Antibody Developability Triage Pipeline

Challenge

Antibody developability failures (poor solubility, aggregation, self-association) are usually caught late, after structure prediction and wet-lab assays, when they're expensive to fix. Most published tools that flag these risks either aren't reproducible from the paper alone or were never checked against whether their own confidence estimates are trustworthy outside the antibodies they were built on.

Approach

Heavy/light chain sequences go through six sequential stages: ANARCI IMGT numbering and QC, ImmuneBuilder structure prediction, a developability panel (a from-paper reimplementation of the 5 TAP descriptors on FreeSASA, Aggrescan3D aggregation propensity via subprocess, and a CamSol-style solubility score), ESM-2 embedding distance against a Thera-SAbDab reference cloud for applicability-domain/novelty scoring, and Spearman correlation benchmarking against measured assay values in GDPa1 and Jain et al. 2017. Each stage reads/writes a shared CSV keyed by antibody ID rather than running as one monolithic script.

What we found

The two from-paper developability descriptors correlate with their intended real assay outcomes: the CamSol-style solubility score against GDPa1's HAC assay (ρ = 0.73, n = 94, p < 0.0001), and the Aggrescan3D aggregation score against GDPa1's HIC assay (ρ = 0.46, n = 242, p < 0.0001). The applicability-domain check — whether the pipeline's confidence should degrade on antibodies further from its ESM-2 reference cloud — did not show a differentiator: mean |ρ| by novelty tertile was 0.12 / 0.15 / 0.14, essentially flat across the full novelty range.

Limits of this demonstration

This is an R&D demonstration on public data, not a client engagement. GDPa1 is a gated dataset — the 246-antibody run's output isn't committed to the repo, so reproducing it requires requesting Hugging Face access (the smaller Jain et al. 2017 cross-check is fully public). The applicability-domain novelty check, the pipeline's own attempt at flagging when its predictions should be trusted less, is a flat null result rather than a working differentiator, reported as-is rather than dropped from the writeup.