Title: TadA-Bench: A Million-Variant Benchmark for Future-Round Discovery Toward Agentic Protein Engineering

Abstract

As artificial intelligence for scientific discovery transitions into an agentic phase, protein-engineering platforms are increasingly tasked with prioritizing upcoming wet-lab experiments over the mere fitting of static data. To support this shift, we present TadA-Bench, a comprehensive benchmark derived from a million variants across 31 directed-evolution rounds of TadA, designed specifically for future-round discovery in agentic protein engineering. This benchmark maintains the chronological order of the experimental campaign and establishes a fixed-data replay objective: models must rank variants that emerge exclusively in later rounds based on data from earlier rounds.

The dataset offers synchronized views of DNA, RNA, and protein sequences. To address the challenge of noisy enrichment measurements, we employ Seq2Graph, a graph-based label-unification pipeline, to generate consistent activity labels across different rounds. While random-split evaluations demonstrate strong interpolation capabilities, the models exhibit significantly weaker performance in future-round ranking and selecting candidates under finite budget constraints. Further controlled analyses indicate that evolutionary coverage holds greater informational value than local data density. Consequently, TadA-Bench serves as a reproducible substrate for wet-lab replay to advance future-round discovery in agentic protein engineering. The associated data and code are publicly available on Hugging Face and GitHub.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC