. 2015b). Thus, increasing the examined sequence length can provide much-needed additional phylogenetic resolution for monitoring virus transmission over short times (Cotten et al. 2013).

other

The analysis reported here investigated RSV A transmission in a community setting, the source of seed viruses and genomic diversification in a subset of samples collected during the household cohort study (Munywoki et al. 2014). We assessed the strength of the phylogenetic signal provided by analyzing the individual RSV genes versus for the whole genome sequences in tracking RSV transmission and the relatedness of the household viruses to contemporaneous strains across the world (Do et al. 2015). Further, due to the close monitoring of this cohort we were able to observe changes occurring at the consensus genome level intra- and inter-host during household transmission of RSV. In this report we show the utility of whole genome sequencing in defining RSV transmission, persistence, evolution and spread in households and at the local community level.

other

The samples analyzed in this study were collected following an informed written consent from each individual participant if aged ≥18 years or through a guardian or parent if aged <18 years and all children assented to participate. The study protocol was reviewed and approved by both the Scientific and Ethics Review Unit (SERU) of the Kenya Medical Research Institute (KEMRI), Nairobi, and Coventry Research Ethics Committee, UK (Munywoki et al. 2014).

other

RNAs were extracted from raw nasal specimens using the QIAamp viral RNA extraction Kit following the manufacturer’s instructions (QIAGEN Ltd, London, UK). Complementary DNA (cDNA), PCR amplification and nucleotide sequencing of RSV genomes were performed as previously described (Agoti et al. 2015b). Briefly, the RSV genome was amplified as six overlapping fragments, which were henceforth pooled and used to prepare Illumina NGS libraries. These were subsequently sequenced using Illumina MiSeq, multiplexing 15 to 20 samples per run, to generate approximately 1-1.5 Million paired-end reads (150 bp × 2) for each sample.

other

Raw sequence data from MiSeq were de-multiplexed into sample specific readsets and processed in QUASR (Watson et al. 2013) to remove low quality reads (median Phred score of <35) and primer and adapter sequences at the end of the individual reads. The resulting reads were de novo assembled using the SPades Program v3.5.0 (Bankevich et al. 2012) into contigs, examined for completeness of the expected open reading frames and, where necessary, partial contigs were further combined using Sequencher v5.0.1. To avoid errors due to crosstalk between multiplexed samples only contigs with a median read coverage of > =500 were used. Genomes with gaps (< 500 nucleotides) were joined with a series of ambiguous nucleotides (Ns) using the most complete genome from the same household as a guide for inferring the length of the gap. Multiple Sequence Alignments (MSA) were generated in MAFFT v6.83 (Katoh et al. 2002).

other

Nucleotides at polymorphic positions on the genomes were checked as follows: A sequence alignment for each household was generated (all sequenced viruses) and any nucleotides showing variation from the group were directly examined. For each observed variant site, a 21-nucleotide (nt) motif spanning the variant nucleotide (normally at the center but adjusted for variants near the termini) was prepared. The frequency of these 21-mers (both forward and reverse complement sequences) in the quality-controlled short read data was then determined using a modified grep script Cartman.py (available at https://github.com/mlcotten/RSV_household_scripts) using ack (http://beyondgrep.com/why-ack/) and the majority nucleotide kept. In addition, all indels were directly examined and all ambiguous nucleotides (R, Y, S, W, M, K) were resolved by a similar direct read counting and with the ambiguous nucleotide replaced by the absolute majority nucleotide. In cases of a position having 2 or more variants with equal counts, the nucleotide variant present in the majority of the genomes from the study was used.

other

The household study was undertaken within Kilifi County of Coastal Kenya in two local administrative units located to the north of the Kilifi Health and Demographic Surveillance System (KHDSS) (Scott et al. 2012). A household (HH) was defined as group of people living in the same compound and eating from the same kitchen (Munywoki et al. 2014). The area is primarily rural, with a number of small markets and the key economic activities include small-scale crop and animal farming, fishing and tourism. Overall, the county experiences a tropical climate with bimodal annual rainfall pattern: main rains April-July and shorter rains October-December. Annual RSV epidemics in this region, as recorded through surveillance in the Kilifi County Hospital (KCH), typically start in October-December of one year and continue to June-August of the following year (Agoti et al. 2015a; Nokes et al. 2009). The GPS locations of study households were recorded and entered in a confidential database. These addresses were validated in Google Earth and then visualized in QGIS v2.2 program (http://www.qgis.org/en/site/) overlaid with regional amenities data including local schools and main roads, Fig. 1. The sampled households occurred within an area of approximately 12 km2. Figure 1.Geographical distribution of the nine studied households which each had at least one assembled genome. Also shown is the Mombasa-Malindi highway, roads and schools in the study area. Light grey lines indicate administrative sub-location boundaries.

other

A detailed description of the household study design was provided in previous publications (Munywoki et al. 2014, 2015a, 2015b). Briefly, 47 households were recruited and closely followed up over a 6-month period between December 2009 and June 2010 to document all respiratory virus infection episodes. Twice weekly throughout the observation period, a nasopharyngeal-flocked swab was obtained from every household member regardless of the symptoms status. More than 80% of the planned samples were collected (Munywoki et al. 2014). The specimens were screened for a range of respiratory viral nucleic acids including RSV using multiplex real-time RT-PCR method (Gunson et al. 2005). A cycle threshold (Ct) of 35.0 or below was considered indicative of infection with the associated virus. In the current analysis all RSV A positive samples (187 RSV A mono-infected and 12 RSV A-B co-infected) from a select 13 households of the 47 were processed for whole genome sequencing and analysis, Table 1. These households were prioritized for analysis because RSV infection (group A or B) was detected in more than one member within a week suggesting a household RSV infection outbreak. The specimens had been collected between March and May of 2010 from 63 subjects. The arising sequence data were analyzed both independently and together with sequence data of RSV A strains from other countries deposited into GenBank. Table 1.Characteristics of the households from which we analyzed RSV A positive samples and sequencing results.HH IDHH size% Female% In schoolMedian age (IQR) in yearsMedian number of samples (IQR)aNumber of RSVA Positive samplesNumber of Genomes53764.924.311.4(3.3–23.5)31(16–42)702466100.050.011.4(1.9–16.5)45.5(45–46)21122050.030.016.6 (4.9–24.9)24(11.5–40)1014633.350.06.3 (2.8–9.4)44.5(43–45)1812191457.150.013.0 (7.6–35.4)41.5(34–43)1026580.060.05.6 (2.7–11.5)46(46–47)9929742.942.97.9 (2.2–27.5)43(42–43)2512311172.727.38.1 (2.3–27.6)31(6–32)115382343.543.512.6 (7.1–27.4)40(36–43)242240540.040.06.1 (2.0–8.9)45(45–45)1210451070.080.011.4 (6.7–18.5)42.5(31–45)60511573.346.79.2 (3.3–28.4)42(28–44)20571643.850.012.9 (7.9–17.5)28(21–29)188Abbreviations: HH for Household, ID for identity and IQR for interquartile range.Near complete RSV genomes were obtained from only 9 of the 13 households we analyzed.aThis refers to number of samples collected per a person in the respective households over the entire study period.

other

A total of 131 virus genomes for which the assembly yielded contigs >5000 nucleotides long were included in the analyses (i.e. gene-by-gene and whole genome analysis). These genomes were derived from 9 households. Of the 131 genomes, 103 were > 14000 nt in length with fewer than 500 ambiguous nucleotides (henceforth referred to as genomes, the only set considered in the whole genome analysis level). The alignment of the full genome was trimmed to include only sequence region covered by all genomes to maximize homology. The aligned sequences were analyzed for recombination using the RDP4 program and no recombination was detected (Martin et al. 2015).

other

Three data sets were prepared for comparison with the household study viruses. First, 11 G gene reference sequences, one for each of the known RSV A genotypes (GA1-7, SAA1-3 and ON1) were prepared and used for genotyping the household viruses on the basis of phylogenetic clustering. Second, 275 RSV A G sequences collated from GenBank that were sampled from different countries across the world between 2009 and 2010 and also from the Coastal Kenya in-patient surveillance at the KCH (Otieno et al. 2016) were prepared and used for determining the number and a probable source of the virus variants that seeded the household infection outbreaks. The third set included 354 nearly complete RSV A genomes retrieved from GenBank. These, inclusive of only genomes with information on country of origin, date of sampling and no recombination detected, were used to determine the global phylogenetic placement of the household viruses genomes.

