ProteinArena - Functional Design - pLDDT: leaderboard

Writing a protein sequence to InterPro function keywords, scored with PDFBench's keyword-guided metrics (InterProScan recovery, ESMFold pLDDT); unlike PDFBench, whose main tables rank protein design models, the contestants here are general LLMs given a plain-language instruction. 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. Functional De Novo Design, Table 1: prompts give the expert-reviewed InterPro function keywords of 870 Swiss-Prot targets under the time-aware protocol (no homology filter is stated for this track); the model must return exactly one uppercase amino-acid sequence of at most 1,024 residues. The paper reports partial coverage only for protein design models (CFP-Gen 319 and ESM3 826 of 870) and prints no per-LLM exclusions or invalid-output handling. Repetition (Rep, Rep2, Rep5), novelty and uniqueness are degeneration and memorization diagnostics and are not admitted. Not on this board: the Natural reference row, the six protein design models (each with its own conditioning format and coverage) and the paper's own unreleased AMix-2 AR and dLLM. Table 10 recomputes the same rows on the 207 targets every baseline covers and stays out. pLDDT: ESMFold-v1 confidence in the predicted structure of the generated sequence (structural plausibility; it does not check the requested function). 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: ESMFold-v1 mean pLDDT (0-100). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.

Top models

Interactive version: theaggregate.ai/benchmark?slug=proteinarena-functional-design-plddt · How It Works · Data refreshed daily, snapshot 2026-11-08.