BACKGROUND: Deep mutational scanning (DMS) provides comprehensive maps of protein variant effects but remains experimentally intensive. Machine learning (ML) approaches have the potential to reduce experimental burden of DMS by predicting the functional impact of substitutions from limited data. RESULTS: We introduced a ML classifier trained on normalised DMS scores from SARS-CoV-2 main protease (Mpro) to categorise amino acid substitutions as functional (wild-type-like) or non-functional. Using brute-force feature selection, we identified minimal subsets of six substitution scores per residue that enable accurate classification of the remaining substitutions, achieving minimum (worst accuracy) scores exceeding 90%. Models including support vector machines, random forests, and logistic regression were evaluated without retraining (zero-shot prediction) against additional SARS-CoV-2 Mpro datasets and against unrelated datasets. The zero-shot performance of the models was strongest for other enzymes and more modest when applied to DMS systems that assess protein folding and/or protein-protein interactions. CONCLUSION: The results show that targeted DMS combined with ML can reduce sequencing and reagent costs while preserving classification accuracy, offering a practical route to accelerate variant effect prediction.

Future advanced AI systems may learn sophisticated strategies through reinforcement learning (RL) that humans cannot understand well enough to safely evaluate. We propose a training method which avoids agents learning undesired multi-step plans that receive high reward (multi-step "reward hacks") even if humans are not able to detect that the behavior is undesired. The method, Myopic Optimization with Non-myopic Approval (MONA), works by combining short-sighted optimization with far-sighted reward. We demonstrate that MONA can prevent multi-step reward hacking that ordinary RL causes, even without being able to detect the reward hacking and without any extra information that ordinary RL does not get access to. We study MONA empirically in three settings which model different misalignment failure modes including 2-step environments with LLMs representing delegated oversight and encoded reasoning and longer-horizon gridworld environments representing sensor tampering.