Learning the limits of protein structure models with NMR chemical shifts

SLAE-ERR predicts its own backbone 15N chemical-shift error with a mean absolute error of 0.76 ppm, using embeddings from SLAE trained on NMR-linked dynamic regions.

• Why it matters: Protein foundation models are hard to evaluate for specific biochemical knowledge because most downstream benchmarks lack experimental, per-atom resolution to reveal failure modes. NMR chemical shifts provide a sensitive structure-linked probe.
• What they did: The authors curated a chemical-shift dataset to reduce experimental confounders and benchmarked diverse protein foundation models on predicting backbone 15N chemical shifts. They then used molecular dynamics to investigate the largest errors in core residues with fast-timescale dynamics and trained SLAE-ERR to predict its own error.
• The result: Across models, the largest prediction errors occur in core residues with evidence of fast-timescale dynamics, and this structured dynamic behavior is linked to model failure. SLAE-ERR’s 0.76 ppm mean absolute error enables self-assessment of prediction error for improved protein design guidance.

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