AI Paper Atlas

Research map / Probabilistic and scientific modeling / Molecules and proteins

Explicit representation of germline and non-germline residues improves antibody language modeling

Jeonghyeon Kim, Nathaniel Blalock, Ameya Kulkarni, Kensuke Nakamura, Philip Romero

ICML 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein fitness and variant effect prediction · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
factorize tokens into germline/non-germline types with an Origin-Conditioned LM head (ablation)antibody language model (ESM2-35M backbone) · GL/NGL linear probePR-AUC: 0.377 -> 0.980Improved
add Origin-Conditioned multi-head instead of simple head on extended vocabulary (ablation)PRISM · Generative PPL (heavy chain)pseudo-perplexity: Ablations 2 and 3 ~50 and ~25 heavy-chain PPL vs 1.54Improved
add unpaired pretraining to the multi-head model (ablation)PRISM · GL/NGL linear probePR-AUC: 0.517 -> 0.980Improved
use factorization rather than ESM2 initialization with antibody fine-tuning (ablation)PRISM vs PRISM-less · CR9114 zero-shot affinitySpearman rho: -0.409 -> +0.393Improved
use factorization rather than ESM2 fine-tuning (ablation)PRISM vs PRISM-less · Trastuzumab zero-shot affinitySpearman rho: 0.176 -> 0.327Improved
ESM2 fine-tuning without factorization gives strong probe separability but no geometric organization (ablation)PRISM-less · GL/NGL embedding separabilityARI: PR-AUC 0.921, ARI 0.006No clear effect
add stop-gradient before conditioning downstream heads on Origin Head (ablation)PRISM · validation NGL perplexityNGL perplexity: ~2x improvementImproved
keep loss reweighting (alpha_NGL=3.0) and 53-token vocabulary but drop the architecture (ablation)PRISM · GL/NGL separabilityARI: 0.162 -> 0.009Hurt
sample from NGL-only token channel (PRISM-NGL)PRISM generation · G6.31, CR9114, TrastuzumabRosetta interface ddG variants improvingImproved
sample from GL-only token channel (PRISM-GL)PRISM generation · G6.31, CR9114, TrastuzumabRosetta stability and CamSolImproved
NGL-biased sampling scored by DMS ridge predictorPRISM-NGL · CR9114, Trastuzumabridge-based predicted bindingHurt
marginalize GL and NGL channels into one score (ablation)PRISM scoring · binding DMS (Trastuzumab, CR9114)Spearman rhoHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy