Amazon Releases Three Papers on AI-Based Antibody Design and Evaluation

The Amazon Bio Discovery team released three papers on AI-based antibody design and evaluation frameworks to address the time and cost bottlenecks in monoclonal antibody development. While existing models suffered from performance degradation on new targets, these studies focused on improving universal prediction and applicability to real-world drug development.

Relative Binding Strength Predictor MochiBind

The first study proposes MochiBind. Instead of absolute binding affinity, it predicts the relative binding strength between two antibodies against the same antigen. It utilizes the ESM-2 protein language model to generate embeddings from sequences alone without structural input, and aggregates global rankings using the TrueSkill algorithm.

• Performance: Achieved an average pairwise accuracy advantage of approximately 10% over structure-based baselines in evaluations on the AlphaBind dataset (including TIGIT, PD-1, HER2, SARS-CoV-1 RBD, etc.).
• Efficiency: Demonstrated fast inference speed, scoring 200,000 pairs in about 13 seconds on a CPU.

CA-MAP Correcting for Lab Variance

The second study is CA-MAP (Context-aware multi-property antibody predictor). It adopts a multimodal structure that leverages context during inference to address measurement variances between laboratories (batch effects). It uses a MAMBA-based state space model, encoding and integrating sequences, property names, and numerical values separately.

• Batch Effect Mitigation: In a simulated batch effect environment, standard fine-tuning saw correlation drop to 0.58, whereas CA-MAP maintained 0.99.
• Zero-shot Prediction: Unlearned properties can be predicted via context, suggesting the potential to replace expensive assays by leveraging correlations.

Nanobody Design for Novel Drug Targets

The third study involves de novo nanobody design targeting cell surface proteins of Desmoplastic small round-cell tumors, a pediatric cancer target. For novel targets lacking structural information, it recommends hotspots through an agent-based workflow and selects candidates using generative models.

• Experimental Validation: Yeast surface display screening results identified 46 out of 116 candidates as strong binders, with zero binding to unrelated proteins confirming specificity.
• Lab-in-the-loop: Applied a closed-loop workflow that utilizes experimental data as training data for the next design cycle.