other

Phylogenies were generated from the nucleotide alignment of both whole genomes and from the excised individual genes. The trees were reconstructed using Maximum Likelihood (ML) method in either MEGA v5.22 (Tamura et al. 2011) or PhyML v3.1 program (Guindon et al. 2010). The best-fitted models of nucleotide substitution for each alignment were determined in IQ-TREE v1.4.3 (Nguyen et al. 2015). All gene-specific ML trees were inferred in MEGA under HKY85 model bootstrapping for 1,000 replicates. Whole genome ML trees were inferred in PhyML v3.1 under GTR + Γ4 model of substitution, with 1,000 bootstraps. A bootstrap value of >70% was considered as statistically significant.

other

The household viruses were genotyped by phylogenetic clustering pattern of their G ORF region with reference G sequences. Representative sequences of all known RSV A genotypes (GA1-7 & ON1) were included. A genome was assigned to a particular genotype if its G sequence clustered with the genotype reference sequence within the same branch with > 70% bootstrap support. To understand the evolution and transmission history of the identified viruses within the same genotype, the sequences were further typed into variants. Viruses were defined as same variant if their divergence was estimated to have occurred no more than a year before their date of collection and this helped identify independent virus introductions into the study area. We inferred these by considering the number of nucleotide differences observed in the G ectodomain region for virus pairs as recently described elsewhere (Agoti et al. 2015a). This method asserts that 4 or more nucleotide differences between viruses in the G ectodomain indicates a distinct virus variant, a criterion that takes into consideration the fragment length, substitution rate and time interval between the samples (Agoti et al. 2015a). The number of variants was also confirmed by the relatedness of the household viruses in the presence of contemporaneous background diversity from multiple countries across the world (Agoti et al. 2015a). A cluster was defined as a group of viruses that do not meet the distinct genotype or variant threshold rules but fall within one tree branch with a bootstrap support of > 50%.

other

The baseline characteristics of the households yielding RSV A positive samples and details on the number of genomes obtained per household are given in Table 1. Nucleotide changes were observed across the entire RSV genome (Fig. 2) in the 8 households with more than one genome sequenced. Within individual households, the number of nucleotide changes between virus genomes was variable and ranged from 0-17 nucleotides. Of the 131 specimens yielding contigs of >5000 nt, 120 from 10 households yielded an intact G coding sequence (CDS) and all these belonged to genotype GA2 and the closely related sub-genotype NA1 (result not shown). These household genomes formed a single monophyletic group within genotype GA2 on the global phylogeny (Fig. 3) that was most closely related to GA2 genotype viruses from Coastal Kenya that had been sampled from young children admitted to KCH in the years 2009 and 2010 . Further, the entire set of RSV A viruses from the households fell within a single variant definition as also determined by their clustering of the G gene genomic region in the global G-gene phylogeny (Supplementary Fig. S1). Figure 2.Nucleotide differences between viruses (total = 130) detected within the individual households. Each panel is a single household. The viruses were compared to the earliest virus genome sequenced from the same household. Vertical colored bars show the nucleotide differences. Red is a change to T, orange is a change to A, purple is a change to C and blue is a change to G. Grey is a deletion or an non-sequenced portion of the genome. Household six is excluded as only a single genome sequence was obtained. A python script to generate this figure is available at https://github.com/mlcotten/RSV_household_scripts. Figure 3.A ML inferred phylogenetic tree showing the global phylogenetic context of the RSV A household study genomes. The taxa of the household study viruses (n= 103) are in red while viruses from the rest of Kenya (inpatient) are colored blue. The taxa of RSV A viruses from around the globe are colored by continent of origin. Asterisk mark has been placed next to major branches with a bootstrap support of >70%.

other

Nucleotide differences between viruses (total = 130) detected within the individual households. Each panel is a single household. The viruses were compared to the earliest virus genome sequenced from the same household. Vertical colored bars show the nucleotide differences. Red is a change to T, orange is a change to A, purple is a change to C and blue is a change to G. Grey is a deletion or an non-sequenced portion of the genome. Household six is excluded as only a single genome sequence was obtained. A python script to generate this figure is available at https://github.com/mlcotten/RSV_household_scripts.

other

The temporal signal in nucleotide divergence of the household viruses was estimated in TempEst v1.4 (Rambaut et al. 2016) using a ML whole genome tree as input. The evolutionary pattern and time to the Most Recent Common Ancestor (tMRCA) of the obtained whole genome sequences were determined in BEAST v1.8.2 under HKY85 model of substitution, (uncorrelated) lognormal relaxed molecular clock and Gaussian Markov random field (GMRF) population skyride (Minin et al. 2008; Drummond and Rambaut 2007; Drummond et al. 2012). The Metropolis Coupled Markov Chain Monte Carlo (MC-MCMC) chain length was set to 50 Million steps sampling after every 2500 steps. The output was examined in Tracer v1.6 (http://tree.bio.ed.ac.uk/software/tracer/), with a 10% burn-in removal, to confirm run convergence (i.e. if the estimated sample size for all inferred parameters was >200). The output trees were summarized in TreeAnnotator (Drummond and Rambaut 2007) (with a 10% burn-in removal) and the resulting Maximum Clade Credibility (MCC) tree was visualized and annotated in FigTree v1.4.2 (http://tree.bio.ed.ac.uk/software/figtree/). A posterior probability of > 0.9 was interpreted as statistically significant.

other

The sequence nomenclature on the phylogenetic trees is country of origin (_sample source for Kilifi indicating if sampled from inpatient (IP) or household (HH))/Unique identifier/Date of specimen collection. The unique identifier for household samples includes the household identifier (first two digits) and subject identifier (the last two digits). All new sequences from this study were deposited in GenBank under the accession numbers KX510136-KX510266.

other

A ML inferred phylogenetic tree showing the global phylogenetic context of the RSV A household study genomes. The taxa of the household study viruses (n= 103) are in red while viruses from the rest of Kenya (inpatient) are colored blue. The taxa of RSV A viruses from around the globe are colored by continent of origin. Asterisk mark has been placed next to major branches with a bootstrap support of >70%.

other

A time-resolved phylogenetic clustering of the 103 household study genomes (Fig. 4, panel A) revealed that all viruses clustered by household of origin, except for those from households 26, 38 and 57. This pattern was also observed with a ML phylogeny (Supplementary Fig. S2) and MJT network that showed household-specific clustering of viruses as well as a varied level of the interconnection of viruses within and between households (Fig. 4, panel B). Viruses from households 5, 31 and 40 formed individual distinct household-specific clusters that included all virus genomes obtained from these households. In contrast, households 26, 38 and 57 had genomes from 2 or more separate branches, suggesting multiple virus introductions into each of these three households. Particularly in household 26, three virus genomes from individual 2605, collected on the 16th, 18th and 22nd March clustered with the other viruses from that household (Supplementary Fig. S2). However the virus genome obtained from 26th March appeared on a lone branch suggesting a second introduction of a genetically varied virus. Genomes from households 14 and 29 were interspersed within the same viral cluster. Household 6 provided only one genome. Figure 4.The sequence relatedness of the household study RSV A viruses. (a) A time-scaled phylogenetic tree of the 103 genome sequenced household study viruses inferred in BEAST program. The genomes are represented by a filled circle colored differently for each household (color scheme similar to Fig. 1). (b) A median-joining (MJ) haplotype network constructed from the 103 household genomes. Each colored vertex represents a sampled viral haplotype, with different colors indicating the different households of origin. The size of the vertex is relative to the number of sampled isolates. Hatch marks indicate the number of mutations along each edge. Small black circles within the network indicate unobserved internal nodes.

