ipSAE extends to ligands: fixing poor ipTM scoring in AF3 and Boltz-2 cofolding models
sokrypton · x · 2026-11-05
Roland Dunbrack Jr. (Fox Chase Cancer Center) published a new bioRxiv preprint extending ipSAE, previously introduced for protein-protein complexes, to interactions involving proteins, nucleic acids, small-molecule ligands, and ions.
- The problem: AF3 and Boltz-2 rely on ipTM for complex confidence, but ipTM performs poorly for ligand interfaces. PAE matrices are asymmetric — AF3 reports the max of directional scores while Boltz-2 reports them separately — and the TM-score d0 parameter depends on total system size rather than the scored component.
- ipSAE makes three explicit choices per molecular pair: which component supplies alignment frames, which tokens are scored, and how d0 is defined.
- Practical takeaway: for ligands, set d0 to 2 and align on protein while scoring the whole ligand; ipSAE=0.5 separates good from bad models in AF3, with 0.8 as the threshold for Boltz-2.
Related event: ipSAE Scoring Extended to Ligands and Protein-Nucleic Acid Complexes(2 posts)→
More from Research
- 2-person lab Linum trains 2B text-to-image model from scratch for just $21K, Apache 2.0 — ycombinator · 2026-11-05
- Trained an MoE across two GPU clusters 25 Gbps apart — the compute islands guide — _arohan_ · 2026-11-05
- AutoDNA: AI agents run multi-step molecular biology experiments with minimal human input — james_y_zou · 2026-11-05
- The Stacks Project, algebraic geometry's giant reference, is now formalized in Lean — LinusEkenstam · 2026-11-05
- Perceptron Egocentric 1.0 turns egocentric video into robot training data at $29/hour — rohanpaul_ai · 2026-11-05
- Study benchmarks divergent creativity of LLMs against 100,000 humans in Scientific Reports — korymath · 2026-11-05