BindCraft has become a go-to tool for protein binder design, using AlphaFold2 hallucination + ipTM-based backpropagation to optimize sequences. But ipTM has limitations: ⚠️ It's a heuristic, not a proper likelihood ⚠️ It yields sparse gradients, making optimization tricky

We fix this with pTMEnergy 💡 It reinterprets AlphaFold’s pAE logits as an energy-based model (EBM) which: ✅ Reflects the true statistical likelihood of the binder-target complex ✅ Enables dense gradients across the entire interface

By integrating pTMEnergy into the hallucination loop, we get BECraft, a drop-in extension to BindCraft that backprops through AlphaFold2 using this new energy. BECraft outperforms BindCraft, RFDiffusion, and ESM3 across 8 diverse targets 🚀

Beyond hallucination, pTMEnergy shines in virtual screening. It outperforms supervised, physics-based, and structure-based baselines at classifying miniprotein binders and RNA aptamers.

💻 BECraft Code: https://t.co/4PCoQJze2R Very grateful to work with Anisha Parsan, @CarolineUhler, and @WengongJin. Huge thank you to the original BindCraft creators, @MartinPacesa @sokrypton @befcorreia, for building a powerful foundation and making the code easy to extend