ASAP: A Machine-Learning Framework for Local Protein Properties

Abstract

Determining residue level protein properties, such as the sites for post-translational modifications (PTMs) are vital to understanding proteins at all levels of function. Experimental methods are costly and time-consuming, thus high confidence predictions become essential for functional knowledge at a genomic scale. Traditional computational methods based on strict rules (e.g. regular expressions) fail to annotate sites that lack substantial similarity. Thus, Machine Learning (ML) methods become fundamental in annotating proteins with unknown function. We present ASAP (Amino-acid Sequence Annotation Prediction), a universal ML framework for residue-level predictions. ASAP extracts efficiently and fast large set of window-based features from raw sequences. The platform also supports easy integration of external features such as secondary structure or PSSM profiles. The features are then combined to train underlying ML classifiers. We present a detailed case study for ASAP that was used to train CleavePred, a state-of-the-art protein precursor cleavage sites predictor. Protein cleavage is a fundamental PTM shared by a wide variety of protein groups with minimal sequence similarity. Current computational methods have high false positive rates, making them suboptimal for this task. CleavePred has a simple Python API, and is freely accessible via a web-based application. The high performance of ASAP toward the task of precursor cleavage is suited for analyzing new proteomes at a genomic scale. The tool is attractive to protein design, mass spectrometry search engines and the discovery of new peptide hormones. In summary, we illustrate ASAP as an entry point for predicting PTMs. The approach and flexibility of the platform can easily be extended for additional residue specific tasks. ASAP and CleavePred source code available at https://github.com/ddofer/asap.