AlphaBind

Every amino acid substitution in a 100-residue antibody multiplies the search space by roughly 2,000, and almost every sequence drawn from that space binds worse than the parent — or fails to behave as a recombinant protein at all. Affinity maturation has therefore been an iterative experiment: build a library, screen it, pick winners, repeat. AlphaBind attacks the same problem by learning the local fitness landscape around a parental antibody well enough that a single round of measurement supports an in-silico search over millions of proposals.

The model is a supervised regressor that maps an antibody sequence and a target sequence to a binding affinity. Its distinguishing asset is what it saw before any particular campaign starts: approximately 7.5 million quantitative affinity measurements drawn from 20 unrelated AlphaSeq yeast-display experiments, covering roughly 300,000 distinct antibody constructs against about 6,000 target constructs. That pre-training corpus is quantitative KD data rather than binary enrichment, which is what lets a regressor rather than a classifier sit at the centre of the pipeline.

AlphaBind was developed by A-Alpha Bio with a collaborator from NVIDIA, and released as a bioRxiv preprint in November 2024 before peer review in mAbs. Sequence embeddings come from ESM-2nv, NVIDIA's BioNeMo build of ESM-2, which is used both to featurize inputs and, at optimization time, to propose mutations.

#Key Features

• Quantitative affinity as the pre-training signal: Pre-training on millions of measured KD values across unrelated antibody-antigen systems transfers to new parental antibodies, and the paper's ablations show removing it degrades candidate quality.
• One round of wet-lab data per campaign: A library of roughly 30,000 variants spanning an 83-residue mutable window supplies the fine-tuning set; no iterative model-experiment loop is required.
• Language-model-guided proposal step: Optimization masks 1-3 positions and samples replacements from ESM-2nv logits rather than uniformly, which reaches edit distances of up to 11 mutations in a third of the generations random sampling needs.
• Multi-objective engineering on top of affinity: Because thousands of high-affinity candidates survive, they can be filtered for developability and germline content without sacrificing binding.

#Technical Details

Antibody and target sequences are embedded with ESM-2nv, concatenated, and reduced through two 1x1 convolutions from 2,560 to 256 dimensions. A transformer encoder with 4 attention heads and 7 layers, followed by average pooling and a linear head, predicts affinity under MSE loss — about 15 million parameters in total. Fine-tuning a campaign-specific model takes roughly an hour on a single H100 over 100 epochs. Optimization runs 60,000 stochastic greedy trajectories for 100 generations, costing about five hours on eight H100s and roughly $200 per parental antibody; candidates are then binned by edit distance from 2 to 11, screened with TAP developability flags, and cut to ~7,500 sequences for AlphaSeq validation and five for biolayer interferometry.

Across three AlphaSeq campaigns — a humanoid anti-TIGIT scFv, an scFv-formatted pembrolizumab, and the camelid VHH72 — all 15 top candidates expressed as scFv-Fc and 10 of 10 measurable ones improved on their parent, with a 74-fold gain for the best anti-TIGIT variant (56.5 pM to 766 fM by kinetic exclusion assay) and 14-fold for VHH72. A fourth campaign fine-tuned instead on ~36,000 published mammalian-display trastuzumab CDRH3 variants, a different assay format entirely, and still produced a 2.8-fold affinity gain. Using the fine-tuned model for developability engineering yielded a variant with both sequence liability motifs ablated, four of nine positions reverted to human germline, and 310 fM affinity as an IgG.

#Applications

AlphaBind targets antibody engineering teams doing lead optimization: taking a hit from phage panning or an already-matured clinical-stage antibody and generating a large, sequence-diverse pool of improved variants from one library build. Because the pipeline hands back thousands of viable candidates rather than a handful, downstream selection can optimize for expression, thermostability, immunogenicity, or germline identity in the same pass as affinity.

#Impact

The code, tutorial notebooks, per-campaign AlphaSeq fine-tuning and validation datasets, and a 65 MB pre-trained checkpoint are all published under the MIT license, which is unusually complete for an industrial antibody-optimization model and makes the ablation results reproducible. The main constraint is honest and stated by the authors: the pre-trained checkpoint is a starting point for target-specific fine-tuning, not a zero-shot affinity predictor, and genuinely generalizable prediction would require substantially more and more diverse pre-training data. Building the container also requires a free NVIDIA NGC account to fetch the ESM-2nv weights.