Study Maps Which AI Antibody Filters Match Wet Lab Outcomes

The short version

Scientists studied which computer structure scores actually help pick AI designed antibodies for lab tests. They worked with more than four thousand antibody pairs that already had wet lab results. The paper is on arXiv and is a preprint, not yet peer reviewed.

What the researchers did

Langzhang Liang, Juncheng Qian, Xiaojie Ma, and coauthors posted Revisiting Structure-Based Filtering for AI-Designed Antibodies: Opportunities, Limits, and What Matters as arXiv 2610.23136. The GitHub repo Antibody-filtering ships code called RIC and NAC plus scripts to replay the filters.

AI tools now propose many antibody shapes before any bench work starts. Labs cannot test every design. Structure predictors draw 3D shapes and attach confidence scores. The team asked which scores truly rank winners on real pairs.

They used 4,184 antibody pairs with known lab outcomes. Four structure predictors ran under two filter setups. The study shows which scores track real success and which ones look good on paper but fail in practice.

Photo: National Cancer Institute / Pexels

What they found

Some predicted structure metrics help sort lists while others add little beyond simple baselines. Regional confidence and repeat run agreement behave differently across predictors. The repo publishes RIC and NAC so teams can apply the same rules to fresh designs.

Exact lift values and cutoffs are in the PDF tables. The work does not promise a perfect filter. It maps where structure triage saves cycles and where it sends good candidates to the trash.

Open the paper before you change gate rules at your company. Numbers can shift when targets or assays differ from the public set.

Photo: CDC / Pexels

Why it matters for you

Biotech groups using generative antibody models must shrink design lists before costly assays. Each false keep burns reagents and time. Each false drop may hide a strong drug candidate.

Investors should ask portfolio companies which filters are evidence based. Regulators will want traceable rules when AI designs enter the clinic.

Keep the arXiv and GitHub links with your wet lab SOPs. Replay RIC and NAC on a blinded slice of your own data before you trust them in production.

What to keep in mind

This is a preprint, not yet peer reviewed. Predictors update quickly, so thresholds may need retuning. Structure tools can be overconfident on rare targets.

Human review for safety and developability stays required. No score replaces full validation.

What a careful reader should do next

Print the table that compares filters side by side. Circle the rows that match your assay type. Ask your CRO whether they already use similar scores or need fresh validation data. Small pilots beat wide rule changes that discard unknown winners.

Key takeaways

• arXiv 2610.23136 covers 4,184 characterized antibody antigen pairs.
• Open code implements RIC, NAC, and baseline scores.
• Some structure signals help triage and some do not.
• Filtering saves cost but does not replace full assays.
• Retune on your own data before wide use.

Sources

• Primary — arXiv, Revisiting Structure-Based Filtering for AI-Designed Antibodies (2026). Dataset size, methods, and conclusions.
• Secondary — GitHub, Jinx-byebye/Antibody-filtering (2026). Code and citation tied to the paper.