Active learning for affinity prediction of antibodies

Alexandra Gessner, Sebastian W. Ober, Owen Vickery, Dino Oglic, Talip Uçar

Abstract

The primary objective of most lead optimization campaigns is to enhance the binding affinity of ligands. For large molecules such as antibodies, identifying mutations that enhance antibody affinity is particularly challenging due to the combinatorial explosion of potential mutations. When the structure of the antibody-antigen complex is available, relative binding free energy (RBFE) methods can offer valuable insights into how different mutations will impact the potency and selectivity of a drug candidate, thereby reducing the reliance on costly and time-consuming wet-lab experiments. However, accurately simulating the physics of large molecules is computationally intensive. We present an active learning framework that iteratively proposes promising sequences for simulators to evaluate, thereby accelerating the search for improved binders. We explore different modeling approaches to identify the most effective surrogate model for this task, and evaluate our framework both using pre-computed pools of data and in a realistic full-loop setting.

Introduction

In drug discovery, lead optimization is a critical phase where candidate molecules, such as antibodies, are refined to improve their binding affinity to a target antigen. This process typically involves modifying amino acids within the complementarity-determining regions (CDRs). However, the search space for these mutations is vast; even a small number of simultaneous mutations leads to a combinatorial explosion of possible sequences, making exhaustive wet-lab testing impossible.

To address this, researchers utilize "dry lab" simulations like Relative Binding Free Energy (RBFE) to predict binding affinity ($\Delta\Delta G$). While these simulations are physically grounded and accurate, they are computationally expensive, often requiring up to 24 hours of GPU time per mutation. The central challenge is to navigate this immense sequence space efficiently, identifying high-affinity binders while minimizing the number of costly simulations.

Methods

We propose an active learning framework based on Bayesian Optimization (BO) to intelligently search the antibody sequence space. The framework operates in an iterative loop consisting of four main components:

  1. Sequence Encoding: Protein sequences are converted into numerical vectors. We evaluated several strategies, including simple one-hot encoding, "bag of amino acids" (counting 5-grams), a novel encoding based on the BLOSUM substitution matrix, and AbLang2, a specialized protein language model.
  2. Surrogate Modeling: A Gaussian Process (GP) is employed to model the relationship between the encoded sequence and its binding affinity. GPs are particularly useful as they provide both a point prediction and a measure of uncertainty.
  3. Acquisition Function: An acquisition function (e.g., Expected Improvement) balances exploration of uncertain regions with exploitation of known high-affinity areas. To handle the discrete nature of the sequence space, a Genetic Algorithm (GA) is used to find the sequence that maximizes this function.
  4. Simulation: The selected sequence is evaluated using a high-fidelity physics simulator (such as NQFEP or Schrödinger Res Scan), and the result is used to update the surrogate model for the next iteration.
The proposed active learning framework for iterative antibody sequence optimization.

Results and Insights

The framework was validated using two datasets: a high-accuracy NQFEP dataset (532 mutations) and a larger, faster Schrödinger Res Scan dataset (~60,000 mutations).

Surrogate Model Performance

A key finding was the superior performance of the Tanimoto kernel when paired with sequence embeddings, particularly AbLang2. While RBF and Matérn kernels are more common in machine learning, the Tanimoto kernel—traditionally used for molecular fingerprints—proved remarkably robust for protein sequence data. Furthermore, the AbLang2 language model consistently outperformed simpler encoding methods in validation tasks, suggesting that pre-training on large protein datasets provides a valuable prior for biological plausibility.

Optimization Efficiency

In validation experiments using the NQFEP dataset, the combination of AbLang2 encoding and the Tanimoto kernel identified the absolute best binder within just 30 iterations. In "full-loop" experiments, where the model could propose any possible mutation, the framework discovered antibodies with predicted $\Delta\Delta G$ values significantly better than any sequence present in the initial 60,000-mutation dataset. This demonstrates that the framework is capable of true optimization rather than just searching a pre-defined list.

Comparison of optimization performance (best ddG found) across different sequence encodings and iterations.

Discussion and Future Work

While the current framework shows significant promise, several areas for improvement remain. The cubic computational cost of Gaussian Processes may limit scalability for very long optimization campaigns, potentially requiring the use of sparse GPs or alternative surrogate models. Additionally, the current model relies primarily on sequence information; incorporating 3D structural data could further enhance predictive accuracy. Finally, moving toward multi-fidelity optimization—using a mix of cheap, low-accuracy simulations and expensive, high-accuracy ones—could further accelerate the discovery process.

Summary updated

Abstract

The primary objective of most lead optimization campaigns is to enhance the binding affinity of ligands. For large molecules such as antibodies, identifying mutations that enhance antibody affinity is particularly challenging due to the combinatorial explosion of potential mutations. When the structure of the antibody-antigen complex is available, relative binding free energy (RBFE) methods can offer valuable insights into how different mutations will impact the potency and selectivity of a drug candidate, thereby reducing the reliance on costly and time-consuming wet-lab experiments. However, accurately simulating the physics of large molecules is computationally intensive. We present an active learning framework that iteratively proposes promising sequences for simulators to evaluate, thereby accelerating the search for improved binders. We explore different modeling approaches to identify the most effective surrogate model for this task, and evaluate our framework both using pre-computed pools of data and in a realistic full-loop setting.
BibTeX
@inproceedings{gessner2024active,
  author = {Alexandra Gessner and Sebastian W. Ober and Owen Vickery and Dino Oglic and Talip Uçar},
  title = {{Active learning for affinity prediction of antibodies}},
  booktitle = {BDU @ NeurIPS},
  year = {2024},
  url = {https://openreview.net/forum?id=rSnUPJ2n1D},
}
Figure