other

The sequence relatedness of the household study RSV A viruses. (a) A time-scaled phylogenetic tree of the 103 genome sequenced household study viruses inferred in BEAST program. The genomes are represented by a filled circle colored differently for each household (color scheme similar to Fig. 1). (b) A median-joining (MJ) haplotype network constructed from the 103 household genomes. Each colored vertex represents a sampled viral haplotype, with different colors indicating the different households of origin. The size of the vertex is relative to the number of sampled isolates. Hatch marks indicate the number of mutations along each edge. Small black circles within the network indicate unobserved internal nodes.

other

In contrast to the genome-based phylogeny, when considering individual gene ORFs, the resolution was reduced and fewer household-specific distinct clusters were identified compared to the full genome analysis. ML phylogenetic clustering of the sequenced viruses by ORF is shown in Supplementary Fig. S3 (whole genome phylogeny included for comparison purposes, panel xi). When we considered the G gene alone (901 nt), just one household had a distinct virus cluster (HH 31); the remaining clusters included viruses from multiple households. Similarly reduced resolution was obtained with the F gene (1727 nt) with only two household-specific clusters (HH 6 and 40), the nucleoprotein (N) gene (1200 nt, with also only two household-specific clusters (HH 5 and 40) and with the L gene (7915 nt), four household-specific clusters were observed (HH 5, 6, 31 and 40). For comparison, the full genome analysis showed seven household specific clusters.

other

The spatial distribution of the nine households is shown in Fig. 1. The geographical distance between the study households ranged from 302 to 3925 meters. There were a variable number of nucleotide differences across the genomes distinguishing clusters of viruses found in one household from the next (range 2-16), Fig. 4, panel B. The RSV A infection was first detected in household 40 (on 15th February) followed by 29 (21st February), 14 (1st March), 57 (3rd March), 5 (9th March), 26 (11th March), 31 (30th March), 6 (9th April) and finally household 38 (19th of April). For some of the study households, the infection periods overlapped. Notably, both HH 14 and 57, being the closest households in geographical distance (∼300 meters apart), had the first RSV infections detected in the first week of March (2 days apart) and virus strains were phylogenetically close when compared to strains from most other households we analyzed (Fig. 4 and Supplementary Fig. S2). This scenario was also observed with HH 6 and 38 (∼400 meters apart). Although these two cases were consistent with the hypothesis that physical distance modulates virus transmission and spread, there were household pairs that showed a contrary relationship, for example some members of household 14 and 29 gave multiple identical full genome sequences despite the two households being 1715 meters apart. Statistical analysis of the entire household dataset did not find a linear relationship between physical and genetic distance for this dataset (R2 = 0.01686).

other

Inferred virus transmission patterns within household 14. (a) Temporal infection patterns. Every rectangular box represent a sample collected from members of the household 14, if there is a circle inside implies the sample was RSV A positive. Unfilled circle implies specimen was not sequenced while filled colored circle implies sample was sequenced (whole genome). (b) A ML phylogenetic tree from whole genome sequences of 12/18 sequences sequenced. Same circle color for sample from the same individual. (c) A median joining haplotype network of 12 genomes. Each vertex presents a sampled viral haplotype, with different colors indicating different individuals who provided the sample. The size of the each vertex is relative to the number of sampled isolates. Hatch marks indicate the number of mutations along each edge. (d) The putative inferred transmission events. Continuous arrow indicates where the transmission link was inferred as highly likely while dotted arrows indicate where multiple alternative scenarios could have been the source of infection.

other

We reconstructed a plausible virus transmission chain between the household members by combining the genetic data with sampling dates. As examples we show analysis for HH 14, a six-member household (Fig. 5) and household 38, a 23-member household (Supplementary Fig. S4). In household 14, of the 18 RSV positive samples identified in this household, 14 assembled into contigs >5000nt and 12 gave near complete genomes. From the sample collection dates, we inferred that the individual designated 1404 introduced the virus into this household since this individual was the only virus positive person in this household on the 1st March (Fig. 5, panel A). Subsequently, the other household members designated 1401, 1402 and 1403 became virus positive within a week after the identification of individual 1404 RSV positivity. The genome data were consistent with individual 1404 (index case) infecting individuals 1402, 1403 and 1401 being identical or displaying only one nucleotide difference across their genomes, Fig. 2, Panel C. Each of the individuals 1405 and 1406 had both only a single virus positive sample collected on 15th March (two weeks after first sample from the index case). Sequencing was unsuccessful with the sample from individual 1405. However, the sample from 1406 had one or two nucleotide changes compared with all genomes in this household. The virus from individual 1406 was genetically closest to virus from individuals 1402 and 1404 but it is more likely that 1406 acquired the infection from individual 1402 who showed prolonged virus shedding. It is also important to note that some viruses identified in household fourteen were identical to those observed in household twenty-nine thus we could not exclude a second introduction of the virus into this household. Figure 5.Inferred virus transmission patterns within household 14. (a) Temporal infection patterns. Every rectangular box represent a sample collected from members of the household 14, if there is a circle inside implies the sample was RSV A positive. Unfilled circle implies specimen was not sequenced while filled colored circle implies sample was sequenced (whole genome). (b) A ML phylogenetic tree from whole genome sequences of 12/18 sequences sequenced. Same circle color for sample from the same individual. (c) A median joining haplotype network of 12 genomes. Each vertex presents a sampled viral haplotype, with different colors indicating different individuals who provided the sample. The size of the each vertex is relative to the number of sampled isolates. Hatch marks indicate the number of mutations along each edge. (d) The putative inferred transmission events. Continuous arrow indicates where the transmission link was inferred as highly likely while dotted arrows indicate where multiple alternative scenarios could have been the source of infection.

other

Individual 1402 was virus positive for the longest period (39 days) compared to other members in this household, Fig. 5, Panel A. Interestingly, the positive sample collected on the 15th April came after several samples collected between 20th March and 13th April had tested RSV negative. The virus from 1402 on 15th April had 3 nucleotide substitutions that distinguished it from all the other viruses sampled from this household, Fig. 5, panels B and C. This scenario could have arisen due to: (i) another virus introduction into the household or (ii) a virus rebound (recrudescence) from initial infection in this individual after accumulating these changes. Combining the genome sequence and temporal diagnostic information we inferred the transmission chain presented in Fig. 5, panel D, for this household.

other

TempEst analysis estimated that the MRCA for the household viruses occurred in December 2009 and their evolutionary rate was 4.948 ×10 − 3 sub/site/year. Notably, the R squared value for the linear model was 0.29 indicating the stochastic nature of variation observable in this limited time period. Different households had differing levels of diversity with only limited temporal relationship to this variation (Supplementary Fig. S5). Using BEAST program, the date of the MRCA for the household dataset was estimated to be 3rd Jan 2010 (95% HPD: 1st November, 2009 to 31st Jan, 2010), corresponding to the beginning of the Kilifi 2009/10 RSV epidemic season. This date was consistent with a single virus variant leading to the RSV A infections in all nine analyzed households. The BEAST-inferred genomic evolutionary rate for the household viruses was estimated as 2.307 × 10 − 3 (95% HPD: 0.935× 10 − 3 to 4.164 × 10 − 3) sub/site/year. This was about 5 fold higher compared to previous estimates for data derived across epidemics (Agoti et al. 2015b). While synonymous nucleotide (dS) changes were found in RSV encoded proteins, non-synonymous nucleotide (dN) changes were observed in only 7 of the 11 RSV proteins (NS2, SH, G, F, M2-1, M2-2, L) with the highest number of dN changes observed in the L protein region (11 independent changes). The NS1, N, P and M were totally conserved at the amino acid sequence level. A summary of the amino acid changes observed between the household genomes for all the ORFs are shown in Table 2. The F protein had the third highest number dN changes (most of these affecting 27-mer amino acid domain (pep27)). Changes in the G protein were spread throughout its length but outside of the central conserved cysteine noose region. All the household genomes contained six highly conserved N-glycosylation sites within their F protein, at positions 27,70,120, 126 and 500. Also six completely conserved N-glycosylation positions were found within the G protein: 85, 103, 135, 251, 273, and 294. All the household viruses were observed to encode uniform F and G protein lengths, 574 and 297, respectively. Table 2.Amino acid changes in the household viruses’ genomes by encoded protein.NS-1NS-2NPMSHGFM2-1M2-2LI43ME2GL13PI59VV125IY36CS224NR50GN34HF114SN40YN236DI49TP119LN44KK591NP143LG130SI83TI955IP146LV516AN970KT148AT529IT1045MD214ET1174ST268AI1588TE1619GL1746SY1762F

clinical case

Our knowledge of RSV transmission in the community, evolutionary patterns and ‘who acquires infection from whom’ (WAIFW) is incomplete (Agoti et al. 2015a; Munywoki et al. 2014). Close contacts within households, workplaces, worship places, market places and other social gathering avenues may provide opportunities for respiratory virus transmission (La Rosa et al. 2013). However, there is little evidence beyond temporal patterns of case occurrence to support that households are a major environment of RSV transmission (Munywoki et al. 2014; Hall et al. 1976). Viral genetic data can provide evidence to support epidemiological linkage of household RSV infected cases and to discount other sources of the infection.

