USP-ddG: a unified structural paradigm with data efficacy and mixture-of-experts for predicting mutational effects on...

Accurately estimating changes in binding free energy (ΔΔG) is critical for understanding protein-protein interactions (PPIs) and guiding rational protein design. Recent deep learning methods have achieved notable progress by pre-training on large-scale structural data. While some approaches explore structural flexibility through energy-based sampling or generative modeling, these strategies typically involve substantial computational cost and overlook data efficacy in terms of training data organization.

We present USP-ddG, a unified structural paradigm for ΔΔG prediction built on a dual-channel architecture. The model incorporates three complementary components: (i) an inverse folding-based log-odds ratio, (ii) the empirical force field FoldX capturing side-chain packing energetics, and (iii) a geometric encoder that leverages Gaussian coordinate perturbation as a regularization strategy to improve robustness. To enhance representation capacity, we introduce a framework that integrates feed-forward network (FFN) and Mixture-of-Experts (MoE) to model domain-invariant and -specific features, respectively. We further propose CATH-guided Folding Ordering (CFO), a data efficacy strategy that organizes samples to mitigate catastrophic forgetting and data distribution bias. USP-ddG consistently outperforms existing state-of-the-art methods on the SKEMPI v2.0 benchmark, including the challenging hold-out CATH test set. It achieves superior accuracy on both single- and multi-point mutations and demonstrates strong performance in antibody affinity optimization against H1N1 and HER2, and in predicting the impact of SARS-CoV-2 variants on hACE2 binding. Ablation studies confirm the contribution of each component. These results highlight USP-ddG as a robust and data-efficient framework for modeling mutational effects on PPIs.

USP-ddG is available at https://github.com/ak422/USP-ddG.