Venus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models

In-silico prediction of protein mutant stability, measured by the difference in Gibbs free energy change (∆∆G), is fundamental for protein engineering. Current sequence-to-label methods typically employ the two-stage pipeline: (i) encoding mutant sequences using neural networks (e.g., transformers), followed by (ii) the ∆∆G regression from the latent representations. Although these methods have demonstrated promising performance, their dependence on specialized neural network encoders significantly increases the complexity. Additionally, the requirement to individually compute latent representations for each mutant site negatively impacts computational efficiency and poses the risk of overfitting. This work proposes the Venus-MAXWELL framework, which reformulates mutation ∆∆G prediction as a sequence-to-landscape task. In Venus-MAXWELL, mutations of a protein and their corresponding ∆∆G values are organized into a landscape matrix, allowing our framework to learn the ∆∆G landscape of a protein with a single forward and backward pass during training. Besides, to facilitate future works, we also curated a large-scale ∆∆G dataset with strict controls on data leakage and redundancy to ensure robust evaluation. Venus-MAXWELL is compatible with multiple protein language models and enables these models for accurate and efficient ∆∆G prediction. For example, when integrated with the ESM-IF, Venus-MAXWELL achieves higher accuracy than ThermoMPNN with 10× faster in inference speed (despite having 50× more parameters than Ther-moMPNN). The training codes, model weights, and datasets are publicly available at https://github.com/ai4protein/Venus-MAXWELL.