other

Our findings support the hypothesis that RSV transmission within households is common as members belonging to the same household were infected with closely related strains, in terms of genomic sequence than viruses found in members from different households. Specifically, household-specific genomic variation was observed in seven of the nine households where we compared associated genomes. Only two households shared a genetically identical strain at full genome level. Notably, this between-household phylogenetic resolution was lost when examining the individual genes (including the G gene), as genetic variations between the sequenced viruses were random and distributed throughout their entire RSV genomes such that examining greater sequence lengths linearly increased the phylogenetic resolution achieved.

other

The genomes of all the household study viruses fell within a single branch on the global phylogeny and G gene analysis suggested that all the nine households were infected by a single virus variant that had entered into this community. Due to limited contemporary sequences from other parts of Kenya or Africa, it was not possible to identify close ancestors of this variant (Agoti et al. 2015a). Furthermore, it was not possible to infer the directions of household-to-household transmissions or pathway of the spread of infections reported here, because only a minority of the households in the study area were sampled. However, some of the households that were physically close happened to be infected by viruses that were also phylogenetically close. This is consistent with the idea that occupants of neighboring households are more likely to come into close contact during daily activities for example journeys to fetch water, to markets and clinics. It is also more likely that children in physically close households go to the same school, which are thought to be respiratory virus transmission hubs.

other

Within two individual households (HH 38 and 57), we observed higher genomic variation. We hypothesize three possible sources for this variation: (1) multiple virus introductions into these households, (2) co-infection of the index case with multiple genetic variants, and (3) diversification of a single virus in the process of replication and transmission through the members of the households. Some of the households had clearer evidence of multiple virus introductions (e.g. household fifty-seven) and this may be a result of factors that cannot be comprehensively investigated from our limited sampling. However, further analysis of these data including inspection of the minor variant populations is necessary to provide additional illumination (Hughes et al. 2012; Grad et al. 2014; Do et al. 2015). It is also possible that some of the observed changes simply reflected PCR and/or sequencing errors. However this is highly unlikely especially where nucleotide changes were observed at the same exact genomic position in multiple samples from the same household or individual despite their independent sample processing (Cottam et al. 2008). Also, importantly, only contigs with high read depth (> = 500) were included into our analysis.

other

The variation of genomes within households aided in identifying members who are likely to have shared an infection source or sequentially transmitted the infection from one to the other (e.g. the chains inferred for household fourteen and thirty-eight). However, it was not possible to elaborate in complete detail the transmission chains within most households even after considering these genomic data. This was partly due to incomplete sequencing (some samples had too low virus load) and also due to fact that the evolutionary rate of the virus was sometimes too low to provide a useful signal. This is likely to be caused by the highly infectious nature of RSV once introduced into a household setting resulting in overlapping infection generations before distinct nucleotide changes accumulate.

other

The evolutionary rates calculated at genome level from the household outbreak were significantly higher than rates derived from long-term data (Tan et al. 2012, 2013; Agoti et al. 2015b;). Our findings support the notion that evolutionary rates for viruses are highly context-specific and decrease when calculated from long-term sampling data (Duchene et al. 2014). This may reflect that deleterious mutations occurring during short-term transmission (and observed in the higher frequency sampling) that are purified from the virus population in the longer term. Multiple nucleotide changes were observed across RSV genome but some genes remained completely conserved at the amino acid sequence level. Although it is unlikely that the amino-acid substitutions observed represented adaptive evolution during short-term transmission of the virus, it will be worthwhile to further investigate their significance in allowing virus survival or escape from pre-existing immune responses.

other

Among respiratory viruses, viral genetic data have been previously utilized for influenza A viruses to define within and between household virus spread. Sequencing of hemagglutinin and neuramidase genes of 2009 pandemic H1N1 viruses found occurrence of only limited genetic diversity for viruses derived from different households early during the outbreak and diversity was negligible for viruses derived from same households (Thai et al. 2014). Deep sequencing of household viruses from Hong Kong revealed that genetic variation was more similar within than between households and associated information on minor variant sharing helped confirm transmission events (Poon et al. 2016).

other

For RSV, our study is the first of its kind using full genomic data to define patterns of its transmission in a community setting. Using temporal infection data alone, it has been previously concluded that young children are most likely to introduce RSV infection into households (Hall et al. 1976; Munywoki et al. 2014; Heikkinen et al. 2015) and the genetic data provided here support this conclusion. Within household RSV transmission has never been inferred to the detail described here. The evidence of multiple virus introductions in some households was particularly intriguing and would been missed if partial sequencing alone was deployed. Our study shows that patterns of shared virus strains between households can vary by the gene analyzed, but it is possible to separate almost all households as infected by a distinct virus strain by analyzing full genome sequences.

other

We are aware of limitations in this study. First, sampling in the households only reached ∼85.6% of the planned level with gaps mostly occurring in adults (Munywoki et al. 2014). Thus, it is possible that we missed important samples in inferring the transmission chains. Second, a significant proportion (34.2%) of the samples failed amplification, especially those with low viral load, hampering the reconstruction of transmission chains. However, this difficulty is common to all such studies (Memish et al. 2014; Bose et al. 2015 ). Third, PCR and sequencing errors were not completely modeled into the interpretation of our data (Orton et al. 2015). Despite our analytical stringency, it is possible that some of the nucleotide changes we observed could be artifacts especially those occurring in single genomes only. Fourth, we only analyzed a small proportion of households in the study area and important information such as contact patterns and school attendance were not factored into the analysis. This made it difficult to infer the broader community transmission pathways and exclude multiple sources of identical virus into a household.

other

In conclusion, our study has shown that the analysis of genome sequences provides better phylogenetic resolution in tracking RSV spread compared to analysis of small partial sequences including the highly variable G gene. Although whole genome analysis alone could not resolve every step in the transmission chains within households, the information derived distinguished many of the between-household transmission links and suggested clear epidemiological linkage of infections of some household members. The findings are consistent with a large percentage of RSV transmissions occurring within the household and thus infection control at the household level should be considered in RSV disease control. Future studies should include mathematical modeling to combine whole genome analysis (both consensus and minor variants data) with other epidemiological information (e.g. symptoms onset, viral load, immunity, social contact patterns, etc.) to allow mapping of WAIFW with regard to RSV spread within households.

