Use AlphaBind Model Finetuning Online

AlphaBind: Antibody-Antigen Binding Affinity Prediction and Sequence Optimization with a Finetunable Deep Learning Model

AlphaBind is a domain-specific deep learning model that can predict and optimize antibody-antigen binding affinity. It was released by A-Alpha Bio as a pre-print with open code and a pre-trained checkpoint; the GitHub repository is MIT licensed. The model is pre-trained on millions of affinity measurements and then finetuned on data for a single parental antibody, so that it can guide the optimization of therapeutic antibody candidates.

Improving the affinity of an antibody that is already a good starting point is slow and expensive when it relies on rounds of random mutation and screening. Generic sequence models carry no direct knowledge of binding strength, so the authors combine protein language model embeddings with a large body of quantitative affinity data to build a model that is specific to this task.

How AlphaBind Predicts Antibody Binding Affinity

Antibody and target sequences are each encoded with ESM-2nv embeddings. The two embeddings are concatenated and passed through a transformer encoder trained with a mean squared error loss to predict affinity.

1. Pre-training: the model learns general rules of binding from about 7.5 million measurements covering roughly 300,000 antibody constructs and 6,000 target constructs. Data on the antibodies and targets later optimized were excluded.

2. Finetuning: the model is then adapted to one parental antibody using a library of about 30,000 variants, including nearly all single missense mutations, some insertions and deletions, and random double and triple mutants. A 90/10 train/validation split was used.

3. Optimization: a stochastic greedy search with ESM-2nv masking ran 60,000 trajectories over 100 generations, giving about 6 million proposals per antibody. Candidates were filtered with TAP developability checks and about 7,500 were selected across edit distances of 2 to 11.

AlphaBind Results: Affinity Gains in Optimized Antibodies

The authors tested optimized variants of several parental antibodies and report the following improvements.

Validation and developability

All 15 of the top candidates expressed, and all 10 tested by BLI had better affinity than the parent. Among candidates binding at 1 uM or tighter, specificity was 80% for VHH72, 92% for Pembrolizumab-scFv and 97% for AAB-PP489. In a developability-focused round, the optimized AAB-PP3117 reached 310 fM as an IgG, removed two liability motifs, reverted 4 of 9 residues to germline and raised expression from 259.1 to 334.1 ug/mL. Ablations showed that removing ESM-2nv embeddings or pre-training significantly degraded performance.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with AlphaBind on Tamarind Bio

• Affinity maturation: improve the binding of a lead antibody, including candidates that were already affinity-matured.

• Multi-property optimization: combine affinity gains with developability goals such as removing liability motifs and improving expression.

• Specificity-aware design: select variants that keep binding to the intended target and avoid a negative control.

• Different antibody formats: the authors applied the approach to scFvs and a VHH, as well as an IgG follow-up.

How to Use AlphaBind on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and open the AlphaBind Model Finetuning tool.

2. Provide finetuning data: supply measured affinity data for the variants of your parental antibody and its target, as the tool describes if exposed.

3. Set training parameters: the authors used 100 epochs with a 90/10 train/validation split; adjust any exposed settings accordingly.

4. Run the finetuning job: the paper reports about 1 hour per antibody on one H100 GPU.

5. Download the outputs: retrieve the finetuned model and predictions once the job completes.

6. Interpret the results: check agreement between predicted and measured affinities on held-out variants before relying on the model.

7. Use downstream: in the authors' workflow, the finetuned model drives an optimization search and the proposals are filtered and tested experimentally.

Things to Keep in Mind

• The model is not zero-shot: it needs local finetuning data for each parental antibody.

• Gains vary by parent antibody, from 2-fold to 74-fold in the study.

• Only one round of optimization was performed.

• Generalization to other classes, such as GPCR-targeting antibodies, pH-dependent binders and broadly cross-reactive antibodies, is untested.

• Reported results are from a pre-print and depend on experimental validation of proposed sequences.

Source: AlphaBind, a Domain-Specific Model to Predict and Optimize Antibody-Antigen Binding Affinity, bioRxiv pre-print, https://www.biorxiv.org/content/10.1101/2024.11.11.622872v1 (DOI 10.1101/2024.11.11.622872). Code: https://github.com/A-Alpha-Bio/alphabind.