Protein structure predictors implicitly define binding energy functions
Track: Machine learning: computational method and/or computational results
Nature Biotechnology: Yes
Keywords: Energy-based models, protein-protein binding, protein-RNA binding, protein folding models
TL;DR: Protein structure predictors can be reinterpreted as energy-based models for unsupervised binding energy prediction.
Abstract: Estimating binding energies is vital for drug discovery, yet supervised methods are hampered by limited experimental data. Recent protein structure predictors (e.g. AlphaFold3) offer unsupervised alternatives via confidence metrics that correlate with binding energies. However, these metrics operate on a fixed scale, limiting their ability to capture fine-grained energy differences. Leveraging the Joint Energy-based Model (JEM) framework, we show that protein structure predictors implicitly define an energy function, and we introduce two new energy-based models derived from the confidence head. Our EBMs consistently improve binding energy prediction, outperforming both traditional confidence metrics and unsupervised baselines, and demonstrate that structure prediction models can be repurposed as powerful unsupervised energy predictors.
Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.
Presenter: ~Wengong_Jin1
Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.
Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.
Submission Number: 44
Loading