Software for increased accuracy prediction of antibody-antigen complexes

Acpharis, Inc., Holliston MA

Investigators

Abstract

Antibodies are important drugs to treat cancer, infectious and cardiovascular diseases, arthritis, inflammation and immune disorders, and are expected to drive further growth of the biotechnology industry. Advances in high- throughput single-cell and VDJ sequencing of B-cell receptor repertoires allow for obtaining large ensembles of antibody sequences relevant to a disease, but experimentally determining the structures of many antibody- antigen (Ab-Ag) complexes and antibody epitopes is very expensive, which calls for computational approaches. However, in spite of the recent progress in machine learning, such methods generally don’t provide the desired accuracy in immunology applications. A team lead by the founders of Acpharis participated in the 2024 CASP16 protein structure prediction contest and demonstrated a breakthrough in the prediction of Ab-Ag structures from sequences, achieving much higher accuracy than any of the other 85 participating groups and even higher than the recently released AlphaFold3 (AF3) program. The key to this breakthrough is the integration of machine learning with physics based sampling implemented as a modified version of the PIPER program developed by the Vajda lab and licensed to Acpharis by Boston University. PIPER is a rigid body docking program, which systematically samples the conformational space of the complex using the fast Fourier transform approach, evaluating the energy for billions of conformations. The scoring function includes van der Waals energy terms, electrostatics energy, and an antibody-antigen specific pairwise statistical potential. Acpharis has already developed a GPU version of PIPER, and a custom version of AlphaFold2 (AF2) using the PyTorch library. Aim 1 of this proposal is to further test and optimize this protocol and to develop a commercial quality software product implementing the algorithm. Testing will involve the most recent public Ab-Ag benchmark sets. In Aim 2 we plan to enhance the accuracy of the AF2 component of the protocol for predicting the conformations of the Complementarity Determining Region, also called the hypervariable loops, by fine-tuning the ML based program with a specialized dataset of high-resolution Ab-Ag crystal structures. This tailored approach aimed to address the generalist nature of AF2 original training, which, while broad-ranging, lacked the granularity necessary for the high-precision demands Ab-Ag interaction prediction. Since the goal is improving the prediction of the frequently flexible hypervariable loops, the training will be based on a benchmark set of Ab-Ag complex structures rather than the structures of separately determined antibodies. In Aim 3 the improved Ab-Ag docking will be utilized to increase the accuracy of the prediction of Ab epitopes currently implemented as the AbEMap server. AbEMap is based on generating a large ensemble of docked Ab-Ag complex structures to determine the contact residues in each complex. For each antigen residue, the energy weighted average of the number of contacts yields a likelihood score of being part of the epitope. The updated AbEMap program will also be developed into a professional quality software product by Acpharis.