Boltz-1 as a force field -- why co-folding models struggle with learning physics and how to fix it

Modern co-folding models have achieved remarkable success in biomolecular structure prediction. However, their ability to generalise to \mbox{out-of-distribution} examples and to capture physical laws remains limited. These limitations may stem from either data or architecture; here, we focus on the latter by examining whether the training objectives and architectural choices of co-folding models hinder the learning of physical laws. Drawing on insights from physics-constrained Machine Learning Interatomic Potentials (MLIPs), we investigate the expressiveness of attention-based modules as implemented in the co-folding models. We evaluate the exemplary co-folding model \mbox{Boltz-1} as an MLIP and find that it underperforms on energy surface learning. Our analysis shows that accurate energy learning requires inter-atomic distances to be encoded appropriately in the attention pair bias, whereas \mbox{Boltz-1} constructs these features in a way that fails to support this task. Based on these insights from MLIPs, we introduce simple architectural modifications, including a revised pair bias encoding, and show that they significantly improve energy landscape learning.