Deep learning based protein-protein interaction prediction using universal protein sequence representations

Abstract

Protein-Protein Interactions (PPI) govern key biological events in the cell and serve as a basis for understanding disease mechanisms and developing treatments. Currently used PPI predictive methods that rely on information from multiple sequence alignments are ineffec-tive on proteins with few known homologs. Recent advances in self-supervised learning per-mit extracting complex features directly from the protein sequence (sequence embeddings) for later use with predictive algorithms. In this thesis, several sequence embedding methods were used in combination with Siamese deep neural network-based classifier architecture for PPI prediction. An average AUROC score of 0.70 on C1 test set suggests that more complex embedding methods such as UniRep and PLUS-RNN are able to extract more in-formation relevant to PPI prediction from the protein sequence. Performance of all methods dropped markedly for C2 and C3 test sets, 0.62 for UniRep and 0.58 for PLUS-RNN, sug-gesting that further improvements are necessary to develop models that are more general in their coverage of the protein sequence space. The results of this work confirm that using pre-trained protein representations with deep learning based classifiers is a viable approach to PPI prediction from sequence alone. In Estonian: Proteiin-proteiini vastastiktoimed (PPI) juhivad olulisi bioloogilisi etappe rakus ning on aluseks haigusmehhanismide mõistmisel ja ravimite tootmisel. Hetkel kasutusel olevad PPI ennustamise meetmed, mis sõltuvad mitme järjestuse joondamise teabest, on ebatõhusad proteiinidel, millel on vähe kaardistatud homolooge. Viimased edusammud iseenesliku õppimise vallas lubavad eraldada keerulisi eripärasusi otse proteiini sekventsist (sekventsi kodeerimine), et neid hiljem ennustavate algoritmidega rakendada. Selle lõputöö käigus kasutati mitmeid sekventsi kodeerimise meetodeid koos Siiami sügava närvivõrgu põhise PPI ennustamise algoritmiga. Keskmine AUROCi skoor 0.70 C1 testandmestikus viitab, et keerulisemad kodeerimise meetodid nagu UniRep ja PLUS-RNN, suudavad proteiini sekventsist rohkem PPI ennustamisele asjakohast informatsiooni eraldada. Kõigi meetodite täpsus langes märkimisväärselt C2 ja C3 testandmestikes, 0.62 UniRepi ja 0.58 PLUS-RNNi puhul. See näitab, et üldisema kattuvusega proteiini sekventsi mudelite arendamiseks on vaja teha edasisi täiendusi. Selle töö tulemused tõestavad, et eeltreenitud proteiini kujutiste kasutamine koos sügavõppel põhinevate klassifitseerijatega, on võimalik lähenemine PPI ennustamisele ainult sekventsi põhjal.