Thal-Kak: Unified Interface Beats AlphaFold3 with Mixed Predictors
DaveJuergens · x · 2026-08-25
The CSSB-SNU lab released Thal-Kak, a unified interface running five all-atom structure predictors (Boltz-2, Chai-1, Protenix, ESMFold2) from a single input file.
Key Findings:
- Ensemble Wins: Splitting a fixed sampling budget equally across the five predictors yielded better structures than any single predictor alone at a matched budget, even beating AlphaFold3.
- The Selection Bottleneck: The accuracy gain relies on an "oracle" view. Confidence scores don't transfer across predictors, and mixed pools lose their advantage under Top-1 selection.
Features:
- Shared MSA and common output schema.
- Google Colab notebook included for easy testing.
More from Research
- Can incident response cases be used to improve model alignment? — repligate · 2026-08-25
- Opinion: Discrepancy theory could optimize resource allocation for AI agents — JosephJacks_ · 2026-08-25
- Geometric modeling of Occam's razor explains deep learning generalization — FrnkNlsn · 2026-08-25
- MVAP-G paper demonstrates fooling Visual Geometry Grounded Transformers — kwangmoo_yi · 2026-08-25
- PhysCaP enables robots to physically probe hidden properties like mass and stiffness — chris_j_paxton · 2026-08-25
- Prime Agent: An Open-Source Harness for Self-Improving RLM — PrimeIntellect · 2026-08-25