← Back
Amazon Science

Advancing AI for biology: Teaching models to design and characterize antibodies

•10 min read•
#llm#agents#deployment
Level:Advanced
For:AI Engineers
✦TL;DR

Researchers at Amazon Bio Discovery introduced MochiBind, a sequence‑only antibody‑antigen binding affinity predictor that reframes the task as pairwise ranking rather than absolute affinity estimation, using ESM‑2 embeddings and a projection layer to infer relative strength. Pairwise predictions are aggregated with the TrueSkill Bayesian rating algorithm to produce a global ranking of thousands of candidates, addressing the out‑of‑distribution performance gap of prior models. The end‑to‑end workflow, integrated with an agent for binding‑site selection, yielded experimentally validated antibody hits against a novel cancer target, demonstrating a faster, cheaper path from target to lead. This approach enables engineers to replace absolute‑affinity models with a scalable ranking pipeline that better aligns with real‑world drug‑discovery workflows.

⚡ Key Takeaways

  • MochiBind predicts which of two antibodies binds more tightly, using pairwise comparisons instead of absolute affinity values.
  • The model leverages a pretrained protein language model (ESM‑2) to embed antibody‑antigen complex residues, then averages residues to a single embedding per complex.
  • A specially trained network layer projects these embeddings into a lower‑dimensional space for relative binding prediction.
  • Global rankings are generated by aggregating pairwise scores with TrueSkill, a Bayesian rating algorithm originally for video‑game player ranking.
  • The system was validated by delivering experimentally confirmed antibody hits against a novel cancer target.
💡 Why It Matters

By shifting to pairwise ranking and Bayesian aggregation, engineers can more reliably prioritize antibody candidates at scale, reducing experimental load and accelerating lead identification in biopharma pipelines.

✅ Practical Steps

  1. Use ESM‑2 to embed antibody‑antigen sequences, compute the mean residue embedding for each complex, and feed it into the trained projection layer of MochiBind.
  2. Generate pairwise binding predictions for candidate pairs and aggregate the results with the TrueSkill algorithm to obtain a ranked list of candidates.
  3. Integrate the ranked list into downstream lab workflows or an agent‑driven design loop to select binding sites and test top hits experimentally.

Want the full story? Read the original article.

Read on Amazon Science ↗

More like this

Censored vs. Truncated Regression in Python, Explained

Towards Data Science•#llm

New in AWS Quick: More capable flows, an updated web experience, and agents that work together

AWS ML Blog•#agents

Three insights you may have missed from theCUBE’s coverage of the Dell Technologies New York Symposium

SiliconANGLE AI•#deployment

How Google Duplex evolved to also call businesses for you in Gemini

Google Research•#agents

EXPLORE AI NEWS

Daily hand-picked stories on LLMs, RAG, agents and production AI — curated for engineers who ship.

BROWSE NEWS

GET THE WEEKLY DIGEST

Join engineers getting the Monday signal-over-noise AI breakdown. No spam, unsubscribe anytime.

LEARN AI ENGINEERING

Curated courses, research papers, repos and tutorials built for engineers leveling up in AI.

START LEARNING