Research map / Probabilistic and scientific modeling / Molecules and proteins
TadA-Bench: A Million-Variant Benchmark for Future-Round Discovery Toward Agentic Protein Engineering
Jin Gao, Juntu Zhao, Zirui Zeng, Jiaqi Shen, Junhao Shi, Dukun Zhao, Yuming Lu, Dequan Wang
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 changed | Where | Measured | Effect |
|---|---|---|---|
| evaluate on random 8:1:1 split instead of fixed future-round split (ablation) | protein language models (ESM2, ESMC, ProtTrans) · TadA-Bench protein view | Spearman: 0.0509 -> 0.8079 (ESMC-600M test) | Improved |
| full fine-tuning instead of frozen encoder (ablation) | ESM2 and NT families · TadA-Bench future-round test | best test Spearman: 0.055 (ESM2), 0.063 (NT) | No clear effect |
| prompt tuning instead of frozen encoder (ablation) | NT-50M and ESMC-300M · TadA-Bench future-round test | best test Spearman: 0.075 (NT-50M), 0.027 (ESMC-300M) | No clear effect |
| increase encoder size from ESM2-35M to ESM2-650M (ablation) | ESM2 frozen probing · TadA-Bench future-round test | Spearman: 0.0300 -> 0.0479 | No clear effect |
| select small candidate budget (96 or 384) with a protein model ranker | ESM2-35M, ESMC-300M · TadA-Bench future-round, top-5% hit rate | hit rate: 3.1% and 4.9% at K=96 vs 5% random | No clear effect |
| train on diversity-selected or whole-round subsets instead of random subsample of the same size (ablation) | protein models on TadA-Bench · TadA-Bench future-round, Spearman vs training-set size | Spearman | Improved |
| increase training-set size by random subsampling (ablation) | protein models on TadA-Bench · TadA-Bench future-round | Spearman | Improved |
| rerun Seq2Graph with 50% of rounds subsampled (ablation) | Seq2Graph · label robustness | Spearman of shared-sequence scores: rho = 0.90 | No clear effect |
| rerun Seq2Graph with 50% of sequences per round (ablation) | Seq2Graph · label robustness | Spearman of shared-sequence scores: rho = 0.95 | No clear effect |
Closest papers by meaning
- Evolution-Inspired Loss Functions for Protein Representation Learning (ICML 2024 · similarity 0.86)
- BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments (ICLR 2025 · similarity 0.86)
- Protein Design with Agent Rosetta: A Case Study for Specialized Scientific Agents (ICML 2026 · similarity 0.86)
- AstaBench: Rigorous Benchmarking of AI Agents with a Scientific Research Suite (ICLR 2026 · similarity 0.85)
- ProteinOPD: Towards Effective and Efficient Preference Alignment for Protein Design (NeurIPS 2026 · similarity 0.85)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.85)
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.
