Geometric Deep Learning Enables Protein Structure Prediction at Atomic Precision

A research team from the Max Planck Institute for Intelligent Systems has developed a geometric deep learning architecture that achieves sub-angstrom accuracy in protein structure prediction, representing a significant advance over existing methods including AlphaFold 3.

The architecture, called EquiFormer-X, leverages SE(3)-equivariant neural networks—networks that respect the rotational and translational symmetries of three-dimensional molecular structures. This geometric inductive bias allows the model to learn more efficiently from limited training data and generalize better to novel protein families.

On the CASP15 benchmark, EquiFormer-X achieved a median TM-score of 0.97 and a median RMSD of 0.8 Angstroms, surpassing all previous methods. Particularly notable was its performance on intrinsically disordered proteins and protein complexes, which have historically been challenging for structure prediction methods.

The implications for drug discovery are significant: more accurate structure prediction enables more precise identification of drug binding sites and more reliable prediction of how small molecule drugs will interact with target proteins. Several pharmaceutical companies have already begun integrating the method into their computational drug discovery pipelines.