PROTEIN DOMAIN PREDICTION USING CONTEXT STATISTICS, THE FALSE DISCOVERY RATE, AND COMPARATIVE GENOMICS, WITH APPLICATION TO PLASMODIUM FALCIPARUM

Abstract

Protein domain prediction is a key step in characterizing protein structure, function, and evolution. However, the genome of Plasmodium falciparum, the causative agent of malignant malaria, remains poorly-annotated using standard approaches to domain prediction. We identified several general problems with domain prediction and developed novel solutions. Firstly, existing domain recognition methods typically evaluate each domain prediction independently of the rest. However, the majority of proteins are multidomain, and pairwise domain co-occurrences are highly specific and non-transitive. Our novel solution, Domain Prediction Using Context (dPUC), successfully harnesses domain co-occurrence to predict weak domains that appear in previously observed combinations, while penalizing domain combinations that have never been observed. Secondly, although the E-value has been the dominant statistic for protein sequence analysis for the past two decades, domain prediction is a multiple hypothesis testing problem and lends itself to the use of q-values, which control the false discovery rate. We develop the first q-value algorithm for domain prediction and introduce several techniques to effectively address the challenges that arise in this application. Follow-up work introduces methods that combine the strengths of dPUC and q-values. All approaches are carefully benchmarked and greatly improve domain prediction compared to standard methods over individual organisms and on a large unbiased and non-redundant database of proteins. Thirdly, domain databases remain incomplete and unsupervised novel domain discovery remains a challenge. We introduce a novel domain discovery framework that takes advantage of advances in remote homology detection using HMM-HMM comparisons, as well as the wealth of information present in the genomes of P. falciparum and 14 of its closest relatives. Our sensitive comparative genomics approach yields 284 novel domain families that greatly increase coverage of proteins that were previously unannotated. In conclusion, our collection of novel techniques are freely available as source code, and are broadly applicable. We have demonstrated their utility in improving our knowledge of one of the most diverged and medically-relevant parasites.