Res ipSAE adhuc loquuntur: What the directionality of predicted aligned error in cofolding models can tell us about scoring diverse molecular complexes

Directional ipSAE scoring of predicted aligned errors in protein–ligand complexes yields wider confidence-score ranges than native ipTM, and classifies ligand placements with balanced accuracies of 0.938 (AF3) and 0.862 (Boltz-2).

• Why it matters: ipTM-based confidence can be biased by asymmetric PAEs, system-size–dependent d0, and lowered scores from disordered or non-interface regions, limiting reliable interpretation across diverse complex types.
• What they did: The method extends ipSAE by explicitly choosing alignment frames, scored tokens, and d0; for protein–ligand pairs it aligns on protein residues, scores all ligand heavy atoms without a PAE cutoff, and uses fixed d0 = 2 [A].
• The result: On a held-out set of 82 AF3/Boltz-2 protein–ligand complexes, ipSAE thresholds of 0.5 (AF3) and 0.8 (Boltz-2) with ligand RMSD ≤ 5 [A] achieve balanced accuracies of 0.938 and 0.862, with protein–DNA support and preliminary protein–RNA results.

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