other

One of the main research areas in metabolomics is to study the metabolic response to one or a few factors of interest in a given biological system. Design of experiment (DOE) has been widely employed to determine such cause-and-effect relationships. There are many statistical tests to analyse these data generated by different types of DOEs in univariate analyses. However, how to incorporate the DOE information into multivariate modelling is comparatively less well explored. There are several multivariate models proposed in the literature in recent years which can make use of a priori knowledge of DOEs, most notably multilevel simultaneous component analysis (MSCA) , analysis of variance (ANOVA)-principal component analysis (ANOVA-PCA) , and ANOVA-simultaneous component analysis (ASCA) . These methods share a common methodology, which is to decompose the independent data matrix X (i.e., the observed data generated by the instruments) to a series of sub-matrices according to the experimental design and perform principal component analysis (PCA) on the decomposed sub-matrices to study the effect of each factor separately. In addition, multi-block models, such as multi-block principal component analysis (MB-PCA) , have also been successfully employed to analyse such datasets by repartitioning X into blocks according to the experimental design, and then performing MB-PCA on the repartitioned multi-block data . In addition, multiple supervised models, mostly based on well-known partial least squares (PLS) have also been proposed in the literature using a similar methodology, such as priority PLS , ANOVA-PLS , ANOVA-target projection (ANOVA-TP) , and multi-block orthogonal PLS . All of these methods have, to date, focused on processing the X matrix: where X is either re-partitioned into blocks according to the experimental design (multi-block approaches) or decomposed into a series of sub-matrices (ANOVA approaches). However, designing the response matrix Y according to the information in the DOE and to build a supervised model to fit the designed Y may also be an efficient method to analyse the data generated by the DOE. In fact, this type of method has already been reported previously, albeit in a rather ad hoc manner . In our present study we aim to investigate such a methodology within the framework of structural modelling and propose a workflow for general use.

other

Historically, the design of output Y is usually categorised into two types: regression and classification. If the coded output is a series of continuous numbers (e.g., different concentrations of a specific metabolite, time points, temperatures, and so on), these numbers can be directly used as Y and the corresponding model is called a regression model (e.g., PLS-R). By contrast, if the target is a number of different groups (classes), such as different types of bacteria or different diseases, then Y is normally coded as a binary matrix in which one column represents one distinct group while each row is the target vector of a sample. A sample of a specific class has its element in the corresponding column coded as “1” and all other elements coded as “0”. The regression models are most suitable for modelling a series of continuous or at least ordinal (e.g., ranks) numbers, while classifications are most suitable for discriminating a set of categories “in parallel”; i.e., there is no particular spatial relationship between these categories.

other

However, there are cases when neither regression nor classification would be able to present the information in a DOE well. For example, if one conducted an experiment in which two different extraction methods (denoted as E1 and E2) were applied to extract metabolites from three different bacterial cells (denoted as B1, B2, and B3); where the objective was to investigate the differences in metabolic profiles of the three different types of cells and also to compare the extraction efficiency of the two extraction methods. To achieve this objective, a two factor full-factorial experimental design would typically be employed and six different combinations of the two factors need to be examined. This could be considered as a multi-class classification problem having six distinct classes, one for a specific combination of E and B. However, strictly speaking, in this particular case the six classes were probably not truly all “in parallel” because there were pairs of classes, which shared a common factor (e.g., E1B1 vs. E1B2). Therefore, it is reasonable to assume that the E1B1 class should be closer to E1B2 than E2B3. Such a spatial relationship, as a part of the prior knowledge of the experimental design, would be ignored by binary coding.

other

It has been recognised that not all problems can be explained well in real numbers (regression) or discrete coding (classification) schemes and, sometimes, more general, structured outputs are needed to cope with these data, which cannot be formulated as simple regression or binary classification problems . The key difference between classical regression/classification modelling and structured output modelling is that instead of using a simple error function (e.g., absolute difference between the predicted and known output for regression, correct or wrong (0/1) in predicting a class membership in classification) to evaluate how closely a predicted output matched the expected (target) output, structured output modelling evaluates such qualities using a set of errors. This error set enumerates all possible scenarios, which could happen in the prediction and ensure that there is a sensible gradient between these errors which can reflect the structured nature of the output . Using the same example given above, suppose there is a properly-trained model and one tests it with a group of blind test samples, if an E1B1-labelled sample had been predicted as E2B1 and another E1B1 sample had been predicted as E2B2, then the former prediction should have a lower error than the latter; i.e., the sensible error gradient between the three possible outcomes would be: a correct prediction in both E and B < one wrong prediction in either E or B < wrong predictions in both E and B. A sensible error set for this particular example could be defined as {0,1,2} and each prediction would have an error of one of those numbers in the set. The design of a structured output and the corresponding error set is largely problem-based and it has to be performed based on the available a priori knowledge, then one would need a modelling technique to build a model to establish the relationship between the structural coded output and the observed data and minimise the error derived from the designed error set. Many machine learning models have been extended to model structured data, such as structured support vector machines (S-SVM) , deep learning neural networks , etc. Compared to those machine learning approaches, PLS has a much larger following in the metabolomics community and almost all major statistical software packages provide easy-to-use PLS routines. As structured output coding only involves designing a more sensible targeted output Y and the (usually human) interpretations on the predicted outputs afterwards, this can be easily adapted by any software package which supports PLS.

other

In this study, we explored such a possibility of using a structured output, designed according to the DOE, for PLS modelling and compared the results with classic binary coding. Two recently published real metabolomics datasets were employed which used two different DOEs with different complexity and characteristics.

other

One dataset (denoted as riboswitch in this paper), was obtained from an experiment to investigate the metabolic effects of producing enhanced green fluorescent protein (eGFP) as a recombinant protein in Escherichia coli (E. coli) cells . A two-factor, full-factorial experimental design was employed and the metabolic profiles of five different E. coli strains, namely BL21(DE3) (wild-type), BL21(DE3) pET-empty (PET), BL21(DE3) pET-eGFP (EGFP), BL21(IL3) pET-empty (iL3PET), BL21(IL3) pETeGFP (iL3EGFP), under four different inducer conditions (control, lac inducer Isopropyl β-d-1-thiogalactopyranoside (IPTG), pyimido[4,5-d] pyrimidine-2,4-diamine (PPDA), and IPTG + PPDA) were measured by Fourier transform infrared (FT-IR) spectroscopy and gas chromatography-mass spectrometry (GC-MS) on cell extracts (the details of this experiment and a more detailed description of the strains can be found in ). Since the FT-IR and GC-MS data showed very similar results, only the GC-MS results were reported in this paper. The data will be uploaded and available at MetaboLights.

other

The other dataset, denoted as propranolol, was the results of an experiment investigating the role of efflux pumps in stress tolerance Pseudomonas putida exposed to toxic hydrocarbons. In this experiment three different strains of P. putida (denoted as DOT-T1E, DOT-T1E-PS28, and DOT-T1E-18) were exposed to four different dosages of propranolol (control, 0.2, 0.4, and 0.6 mg/mL), and three time points were monitored over a period of one hour (start (0), 10, and 60 min after exposure to the drug). At each time point, a sample of cells were collected, extracted, and analysed by GC-MS. The details can be found in and these data are available at MetaboLights under study identifier MTBLS320 . In our previous publications, we employed a series of different MB-PCA models to examine the effect of each factor separately while, in this study, we designed structured outputs to capture the essence of the experimental designs, and used PLS to model these structured outputs so that the effect of the factors of interest can be analysed simultaneously using a single model.

other

Since the results from the PLS modelling were in agreement with our previous reports, the detailed biological interpretations and significance of the results of the two datasets can be found in our previous publications (riboswitch and propanolol ) and will not be repeated again in this paper.

other

The structured output of this data set is illustrated in Figure 1 and a detailed description is given in Section 4.3.1. The PLS models were validated using a double cross-validation procedure and the reported results were the average of 1000 random splits of training and test sets as described in Section 4.3.2.

