bioRxiv preprint extends ipSAE confidence scoring to diverse molecular complexes
victorgreiff · x · 2026-11-04
Roland Dunbrack Jr. (Fox Chase Cancer Center) posted a bioRxiv preprint extending his ipSAE framework — originally developed for protein–protein complexes — to scoring interactions involving proteins, nucleic acids, small-molecule ligands, and atomic ions in cofolding models.
Key points:
- AF3's ipTM and Boltz-2's PAE-based scores have structural flaws: the PAE matrix is directional/asymmetric, AF3 only reports the max of the two directional ipTM scores while Boltz-2 reports them separately, and the d0 TM-score scaling depends on overall system size rather than the scored component; unrelated domains and disordered regions further depress ipTM.
- ipSAE fixes this by making three explicit choices per molecular pair: which component supplies alignment frames, which tokens are scored, and how d0 is defined.
More from Research
- CWE-Bench, a frontier cyber benchmark where agents find and fix real vulnerabilities, goes open-source — mmitchell_ai · 2026-11-04
- FLoRe teaches humanoid robots force-reactive loco-manipulation from robot-free human demos — HaozhiQ · 2026-11-04
- Telling agents their token throughput in the prompt improves time forecasting — maksym_andr · 2026-11-04
- NRGPT paper reimagines GPT inference as gradient descent on an explicit energy landscape — burkov · 2026-11-04
- Fireworks event slides show sliding-window bit-sharing quantization trick — sophiamyang · 2026-11-04
- UCLA Study Seeks Users of 'Griefbots' to Interview on AI and Mourning ($30 Gift Card) — Interesting-Cat-7838 · 2026-11-04