Deep Learning Models for Protein Structure Prediction: A Comprehensive Survey

Abstract

A major challenge in computational biology is determining the 3D structure of proteins from their 1D amino acid sequence. Accurate prediction of protein structures improves the understanding of protein function and facilitates applications in drug discovery, protein engineering, and structural biology. Modeling proteins using template-based approaches is difficult when homologous structures are unavailable in the Protein Data Bank (PDB). Recent advances in deep learning (DL) have significantly improved the accuracy of protein structure prediction (PSP). The current approaches employ convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph neural networks (GNNs), and Transformer architectures to learn complex sequence patterns, evolutionary relationships, and long-range residue interactions that govern protein folding. By predicting long-range residue contacts, inter-residue distances, and geometric constraints, these approaches enable the generation of highly accurate three-dimensional protein structures. DL has transformed PSP, with models such as AlphaFold2, AlphaFold3, Boltz-1, and other recent architectures achieving near-experimental accuracy on many CASP targets. The Critical Assessment of Protein Structure Prediction (CASP) is the standard biennial community-wide benchmark for evaluating PSP methods. One of the biggest challenges remaining in this area is modeling regions of proteins that are intrinsically disordered and capturing the full range of protein dynamics and the relationships of different protein domains that exist in different proteins. This survey provides a comprehensive review of recent DL-based PSP methods, analyzes their architectural evolution, compares their strengths and limitations, and discusses emerging research challenges and future directions.