other

The confusion matrix of strain predictions are given in Table 1, while that of inducer condition predictions are shown in Table 2. The overall correct classification rate (CCR) for strain prediction was 80.21%, while that of the inducer condition was 58.20%, suggesting that the strain difference was a more dominating factor compared to the inducer conditions. A more detailed pattern can be revealed by inspecting the two confusion matrices. In the confusion matrix of strain prediction, wild-type and EGFP strains had obtained very high predictive accuracy (approximately 97% and 89%, and also very low off-diagonal misclassification errors of 0.08% and 0%, respectively), indicating that these two strains were most different to each other, while there was minor overlapping between iL3EGFP and iL3PET, which resulted in 8%–19% misclassification errors. This is consistent with the pattern revealed by MB-PCA reported previously . Additionally, in the confusion matrix of inducer condition prediction, there were seemingly two groups. One group consisted of No inducer and PPDA, there was moderate overlapping between these two conditions, and this resulted in 16%–19% misclassification error. IPTG and IPTG + PPDA formed another group and these two conditions overlapped heavily with each other, with misclassification errors as high as 34%. However, the errors between these two groups of conditions were generally lower (no more than 10.3%). This suggested that PPDA had a weak effect on the metabolic profiles of the bacteria cells. IPTG, however, had a very significant impact on the metabolic profile so that IPTG and IPTG + PPDA samples were better separated from those of the other two conditions. Again, this is consistent with the previously reported MB-PCA pattern. We also examined the predictive accuracy of inducer condition prediction on each strain separately, and the CCRs were 45.00%, 40.88%, 66.90%, 50.25%, and 59.58% for wild-type, PET, EGFP, iL3EGFP, and iL3PET, respectively. The much higher CCR in EGFP strains suggested that this strain was more sensitive to the addition of inducers than other strains. The CCRs of strain prediction under each inducer condition were 70.10%, 95.70%, 83.46%, and 72.62%, for control, IPTG, IPTG + PPDA, and PPDA, respectively. Interestingly, PPDA and control had similar CCR, while IPTG had the highest CCR. This also indicated that IPTG had the strongest effect on the metabolic profiles of the bacterial cells and different bacterial strains also responded differently to this inducer. These two sets of CCRs had revealed the delicate interactions between the bacterial strains and inducer conditions. We also employed PLS-DA models using classic binary coding and trained a 20-class classification PLS-DA model. The rearranged confusion matrices (in order to make them comparable with Table 1 and Table 2) are presented in Table 3 and Table 4. It can be seen that similar conclusions can be drawn from those confusion matrices. However, it is interesting to compare the resultant predictive accuracies of these two types of coding. The binary coding resulted in a similar predictive accuracy in predicting inducer conditions (59.37% vs. 58.20%) while that of strain prediction was much lower than the results of the PLS model using structured output (67.78% vs. 80.21%). This may suggest that by using binary coding, the model had focused on separating classes of the “weak” factor (different inducer conditions in this study) and somehow compromised the performance in separating classes of the “strong” factor (strain differences in this study). For binary coding the Y vectors in PLS-DA, it is also possible to train two separate PLS-DA models, one for each factor. Thus, we have also trained two PLS-DA models, one for strain classification and another for inducer classification. The PLS-DA model focused for strain classification and the average CCR was 78.51%, which is a significant improvement compared to a full 20-class model, but still slightly worse than the model using structured output PLS. For the inducer condition model the averaged CCR was 54.9%, which was the worst prediction accuracy for the three types of coding methods investigated.

other

For reasons of brevity we denote the three P. putida strains: DOT-T1E, DOT-T1E-PS28, and DOT-T1E-18 as S1, S2, and S3, respectively; the four propranolol dosages: control, 0.2, 0.4, and 0.6 mg/mL are denoted as D0, D1, D2, and D3; the three monitored time points: 0, 10, and 60 min after exposure are denoted as T0, T1, and T2. The structured output coding employed for this study is illustrated in Figure 2 and a detailed description is given in Section 4.3.1. The PLS models were also validated by using the same double cross-validation procedure used in the riboswitch dataset analysis.

other

The confusion matrices of strain predictions, dosages, and time point predictions are given in Table 5, Table 6 and Table 7. The confusion matrix of strain classification suggested that S3 (DOT-T1E-18) was most different to the other two strains. The predictive accuracy in dosages prediction showed a gradient from control to D3, suggesting that the effect of the drug increased gradually as the concentrations increased. It appeared that there might be more overlap between D2 and D3, although this was rather inconclusive, probably due to the limited number of samples available. Lastly, the predictive accuracy in the time points suggested that T0 and T1 were more overlapped with each other, and this suggested that the effect of the drug may not be very significant at 10 min after exposure, and it became more pronounced at 60 min after exposure. To examine the further effect of the way coding on the final pattern is revealed by the model we also performed another analysis using an evenly spaced time point coding in which T0, T1, and T2 were coded as 0, 3, and 6 (other coding remained the same). The resultant confusion matrix for time prediction is listed in Table 8. It can be seen that a similar pattern still persisted, except that more T0 samples had been assigned to T1. The unbalanced error between T0 and T1 can be explained by the fact that there were much fewer T0 samples than T1 (T0 only had control samples).

other

We also performed a PLS-DA using binary coding and treated each unique combination of strain, dosage, and time as a distinct class. The results did not show any sensible pattern and the predictive accuracies were no better than a pure random classifier (data not shown). This could be caused by a large number of classes, limited number of samples, and a mixture nature of ordinal and categorical data in the experimental design. This, again, highlighted the advantage of using a structured output to analyse data from complex experimental designs.

other

Identifying significant metabolites is also an important aspect in metabolomics studies and PLS models can provide many statistics which can aid in the discovery of important (input) variables which have contributed to the separation between classes. In this study we employed variable importance in projection (VIP) scores for significant metabolites identification.

other

The VIP score plots of the riboswitch dataset are provided in Figure 3. The top 10 most significant metabolites that could be definitive identified according to MSI (variable 21 and 53, although significant, could not be confidently identified through mass spectra matching) across two blocks in the plot were identified as: glycerol (variable 27), leucine (33), l-isoleucine (37), threonine (49), l-aspartic acid (71), glutamine (81), phosphoric acid (103), lysine (119), silanamine (170), and inosine (176). The box-whisker plots that are presented in Supplementary Materials Figures S1–S10 can be used for visualising the patterns of these metabolites. It is easy to see that most of these metabolites are amino acids and this suggests that the amino acid metabolism had been significantly affected by exposure to different inducers and, also, different strains responded to these inducers differently because these amino acids had significant VIP scores in both blocks. This finding is consistent with our previous report . Among them it is perhaps not surprising that the levels of amino acids, such as threonine, are the highest under inducing conditions (IPTG and IPTG + PPDA) in the protein-producing strain (GFP), while being at its lowest in the wild-type strain, as the protein producers demand higher levels of such amino acids as they require them for synthesis of the recombinant protein. However, the silanamine levels displayed the complete opposite trend, as it is at its highest level in the wild-type while being at its lowest level under the inducing conditions in the protein-producer strains. These findings suggested that, as the wild-type strain is not restricted by GFP production, it may catabolise the available amino acids, as carbon and/or nitrogen sources, to support cellular growth, which results in lower amino acid and higher ammonia levels. By contrast, as the amino acid pools in the protein-producer strains are constricted by the demand for recombinant GFP production, under inducing conditions (IPTG, IPTG + PPDA) these strains may direct the central metabolism towards the required amino acid biosynthetic pathways to keep the homeostasis of these metabolites, which may also result in lower ammonia production and, subsequently, lower silanamine levels.

other

The VIP scores plots from PLS-DA analysis on the propranolol dataset are shown in Figure 4. The top 10 most significant and identified metabolites are: alanine (14), butanoic acid (26), leucine (37), serine (54), silanamine (86), phenylalanine (95), ornithine (100), trehalose (188), metoprolol (189), and 5′-adenylic acid (199). The box-whisker plots of alanine, leucine, serine, pheylalanine, and ornithine had already been published in and, therefore, we do not repeat these here. The box-whisker plots of the remaining metabolites are provided in Supplementary Materials Figures S11–S15. It was rather surprising to see that metoprolol had obtained a very significant VIP score and, in fact, is the most significant small molecule compared to all other metabolites. Metoprolol is not a substance of natural occurrence in bacteria, but does have a similar chemical structure to propranolol; we, therefore, suspected that it was converted from the propranolol that had been used to challenge these P. putida. We conducted a GC-MS analysis on the propranolol standard and found out that a metoprolol peak was indeed detected within the standard and had a higher peak area than propranolol itself (Supplementary Materials Figure S16). We then conducted LC-MS on the same standard and could only observe a propranolol peak. Thus, we concluded that the majority of propranolol had undergone chemical conversion to metoprolol either during the derivatization process or in the electro-ionisation source used in the GC-MS analysis. This explained the fact that the VIP score of metoprolol had been the highest in the dosage block. It is also worth noting that the propranolol peak (180) had also obtained a significant VIP score of 5.193, even though it is not one of the top 10 highest scores. Additionally, the box-whisker plot of propranolol showed almost exactly the same pattern as that of metoprolol, but at a lower scale (data not shown). The energy-related metabolites, such as trehalose and 5′-adenylic acid (AMP) were also found to be significant. This could be caused by the energy needed to drive the efflux pumps in these P. putida strains in response to the exposure of the drug. Such a pattern had also been observed in one of our previous studies .

other

In this paper we have demonstrated that PLS can also be used to model structured outputs and provide improved results over classical binary output coding for modelling data from complex DOEs. It is also easy to implement this methodology, as one only needs to design a structured output Y based on the experimental design. After modelling these can then be interpreted by producing a series of confusion matrices to gain insights into the modelled patterns in the data. While it is also possible to inspect PLS scores to visualise the pattern, this is usually not easy for PLS using a structured output as the high complexity in Y usually requires a large number of latent variables (PLS component) to model such complexity sufficiently. Thus, it is not realistic to expect that the overall pattern can always be well represented by, first, a few PLS components, and one may be tempted to plot any latent variables against each other to get the “desired” picture (a practice that is not very objective). Another concern in visualising PLS scores is that it can only present the results of one specific split of the training and test sets while, for robust modelling, it is better to test multiple combinations of training and test sets to get a robust estimation of the errors and prevent getting over optimistic results because of a “lucky” split. Finally, on this point the need to interpret the Y predictions rather than the PLS scores has been highlighted in .

other

The results from the riboswitch data suggested that when there is a mixture of strong and weak factors, using binary output coding may result in a model which is focused on separating the classes of the weak factor and compromise its capability in predicting strong factors. Using a structured output coding has resulted in a more balanced model which might still not be able to improve the results of weak factors, although it can significantly improve the results of the strong factors. The results of the propranolol dataset have shown that when there is a mixture of values both of an ordinal and categorical nature in the experimental design, structured output coding might be required to be able to successfully build a sensible model. Although it is worth mentioning that the currently proposed method does not explicitly code the interactions between factors. Such effects are inferred from examining the conditional confusion matrices. How to code in interaction terms to Y and interpret the results is an interesting open research question. Finally, we have demonstrated that the variable importance statistics in PLS modelling, such as VIP scores, can also be used for significant metabolite identification. This can be considered as a major advantage of PLS compared to more “black-boxed” machine learning techniques, such as S-SVM or neural networks.

other

It is also important to note that compared to those methods which have dedicated structural data modelling, such as S-SVM, PLS has a major limitation in that the model has no flexibility in choosing how to represent the errors, i.e., the difference between coded and predicted outputs. The solution found by PLS models is to maximise the covariance between coded (known target) Y and the observed X while methods like S-SVM allows the user-defined error set to be used directly and optimise the model towards minimising such errors. This means that, for PLS, the scale of the coded output will have an influence on the final solution and the results will favour minimising the error of the columns in Y having larger variance. Therefore, if the structured output has multiple blocks, it is important to ensure that different blocks have comparable variance to prevent the block with the largest variation from dominating the results.

other

Since these two datasets had already been published elsewhere, only brief descriptions are provided here, and more detailed information about the motivation of the experimental design, bacterial characteristics, sample analysis, and biological interpretations can be found in .

other

Five E. coli strains were used in this study, coded as wild, PET, EGFP, iL3PET, and iL3EGFP, a detailed description of these strains can be found in . All strains were streak plated three times on LB agar prior to every experiment to ensure the purity of the stocks. One-hundred milligrams per litre of ampicillin and/or 10 mg/L kanamycin were added to the LB broth/agar as selectable plasmid markers where necessary. Starting inocula were prepared by inoculating 25 mL of LB broth with a single colony of the appropriate strain followed by overnight incubation at 37 °C with 200 rpm shaking in a Multitron standard shaker incubator (INFORS-HT Bottmingen Switzerland). Different inducing conditions examined in this study are described in Table 9.

other

Fifty millilitres of LB broth, three biological replicates per condition, was inoculated with the appropriate strains using the overnight grown cells to a final OD600nm = 0.1, followed by incubation at 37 °C at 200 rpm shaking for 3 h. Upon reaching the OD600nm = 0.5 the samples were exposed to one of the inducing conditions (Table 9), and the incubation temperature was decreased to 20 °C at 200 rpm for 8 h in shaking incubators, which sums up to a total of 11 h of incubation. Fifteen millilitre samples from each flask were quenched using 30 mL, 60% aqueous methanol (−48 °C) following procedures described in previous studies . The extraction protocol was also adapted from with the exception of centrifugation speed being set at 15,871× g. All extracts were normalized according to OD600nm followed by combining 100 µL from each of the samples in a new tube, to be used as the quality control (QC) sample. One-hundred microlitre internal standard solution (0.2 mg/mL succinic-d4 acid, 0.2 mg/mL glycine-d5, 0.2 mg/mL benzoic-d5 acid, and 0.2 mg/mL lysine-d4) was added to all the samples (including QCs) followed by an overnight drying step using a speed vacuum concentrator (Concentrator 5301, Eppendorf, Cambridge, UK).

other

The derivatized samples were randomised and analysed using a Gerstel MPS-2 autosampler (Gerstel, Baltimore, MD, USA) used in conjunction with an Agilent 6890N GC oven (Wokingham, UK) coupled to a Leco Pegasus III mass spectrometer (St. Joseph, MI, USA) following previously published methods . Collected data were deconvolved using Leco ChromaTOF software v3.32 and initial identification was carried out according to metabolomics standards initiative (MSI) guidelines followed by removal of mass spectral features with high deviation within the QC samples . The chromatographic peaks corresponding to PPDA and IPTG were also removed from the data before subjecting the data to PLS to eliminate any variation that might result from the presence of these compounds.

other

Three bacterial strains of P. putida DOT-T1E were used in this study, denoted as DOT-T1E, DOT-T1E-PS28, and DOT-T1E-18; their relevant characteristics, and references for further information on each strain can be found in . All strains were sub-cultured in triplicate to obtain axenic cultures. Individual colonies were then picked and transferred from plates into 250 mL flasks containing 50 mL of autoclaved Lysogeny broth (LyB) medium and incubated at 24 h at 30 °C in an orbital incubator (Infors HT Ltd., Reigate, UK) shaken at 200 rpm.

other

Cells were grown in 50 mL of LyB medium for 5 h at 30 °C and 200 rpm. Once cell cultures reached the mid-exponential phase, samples were divided into two groups. One group was kept as a control and to the second group propranolol was added at three different concentrations (0.2, 0.4, and 0.6 mg/mL). These cultures were then incubated for an additional 8 h. Fifteen millilitre samples were quenched at three time points 0, 10, and 60 min before and after the addition of propranolol (0 min refers to the point immediately before the addition of propranolol). This procedure was performed with four biological replicates.

other

The samples (15 mL) were plunged into a double volume of 60% cold methanol (−50 °C) in a 50 mL tube. The quenched culture mixture was centrifuged (3000× g, 10 min, 1 °C), and then the supernatant was discarded, while the cell pellets were stored at −80 °C until required for metabolite extraction.

other

The biomass pellets were resuspended in 750 μL of freshly prepared cold methanol (80%). The solution was then transferred to a 2 mL Eppendorf microcentrifuge tube. This was followed by a freeze-thaw cycle in order to extract the intracellular polar metabolites from the cells. Samples were centrifuged at (13,500× g, 3 min, 4 °C) and the supernatant was transferred to new tubes and stored on dry ice. The extraction was performed again on the remaining pellet and both supernatants were combined and again stored on dry ice. A final aliquot (1400 μL) of metabolite extracts were normalised using 80% methanol according to OD at 660 nm. A quality control (QC) sample was prepared by transferring 100 μL from each of the samples to a new (15 mL) centrifuge tube. This was followed by the addition of (100 μL) of internal standard solution (0.2 mg/mL glycine-d5, 0.2 mg/mL benzoic-d5 acid, 0.2 mg/mL lysine-d4, and 0.2 mg/mL succinic-d4 acid) to all samples. The samples were lyophilized for 16 h by speed vacuum concentrator (concentrator 5301; Eppendorf, Cambridge, UK), and then the pellet was stored at −80 °C for further analysis.

