BindEnergyCraft

BindEnergyCraft, abbreviated BECraft, is a de novo protein binder design pipeline from researchers at the Broad Institute of MIT and Harvard and Northeastern University. It targets a specific weakness in hallucination-based binder design, the family of methods that generate binders by backpropagating through a frozen structure predictor to maximize a confidence metric — most commonly the interface predicted TM-score, ipTM.

The objection the authors raise is that ipTM is a heuristic, not a likelihood. It summarizes the predictor's confidence into a single scalar, which says nothing about how probable the binder–target complex is under the distribution the predictor actually learned, and it produces sparse gradients that make optimization inefficient. BECraft's answer is to extract a genuine statistical energy from the same frozen network. Applying the Joint Energy-based Modeling framework to a structure predictor's predicted aligned error logits yields pTMEnergy, a free-energy-style quantity computed over the full distribution of inter-residue error rather than its point summary.

The change is deliberately minimal in engineering terms: BECraft keeps BindCraft's optimization loop intact and substitutes pTMEnergy for the ipTM term. BindCraft is not separately cataloged on bio.rodeo, but BECraft is distributed as a drop-in extension of it. Nothing is retrained — the structure predictor stays frozen, and only the candidate binder sequence is updated — which places BECraft in the same category as other guided-sampling pipelines built over pretrained backbones, and puts it in direct competition with RFdiffusion and ESM-3 on binder generation.

#Key Features

• Energy from confidence: pTMEnergy is derived from a structure predictor's predicted aligned error logits under the Joint Energy-based Modeling reformulation, converting a confidence readout into a statistical energy without any architectural change or retraining.
• Drop-in objective swap: The design loop, including logit-space gradient descent, temperature annealing, straight-through estimation, and greedy mutation, is inherited from BindCraft; only the ipTM loss term is replaced.
• Denser optimization signal: Because the energy integrates over the whole inter-residue error distribution rather than collapsing it, it yields more informative gradients than ipTM for sequence optimization.
• Fewer structural clashes: Designs produced under the energy objective show reduced clashing relative to ipTM-optimized designs at matched or higher success rates.
• Doubles as a screening score: pTMEnergy is a scoring function in its own right and sets a new state of the art on structure-based virtual screening for both miniprotein binders and RNA aptamers.

#Technical Details

The frozen energy model is AlphaFold2-Multimer, from which pTMEnergy is computed by weighting the predicted aligned error distance bins by physical plausibility and aggregating the resulting log-sum-exp over inter-chain residue pairs; the term enters the BindCraft objective with a weight of 0.05. Binders were designed against eight targets — ALK, H3, IL-2Rα, IL-7Rα, InsulinR, LTK, TrkA, and VirB8 — generating 100 sequences of 55 to 65 residues per target with no hotspot constraints. BECraft attains the highest in silico success rate on 7 of 8 targets, with gains up to 16 percentage points over the next best method: on IL-7Rα it reaches 0.66 against BindCraft's 0.51, and on VirB8 0.81 against 0.72, while filtered RFdiffusion peaks at 0.35 and ESM3 registers 0.00 on every target. As a screening function pTMEnergy reaches AUPRC 0.467 on miniprotein binders versus 0.434 for ipTM, and 0.352 on GFP-binding RNA aptamers versus 0.253 for ipTM, ahead of geometric and transformer-based baselines.

#Applications

BECraft suits groups already running hallucination-based binder campaigns, since adopting it means changing an objective rather than adopting a new model or retraining anything. pTMEnergy is separately useful as a ranking function for structure-based virtual screening, including for RNA aptamers, where it materially outperforms ipTM.

#Impact

The broader claim in this work is conceptual: structure predictors are implicitly energy-based models, and reading their confidence heads as such recovers a better-grounded objective than the ad hoc metrics the design field has converged on. That reframing is portable to any pipeline currently optimizing a confidence score. Two limitations are important. All results are computational — the authors state explicitly that experimental validation is still needed to confirm the effects of energy-based binder design — and the preprint has not been peer reviewed. The code repository carries no license file, leaving reuse terms undeclared, and no model card or data card accompanies the release; the method itself ships no new weights, depending entirely on the existing frozen structure-predictor checkpoints.