Machine Learning Models for Protein Structure Prediction

Abstract

Predicting three-dimensional protein structures from amino acid sequences has been a grand challenge in computational biology for over fifty years, and the recent success of deep learning approaches has transformed both the accuracy and the speed of structure prediction, yet the adoption of these models into routine structural biology workflows varies across method categories and application contexts. We evaluated 206 machine learning protein structure prediction programmes active across computational biology centres affiliated with Advanced Computing University in Paris between 2016 and 2021, spanning five method categories: end-to-end deep learning predictors, co-evolutionary contact-based methods, physics-informed hybrid models, fragment assembly with ML scoring, and homology modelling enhanced by neural networks. A Protein Structure Prediction Quality Index (PSPQI) was constructed from five sub-scores -- backbone accuracy, side-chain packing quality, prediction confidence calibration, computational efficiency, and experimental validation rate -- with weights from regression against adoption into active structural biology research pipelines. PSPQI correlated with pipeline adoption at r = +0.84 and discriminated adopted from non-adopted methods with an AUC of 0.886. End-to-end deep learning predictors scored highest (mean PSPQI 0.828), while fragment assembly methods trailed at 0.596. Only 35.4 percent of programmes exceeded the 0.75 threshold. Backbone accuracy carried the largest regression weight (beta = +0.282), followed by prediction confidence calibration (beta = +0.228).