Machine learning solutions for predicting protein–protein interactions

Proteins are social molecules. Recent experimental evidence supports the notion that large protein aggregates, known as biomolecular condensates, affect structurally and functionally many biological processes. Condensate formation may be permanent and/or time dependent, suggesting that biological processes can occur locally, depending on the cell needs. The question then arises as to which extent we can monitor protein-aggregate formation, both experimentally and theoretically and then predict/simulate functional aggregate formation. Available data are relative to mesoscopic interacting networks at a proteome level, to protein-binding affinity data, and to interacting protein complexes, solved with atomic resolution. Powerful algorithms based on machine learning (ML) can extract information from data sets and infer properties of never-seen-before examples. ML tools address the problem of protein–protein interactions (PPIs) adopting different data sets, input features, and architectures. According to recent publications, deep learning is the most successful method. However, in ML-computational biology, convincing evidence of a success story comes out by performing general benchmarks on blind datasets. Results indicate that the state-of-the-art ML approaches, based on traditional and/or deep learning, can still be ameliorated, irrespectively of the power of the method and richness in input features. This being the case, it is quite evident that powerful methods still are not trained on the whole possible spectrum of PPIs and that more investigations are necessary to complete our knowledge of PPI-functional interactions