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

Abstract

Protein protein interactions (PPI) are crucial for protein functioning, nevertheless predicting residues in PPI interfaces from the protein sequence remains a challenging problem. In addition, structure-based functional annotations, such as the PPI interface annotations, are scarce: only for about one-third of all protein structures residue-based PPI interface annotations are available. If we want to use a deep learning strategy, we have to overcome the problem of limited data availability. Here we use a multi-task learning strategy that can handle missing data. We start with the multi-task model architecture, and adapted it to carefully handle missing data in the cost function. As related learning tasks we include prediction of secondary structure, solvent accessibility, and buried residue. Our results show that the multi-task learning strategy significantly outperforms single task approaches. Moreover, only the multi-task strategy is able to effectively learn over a dataset extended with structural feature data, without additional PPI annotations. The multi-task setup becomes even more important, if the fraction of PPI annotations becomes very small: the multi-task learner trained on only one-eighth of the PPI annotations—with data extension—reaches the same performances as the single-task learner on all PPI annotations. Thus, we show that the multi-task learning strategy can be beneficial for a small training dataset where the protein’s functional properties of interest are only partially annotated.

Ten quick tips for sequence-based prediction of protein properties using machine learning

“…Firstly, there isn’t always a clear reason for a head-on comparison with other methods. You may, for example, be setting out to find the added value (or not) of specific parts of your training procedure (e.g., [ 15 ]) or of the architecture (e.g., [ 16 , 17 ]). First point of business will be to identify the current state of the art, which you can usually find in a recent benchmarking review.…”

“…We also introduced a broader benchmark set ProteinGLUE including mutiple prediction tasks: secondary structure, solvent accessibility, PPI, epitopes, and hydrophobic patch prediction [ 16 ]. Many method papers will also include an update of latest developments (e.g., [ 15 , 44 ]).…”

“…Predicting protein functional properties is still one of the most important tasks for bioinformaticians (e.g., [10][11][12][13][14][15][16][17]). Here, we collect 10 useful tips or guidelines representing best practices specifically for methods that generate predictions of protein functional structural properties using protein sequence data as input; Fig 1 illustrates several examples.…”

Collectively encoding protein properties enriches protein language models

“…Generally, an MTL model can be trained by linearly combining loss functions from different tasks into a single total loss function [15]. In this way, the model can learn a shared representation for all tasks by stochastic gradient descent (SGD) with back-propagation [15,43]. Ordinarily, assuming that there are M tasks in all, the global loss function can be defined as where L i represents task-specific loss function, and w i denotes weights assigned for each L i .…”

Comparative Study on Feature Selection Methods for Protein

“…Many techniques for protein structure prediction have been intensively studied in recent years, and scientists have developed an increasing number of creative models to boost prediction performance. Database annotation and sequence-based approaches are the two main approaches used in this area [2]. In order to make predictions, sequence-based attempts to extract unique features from protein sequences.…”