Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Antibody Humanness Prediction using Patent Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Pre-trains a small transformer antibody encoder with weakly-supervised contrastive learning on patent-derived target groups, then fine-tunes it with cross-entropy for humanness prediction (FT-SelfPAD), outperforming AbNatiV, OASis and Hu-mAb on three immunogenicity-related datasets.
Problem
Predicting antibody humanness as a proxy for immunogenicity of therapeutic antibodies, where existing tools rely on OAS-based k-mer or multi-model approaches and limited reference diversity.
Approach
Use noisy functional identifiers (potential targets) from the Patented Antibody Database as weak supervision: contrastive pre-training groups sequences sharing targets, then the same encoder is fine-tuned with a supervised humanness classifier head plus shift/swap noise.
Key results
- On the PAD test set, FT-SelfPAD reaches 97.59 F1 and 99.12 ROC AUC, versus 91.59 F1 and 91.52 ROC AUC for AbNatiV.
- On 553 therapeutics (paired), FT-SelfPAD achieves 93.33 F1 and 94.47 ROC AUC, the best among the compared methods.
- On the 25 humanization pairs, FT-SelfPAD classifies all sequences correctly (100 on all metrics for paired data).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · PAD test set | ROC AUC: 91.52 -> 99.12 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 553 therapeutics, paired | ROC AUC: 90.02 -> 94.47 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 553 therapeutics, single chain | ROC AUC: 86.91 -> 91.84 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 217 immunogenicity, paired | ROC AUC: 79.90 -> 82.86 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 217 immunogenicity, single chain | ROC AUC: 77.86 -> 78.94 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 25 humanization data, paired | ROC AUC: 99.52 -> 100 | Helped |
| use FT-SelfPAD instead of the strongest baseline, AbNativ | humanness prediction · 25 humanization data, single chain | ROC AUC: 96.12 -> 100 | Helped |
| add pre-training rather than train from scratch (ablation) | SelfPAD on 217 immunogenicity data · 217 immunogenicity data | ROC AUC: 80.32 -> 82.86 | Helped |
| add random initialization of four pre-trained layers before fine-tuning (ablation) | SelfPAD on 217 immunogenicity data · 217 immunogenicity data | ROC AUC: 81.90 -> 82.86 | Helped |
| add swap noise and shift noise (ablation) | SelfPAD with layer re-initialization on 217 immunogenicity data · 217 immunogenicity data | ROC AUC: 81.53 -> 82.86 | Helped |
| add swap noise (ablation) | SelfPAD with layer re-initialization and shift noise on 217 immunogenicity data · 217 immunogenicity data | ROC AUC: 81.53 -> 82.86 | Helped |
| add shift noise (ablation) | SelfPAD with layer re-initialization and swap noise on 217 immunogenicity data · 217 immunogenicity data | ROC AUC: 82.37 -> 82.86 | Helped |
Surprising findings
- Ablation on the 217 dataset shows fine-tuning from scratch without pre-training performs worse than with pre-training, though the gap is modest (ROC AUC 80.32 vs 82.86).
Limitations
- PAD function identifiers are noisy and may corrupt the contrastive representation.
- PAD is biased toward lab-optimised rather than natural sequences, and OAS-like natural diversity is limited.
- Humanness is only a proxy for immunogenicity, and humanness labels in evaluation datasets are approximations.
Experimental setup
Datasets: Patented Antibody Database (PAD) 553 therapeutics dataset 217 immunogenicity dataset 25 humanization dataset
Benchmarks: PAD test set 553 therapeutics 217 immunogenicity 25 humanization data
Metrics: F1 recall precision accuracy ROC AUC PR AUC
Baselines: AbNatiV OASis Hu-mAb
Builds on: SimCLR-style NT-Xent contrastive loss AbNatiV OASis Hu-mAb
Closest papers by meaning
- Pareto-Optimal Energy Alignment for Designing Nature-Like Antibodies (NeurIPS 2025 · similarity 0.83)
- Bayesian Optimization of Antibodies Informed by a Generative Model of Evolving Sequences (ICLR 2025 · similarity 0.83)
- Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints (ICML 2024 · similarity 0.82)
- Immunogenicity Prediction with Dual Attention Enables Vaccine Target Selection (ICLR 2025 · similarity 0.82)
- Antigen-Specific Antibody Design via Direct Energy-based Preference Optimization (NeurIPS 2024 · similarity 0.82)
- Retrieval Augmented Diffusion Model for Structure-informed Antibody Design and Optimization (ICLR 2025 · similarity 0.81)
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.