other

Prior to GC-MS analysis the samples were derivatized using the same method used in riboswitch experiment. These samples were then randomised and analysed by using gas chromatography electron ionisation time-of-flight mass spectrometry (GC-TOF-MS) using an Agilent 6890 GC instrument coupled to a LECO Pegasus III TOF mass spectrometer (Leco, St. Joseph, MI, USA), as described previously . A GC column (VF-17MS column, 0.25 mm ID × 30 m × 0.25 μm film thickness, Varian, cat. No. CP8982) was employed at a constant helium carrier gas flow of 1 mL/min, with a temperature program starts at 70 °C and end at 300 °C. The mass spectrometer source was operated at a temperature of 250 °C in electron ionization (EI) mode, with an electron energy of 70 eV and the detector is operated in the range 1400–1800 V. Raw data processing was undertaken using LECO ChromaTOF v3.26 in order to construct a data matrix of metabolite peak vs. sample and infilled with peak areas for metabolites that were detected. A reference database was prepared that contained retention times, quant mass, peak area, retention index value, and peak number related to each peak by analysing QC samples. The identification of analytes was based on both spectral similarity and matched with retention indices. An in-house library, as well as the NIST library, was used for identification, and we also followed the same MSI guidelines for metabolite identification as for the riboswitch experiment.

other

All of the data analysis were performed using in-house scripts written in MATLAB 2014a (Mathworks, MA, USA) environment and these scripts are freely available online at . Both datasets were firstly aligned using QCs , and the missing values were imputed by using KNN-imputation method and then subjected to PLS analysis.

other

The structured coding for this dataset is relatively straightforward. It can be considered as a classification model that will be calibrated to predict two class membership simultaneously. Therefore, the structure coded output is a combination of two binary matrices, one for bacterial strains and another for inducer conditions, as illustrated in Figure 1. The predicted output was firstly “crisped” so that an unambiguous class membership for each class (strain and inducer condition) could be more readily assigned. This is done by assigning the class membership to the corresponding column which achieved the highest number compare to other columns within the block. The error set was a record of the count of the number of blocks which had been misclassified which varied from 0 (correct in both blocks), 1 (wrong in either strain or inducer condition block), to 2 (wrong in both blocks).

other

The structured coding for propranolol dataset was more complicated as it consisted of three blocks: one classification block (strains) and two ordinal blocks (dosage and time). These three blocks were also at different scales in which the strain block could be presented by a dummy binary matrix in any scale, the dosage varied from 0–0.6 mg/L while the time block varied from 0–60 min. These three blocks need to be properly balanced to avoid one block with the largest variance dominating other “minor” blocks because PLS can only minimise the differences between the predicted and known outputs in least square sense. In this study we employed the following coding scheme: (1) the two discrete numbers in the binary matrix part were coded as “0” and “6”; (2) the four different dosages were coded 0, 0.2, 0.4, and 0.6 mg/L as 0, 2, 4, and 6, respectively; and (3) the three time points were coded as 0, 1, and 6 respectively as illustrated in Figure 2.

other

This coding method ensured that the “worst” predictions in different blocks would result in similar errors, e.g., a “complete” misclassification error in strain prediction, D0 had been predicted as D3 or T2 had been predicted as T0. The coded values in the dosage column were evenly spaced as in the experiment the dosages applied were also evenly spaced while the unevenly spaced intervals in the time column had reflected the real-time differences, but at a different scale, which is comparable with other blocks.

other

The interpretation of the predicted outputs was also performed on three blocks separately, and the classification block was interpreted in the same way as what had been done for the riboswitch dataset. The interpretations on the two ordinal blocks also followed a similar “crisping” of the output. For each sample, the predicted value was compared to all available known values, i.e., 0, 2, 4, and 6, for the dosage column, and 0, 1, and 6 for the time column. The one with the smallest absolute difference was then assigned to the corresponding sample. For example, if a sample had a predicted dosage as 2.5 and time as 0.8, this sample would be assigned as D1 and T1. The reason is that it was not easy to infer the spatial distribution of different types of samples using regression-based indicators, such as root mean squared error (RMSE), R2, or Q2 . Additionally, the number of different points were very limited (four for dosage and only three for time), a plot of predicted vs. known values could not show a clear monotonic changing trend, either. By assigning the “raw” outputs in prediction to the nearest target, a confusion matrix can be calculated and the types of samples with larger misclassification error between each other can be considered as closer-related types, and vice versa.

other

For PLS with multiple columns of outputs, there is a VIP score vector for each column in Y. To simplify the task of inspection the VIP scores were summarised according to the blocks. This is done by taking the maximum of the VIP scores of the Y variables within the group. For the riboswitch dataset, the VIP scores for strain classification were the maximum VIP scores of the first five columns, and those for inducer condition classification were the maximum VIP scores of the last four columns. For the propranolol dataset, the VIP scores for the strain classification were the maximum VIP scores of the first three columns; those for dosage and time modelling were the VIP scores of the fourth and fifth column, respectively.

other

In conclusion, we have demonstrated that it is possible to implement a structured output for modelling metabolomics data when multiple interacting factors are present in the experimental design. We believe that this approach would have general utility in metabolomics data analysis as well as in other areas where the analysis of complex multivariate data is needed.

other

A double cross-validation (CV) procedure was employed to train and validate the PLS models. The split of training and test set was based on biological replicates. For example, in the riboswitch experiment, there were three biological replicates for each of the 5 × 4 = 20 different combinations of the two factors (five strains, and four inducer conditions). This resulted in 20 × 3 = 60 samples in total. For one split of the training and test set, one biological replicate from each unique combination of strain and inducer condition was randomly selected and removed from the data to form a blind test set which had 20 samples. The model was then built on the remaining 40 samples and the optimal number of LVs was chosen by performing a LOOCV on the training set. The CV errors were calculated using the corresponding error set. The final PLS model was then built on the whole training set using the optimal number of PLS components, which had minimal cross-validation error, and this model was then applied to the blind test set to generate a predicted output. This output was then compared to the known structured output, as described before, to calculate the misclassification or misassignment errors. A confusion matrix was then calculated in which each row represents the percentage of a type of samples that had been predicted as one of all the available types. This procedure had been repeated 1000 times, where in each iteration a different training and test set had been randomly chosen as described above. The confusion matrices of these 1000 iterations were averaged to generate the final confusion matrix. Additionally, in order to assess the statistical significance of the effect of each factor, we conducted permutation tests . For each split of the training and test sets, the labels of the samples were randomly permuted and the PLS model was built and validated using the same procedure as above. The predictive error in the test set was then calculated and recorded. An empirical p-value was derived by calculating the ratio of the number of cases when the errors of the models using the known labels (observed errors) had been higher than the ones using randomly-permuted labels (null errors) over the all 1000 iterations. A low p-value (e.g., p < 0.01, meaning that there were less than 10 cases out of 1000 in which the observed errors were higher than the corresponding null errors) would suggest the effect of the factor was statistically significant, while the effect would be considered as insignificant if the p-value was high (e.g., p > 0.05).

other

Mean centre pre-processing was applied for the PLS modelling . On the training set, the means of both X and Y matrices (denoted as x¯ and y¯, respectively) were calculated, recorded, and subtracted from X and Y, respectively. The PLS model was then built between the mean centred X and Y. Then, in the validation or blind test, x¯ was subtracted from X in the validation/test set and the trained PLS model was applied to calculate the predicted Y (denoted as Y^). The final prediction was Y^ with y¯ added back into it.

other

An important property of the cochlea is the ability to “amplify” the mechanical vibrations at the basilar membrane (Dallos, 2008). This process is under the control of the medial olivocochlear (MOC) system via efferent fibers that innervate the outer hair cells. Activation of these efferents, called the MOC reflex (MOCR), hyperpolarizes the outer hair cells (Fuchs, 2002) and decreases the cochlear gain in anesthetized animals (Buno, 1978; Dolan and Nuttall, 1988; Liberman, 1989; Warren and Liberman, 1989; Kawase and Liberman, 1993; Guinan and Stankovic, 1996).

study