Machine-learning in building bioinformatics databases for infectious diseases Victor Tong Institute for Infocomm Research A*STAR, Singapore ASEAN-China International Bioinformatics Workshop Apr 2008
Overview Definitions and background Architectures of existing immunological databases Machine-learning for biological databases Conclusion
Biology produces more data than we can process >3000 HLA alleles different T-cell receptors linear 9mer epitopes Post-translational spliced epitopes Data are stored in databases, literature, laboratory records, clinical records, … A major issue: turning data into knowledge The information centric world
Impractical to do manual curation ≥ 16 million PubMed abstracts ~80K immunology related references Large amounts of data that are difficult to interpret Protein-protein interaction extraction from text Bioinformatics: systematic construction and updating of databases Use of bioinformatics
Ad hoc bioinformatics Biological system Computational analysis Biological interpretation
More systematic use of bioinformatics Biological system Computational analysis Biological interpretation Formal description Mathematical problem Conversion of results
Knowledge discovery from databases is the process of automated extraction of useful information or knowledge from individual or multiple databases
1) Data explosion Current databases: Volume of data increasing exponentially GenBank, SWISS-PROT, IMGT, PubMed, etc New databases: Growth in numbers Increase in size More complex Biologists: Maintain personal data bank Information relevant to their research Define objectives for data mining and analysis
2) Data quality Nature of biological data: Fuzzy and complex Varying interpretations Problems with raw data: Inconsistent Inaccurate Redundant Irrelevant Incomplete Incorrect Data cleaning: Limit on the percentage error that can be tolerated in the data Prevent propagation of errors to our databases Prevent depreciation of data quality
3) Database creation and maintenance Software tools and programming efforts: Data collection Constructing databases Integrating data mining tools Updating the databases Nature of the databases: Short lifespan Hard to maintain
4) Data integration Disparities in data sources: Data structures Data formats Views Search mechanisms Location
Overview Definitions and background Architectures of existing immunological databases Machine-learning for biological databases Conclusion
Web-resources for immune epitope information Immune Epitope Database and Analysis Resource (IEDB) Contains B-cell epitopes, T-cell epitopes, MHC ligands for humans, non- human primates, rodents, and other animal species. URL: The international ImMunoGeneTics information system (IMGT) Specializes in Ig, T-cell receptors, MHC, Ig superfamily, MHC superfamily, and related proteins of the immune system of human and other vertebrate species URL: SYFPEITHI Contains ~3,500 T-cell epitopes, MHC ligands and peptide motifs for humans and rodents URL:
Web-resources for immune epitope information MHCBN Contains T-cell epitopes, TAP ligands, MHC binding peptides and MHC non-binding peptides for humans and rodents URL: MPID-T Contains 3D structural information of 187 T-cell receptors, MHCs and interacting epitopes for humans and rodents, spanning 40 alleles URL: AntiJen/JenPep Contains T-cell epitopes, MHC ligands, TAP ligands and B-cell epitopes. URL:
The IEDB class diagram
Relationships between an epitope & contexts
Overview Definitions and background Architectures of existing immunological databases Machine-learning for biological databases Conclusion
Naϊve Bayes classifiers Attribute values are conditionally independent given the target value Goal: to assign a new instance v j the most probable target value V target given a set of attribute values The target class may be defined as: V target = argmax P ( v j ) Π P ( a i | v j )
Comparison of popular text classification algorithms Dataset 20,910 PubMed abstracts 181,299 unique words AROC NBC: ANN: SVM: DT: Wang et al., BMC Bioinformatics 2007, 8:269
Feature selection (FS) Data source PubMed abstracts Medical Subject Headings (MeSH) - National Library of Medicine's controlled vocabulary used for indexing articles, for cataloging books and other holdings Publication title Author(s) etc
Feature selection (FS) Algorithms Document frequency (DF) – ranks features based on the number of abstracts they appear in Information gain (IG) – measures the number of bits of information obtained for category prediction based on their occurrence in a document IG(u) = -∑ P(ci) log P(ci) + P(u) ∑ P(ci|u) log P(ci|u) + P(t) ∑ P(ci|ū) log P(ci|ū) where u is the feature of interest, ci (i = 1, …, m) denotes the set of categories the documents belong to
Feature condensation (FC) Stemming To reduce words to their common root e.g. “binding, binds, bind” to bind Porter stemmer – A ROC = to A ROC = Domain specific vocabulary may be reduced to unsuitable terms
Feature extraction (FE) Rules to capture immune related expressions and group them together Reduction of feature space (i.e. no. of unique words) Enrichment of information content Better performance?
Feature extraction (FE) Examples: Sequence length – identify sequence length and replace with “~range 50~” if sequences to be mapped stretches 50 amino acids MHC alleles – identify MHC alleles and replace with “~mhc_allele~” Protein sequences – identify sequences as a) exclusively containing characters representing the 20 aa, b) in upper case, length > threshold, and replace with “~sequence~”
Performance comparison Wang et al., BMC Bioinformatics 2007, 8:269
Overview Definitions and background Architectures of existing immunological databases Machine-learning for biological databases Conclusion
Machine-learning algorithms enable systematic approach to database construction and facilitates scientific discovery It must be performed with due care and must be scientifically and technically sound