DeePEn - A Depth sensitive benchmark for Protein Engineering

DeePEn shows that protein variant-effect prediction performance steadily deteriorates as mutational distance from the wild type increases across four deep mutational scanning datasets.

• Why it matters: Real-world protein engineering must generalize to held-out variants far from the wild type, but existing benchmarks do not robustly reflect this depth-sensitive challenge.
• What they did: DeePEn retrospectively evaluates models by defining mutational distance as the number of simultaneous point mutations (edit distance) and using experimentally labeled DMS variants from ProteinGym.
• The result: Across general and biophysics-informed protein language models and a non-transformer network, accuracy declines with distance, motivating multi-metric benchmarking for distant-variant prediction.

AI summary of the abstract · may contain errors · not medical advice · read the paper.