[Paper Review] WideDTA: prediction of drug-target binding affinity

Hakime Öztürk, Elif Özkırımlı+1 more|arXiv (Cornell University)|Feb 4, 2019

Computational Drug Discovery Methods9 references152 citations

TL;DR

WideDTA uses word-based representations from protein sequences, ligand SMILES, PROSITE domains/motifs, and ligand MCS to predict drug-target binding affinity, outperforming DeepDTA on benchmark datasets. Domain/motif and MCS add limited benefit on kinase-focused data.

ABSTRACT

Motivation: Prediction of the interaction affinity between proteins and compounds is a major challenge in the drug discovery process. WideDTA is a deep-learning based prediction model that employs chemical and biological textual sequence information to predict binding affinity. Results: WideDTA uses four text-based information sources, namely the protein sequence, ligand SMILES, protein domains and motifs, and maximum common substructure words to predict binding affinity. WideDTA outperformed one of the state of the art deep learning methods for drug-target binding affinity prediction, DeepDTA on the KIBA dataset with a statistical significance. This indicates that the word-based sequence representation adapted by WideDTA is a promising alternative to the character-based sequence representation approach in deep learning models for binding affinity prediction, such as the one used in DeepDTA. In addition, the results showed that, given the protein sequence and ligand SMILES, the inclusion of protein domain and motif information as well as ligand maximum common substructure words do not provide additional useful information for the deep learning model. Interestingly, however, using only domain and motif information to represent proteins achieved similar performance to using the full protein sequence, suggesting that important binding relevant information is contained within the protein motifs and domains.

Motivation & Objective

• Predict protein-ligand binding affinities using text-based representations of proteins and ligands.
• Assess whether adding domain/motif and maximum common substructure information improves prediction.
• Compare word-based WideDTA to prior character-based models and traditional methods on Davis and KIBA datasets.

Proposed method

• Represent protein sequences as 3-residue words (PS).
• Represent ligand SMILES as 8-character words using a sliding window (LS).
• Extract protein domains/motifs from PROSITE and represent them as 3-residue words (PDM).
• Extract ligand maximum common substructures and represent them as words (LMCS).
• Process each information source with two 1D CNN layers plus max-pooling to obtain features, then concatenate and pass through three dense layers with dropout.
• Train and evaluate on Davis and KIBA datasets using Concordance Index (CI), MSE, and Pearson correlation; compare to KronRLS, SimBoost, and DeepDTA.

Experimental results

Research questions

• RQ1 Does a word-based representation of proteins and ligands improve binding affinity prediction compared with character-based methods?
• RQ2 Do domain/motif information (PDM) and LMCS words provide additional predictive value beyond full protein sequence and LS/LMCS alone?
• RQ3 What is the performance of WideDTA versus state-of-the-art methods on the Davis and KIBA benchmarks?

Key findings

• WideDTA with all four modules achieves the best Davis results: CI 0.886 and MSE 0.262.
• On KIBA, the best WideDTA setup achieves CI 0.875 and MSE 0.179.
• Using PS + LS alone already outperforms DeepDTA on both datasets (Davis CI 0.874; KIBA CI 0.874).
• In Davis and KIBA, adding PDM did not significantly improve performance; PDM alone performed similarly to full sequences in some cases.
• LMCS provided marginal gains for Davis but was less favorable on KIBA compared to LS.
• Compared to KronRLS and SimBoost, WideDTA variants with LS/PS (and PDM/LMCS combinations) consistently perform better; on Davis, best WideDTA CI 0.886 vs KronRLS 0.871 and SimBoost 0.872; on KIBA, best WideDTA CI 0.875 vs DeepDTA 0.863.