SPOC Reranks ESMFold2 PPI Screens: Matches AlphaFold2-Multimer Accuracy, 2.7x Faster
proteinrosh · x · 2026-09-20
This work tackles a key bottleneck in proteome-scale structural PPI screening: structure predictors often produce plausible interfaces even for non-interacting pairs, so structural confidence alone can't reliably indicate biological interaction — especially under extreme class imbalance.
- SPOC scoring method: a Random Forest domain-domain interaction classifier built on ESMFold2 predicted structures + public biological databases (39 features: 25 structural + 14 biological), shipped as a self-contained GitHub repo with inference.py and notebooks reproducing the paper's figures
- Scalable pipeline around MSA-free ESMFold2-Fast: cutting recycling to 3 iterations preserves accuracy — 57% acceptable-or-better DockQ (≥0.23) on FoldBench PPIs, comparable to AlphaFold2-Multimer via LocalColabFold (56%)
- Speed: 266 min vs 718 min; PAE-guided domain parsing works around GPU memory/length limits (≤350 aa per chain, ≤750 aa complex)
Takeaway: reranking with joint structural and biological evidence maintains PPI screening accuracy at a fraction of the runtime.
More from Research
- XGEN Labs unveils generative world simulation JING+DAO, tops WBench leaderboard — hey_abusiddik · 2026-09-20
- Schmidhuber: LLMs aren't truly creative because they lack compression progress — SchmidhuberAI · 2026-09-20
- Jev tested on 8,054 NASA Kepler signals: 54.2% accuracy, loses to a simple 3-rule baseline — This_Cell_1829 · 2026-09-20
- François Fleuret nicknames his training curves; researchers admit they curse baselines too — giffmana · 2026-09-20
- HN: How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip — petrusenko_max · 2026-09-20
- FlashNorm: two lines of algebra buy 33-35% speedup — and a CUDA race that made the model echo the past — AI Engineer · 2026-09-20