PIPENN: protein interface prediction from sequence with an ensemble of neural nets

Abstract

Motivation The interactions between proteins and other molecules are essential to many biological and cellular processes. Experimental identification of interface residues is a time-consuming, costly, and challenging task, while protein sequence data is ubiquitous. Consequently, many computational and machine learning approaches have been developed over the years to predict such interface residues from sequence. However, the effectiveness of different deep learning architectures and learning strategies for protein–protein, protein–nucleotide, and protein–small molecule interface prediction, has not yet been investigated in great detail. Therefore, we here explore the prediction of protein interface residues using six deep learning architectures and various learning strategies with sequence-derived input features. Results We constructed a large data set dubbed BioDL, comprising protein–protein interactions from the PDB, and DNA/RNA and small molecule interactions from the BioLip database. We also constructed six DL architectures, and evaluated them on the BioDL benchmarks. This shows that no single architecture performs best on all instances. An ensemble architecture, which combines all six architectures, does consistently achieve peak prediction accuracy. We confirmed these results on the published benchmark set by Zhang & Kurgan (ZK448), and on our own existing curated homo- and heteromeric protein interaction data set. Our PIPENN sequence-based ensemble predictor outperforms current state-of-the-art sequence-based protein interface predictors on ZK448 on all interaction types, achieving an AUC-ROC of 0.718 for protein–protein, 0.823 for protein–nucleotide and 0.842 for protein–small molecule. Availability Source code and data sets at https://github.com/ibivu/pipenn/

Multi-task learning to leverage partially annotated data for PPI interface prediction

“…Heffernan et al 21 used LSTM bidirectional RNNs and showed that this method is useful to capture long range interactions, especially for residues with large numbers of long-range contacts. We recently compared the usage of different neural network architectures for the prediction of protein interfaces 23 . Furthermore, transformers have been successfully used in the language of proteins 24 , 25 .…”

“…In this study, we predict PPI interface residues based on the primary sequence in a partner-unspecific way. The most recent other sequence-based partner-unspecific models are SSWRF 30 , SeRenDIP 7 , 31 , SCRIBER 9 , and PIPENN 23 . The SSWRF method uses an ensemble support vector machine and a sample-weighed random forest to predict the PPI interface 30 .…”

“…However, this is not entirely trivial as one needs to define the interface of the of the binding molecules. Typically, some threshold is used to select amino acids in close proximity 9 , 23 . Some larger published datasets are available, notably ‘ZK448’, a 448 protein test-set by Zhang and Kurgan 9 , ‘BioDL’ containing in total 4620 proteins with PPI annotations by Stringer et al 23 , and ‘Homomeric & Heteromeric’ with 546 proteins by Hou et al 7 , 31 , which each come with their sets of precalculated features.…”

“…Typically, some threshold is used to select amino acids in close proximity 9 , 23 . Some larger published datasets are available, notably ‘ZK448’, a 448 protein test-set by Zhang and Kurgan 9 , ‘BioDL’ containing in total 4620 proteins with PPI annotations by Stringer et al 23 , and ‘Homomeric & Heteromeric’ with 546 proteins by Hou et al 7 , 31 , which each come with their sets of precalculated features. However, not all datasets include all features, and generation of missing features may be a time-consuming task.…”

“…However, not all datasets include all features, and generation of missing features may be a time-consuming task. State of the art performances for partner-unspecific PPI interface prediction range between 0.68 and 0.78 AUC ROC, depending on the exact dataset and model used 23 . Some types of PPIs are more difficult to predict than others: heteromeric interfaces tend to be more difficult than homomeric interfaces 7 .…”