ProteinArena - EC Prediction - Level 2 Accuracy: leaderboard
Enzyme Commission number prediction by general LLMs from the raw amino-acid sequence, with no training split or database search, unlike CARE, which pairs its EC classification test sets with training splits for the methods it benchmarks. ProteinArena, AMix-2 paper v1 (arXiv 2605.30963, 29 May 2026). Test proteins are Swiss-Prot entries first public on or after 1 January 2025, and models work without retrieval, database lookup or workflow orchestration. The seven frontier LLMs ran at their default temperature, with top_p only where an API or reasoning mode needs it (0.95 for Claude) and max_tokens 8,192, raised to 32,768 when an answer could be longer. Thinking mode and effort are not reported. Unlike the paper's protein baselines, the LLMs' training data was not checked against the 2025 cutoff. EC Prediction, Table 2: the primary low-homology split (under 30% sequence identity to any sequence released by 31 December 2024). Given the amino-acid sequence, the model must give the full four-level EC code (for example 2.7.7.60); this board scores Level 2 of the four (Level 1 is x.-.-.-, Level 4 the full x.x.x.x code). Percent of proteins; the paper does not print the protein count. Not on this board: the paper's own AMix-2 AR and dLLM (unreleased; AR is an architecture ablation), the BLAST and Foldseek retrieval tools, and ESM2, ESM3, ProTrek and SaProt, which are scored through task heads the authors trained. The ≥30% identity tiers (Tables 11 and 12) are supplementary near-neighbour analyses and stay out. Question access: The paper says ProteinArena will be released; on 7 November 2026 the project page still showed 'Coming Soon' and no ProteinArena or AMix-2 repository existed on Hugging Face (GenSI, AMix-Bio). The test proteins are public Swiss-Prot entries, but the scored selection is not published.
Metric: Level-2 EC prediction accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 61.62 | |
| 2 | Gemini 3.1 Pro (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 34.34 | |
| 3 | Claude Opus 4.6 (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 21.21 | |
| 4 | DeepSeek V4 Pro (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 19.19 | |
| 5 | GPT-5.5 (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 18.18 | |
| 6 | GLM-5.1 (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 5.05 | |
| 7 | Qwen 3.5 27B (Benchmark Sweep Run: source=ProteinArena; default temperature; thinking/effort unreported) | 2.02 |
Interactive version: theaggregate.ai/benchmark?slug=proteinarena-ec-prediction-level-2-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-08.