Using AlphaFold2 for Antibody-Antigen Structure Re-Scoring

Estimated Reading Time: 4 minutes

TL;DR

This week we briefly review a recent article on applying AlphaFold for antibody-antigen modeling.

Antibody-Antigen Structure Prediction from Molecular Docking Using a Deep Learning Model

Antibodies are an increasingly important class of therapies. Improved understanding of how antibodies bind to antigens would enable us to more effectively design novel antibody therapeutics. For this reason, considerable effort has gone into computational methods to model antibody-antigen binding. Classical physics based approaches, such as docking, attempt to find a binding pose (that is, a configuration of the antibody and antigen) that minimizes the binding energy. Newer machine learning approaches such as AlphaFold2 attempt to directly sample antibody-antigen binding poses using a data-driven approach rather than a physical energy minimization approach.  

Physics-based docking techniques tend to sample higher quality antibody-antigen structures than AlphaFold2 due to the lack of co-evolutionary constraints to guide the machine learning algorithm. AlphaFold2’s success in other protein folding applications depends in large part on co-evolutionary information from multiple-sequence alignment, which matches novel protein sequences against known template sequences, but this technique doesn't apply to antibody-antigen complexes whose binding mechanism is driven by somatic hypermutation and affinity maturation.

A recent paper explores the idea of combining physics-based and ML-based methods by using AlphaFold2 to predict the binding affinity of antibody-antigen structures sampled by a classical docking algorithm. The underlying reasoning is that although physics-based methods may sample better structures, their predicted energies are often miscalibrated. Machine learning methods often succeed at calibrating tightly, even if they don’t learn the full physics of the system. AlphaFold2 is consequently used to compute a learned binding score on top of a docking-sampled structure to better differentiate binding vs non-binding antibody-antigen pairs.

The method creates sets of antibody-antigen docking solutions consisting of positives (resembling known crystal structures generated by X-ray crystallography) and negatives (low energy structures similar to known crystal structures), and re-assesses them using AlphaFold2 with a novel composite score. This AF2_Composite scoring system is the sum of two metrics, normalized pLDDT and pTMscore (these are both metrics computed by AlphaFold2). The approach is independent of choice of specific docking method, making it a potential broadly applicable method for improving antibody-antigen structure predictions.

The ability of AlphaFold2 to improve the accuracy of docking predictions is examined using decoy sets generated from 3 complementary docking methods, ProPose, ZDOCK and PIPER. AlphaFold2 successfully distinguishes near-native structures from challenging decoys, particularly in case of “unbound-backbone” structures in which the entire antibody conformation changes during binding. Such structures are more relevant for real-world applications. However, as the figure below shows, AlphaFold2 does not provide across-the-board improvements, but may still prove useful in enhancing the accuracy of antibody-antigen structure predictions even when co-evolutionary data is unavailable. That said for the time being, physics-based methods still remain crucial for sampling realistic antibody-antigen structures.

Success rates for various docking methods followed by AlphaFold2 rescoring as a function of the number of top-ranked models. (A) is for structures released before AlphaFold2 and (B) for structures released after AlphaFold2.  AlphaFold can help improve early success rates especially for unbound-backbone structures but the improvement is not across the board.

Interesting Links from Around the Web

• https://spectrum.ieee.org/aneutronic-fusion: A great review of a number of competing approaches to practical fusion.

• https://spectrum.ieee.org/solar-powered-cars: A discussion of the prospects for solar powered cars.

• https://www.nature.com/articles/d41586-023-03183-3: Drugs like Ozempic are changing how obesity is treated, but there are still open questions on potential side effects that need to be investigated further.

• https://www.quantamagazine.org/echoes-of-electromagnetism-found-in-number-theory-20231012/: Neat mathematical links between mathematical physics and number theory

• https://www.science.org/content/blog-post/target-based-drug-discovery-waste-time: A thought-provoking argument that target-based drug discovery is much less potent a strategy than previously thought.

Feedback and Comments

Please feel free to email me directly (bharath@deepforestsci.com) with your feedback and comments! 

About

Deep Into the Forest is a newsletter by Deep Forest Sciences, Inc. We’re a deep tech R&D company building Chiron, an AI-powered scientific discovery engine for the biotech/pharma industries. Deep Forest Sciences leads the development of the open source DeepChem ecosystem. Partner with us to apply our foundational AI technologies to hard real-world problems in drug discovery. Get in touch with us at partnerships@deepforestsci.com! 

Credits

Author: Bharath Ramsundar, Ph.D.

Editor: Sandya Subramanian, Ph.D.

Research and Writing: Rida Irfan