Designing small-molecule drugs that bind a target protein without also engaging other proteins is a central challenge in drug development. Off-target interactions are a major source of side effects and contribute substantially to clinical attrition. A new paper on arXiv addresses this issue by examining how generative models used in structure-based drug design handle selectivity.
The authors argue that while these models have become powerful tools for generating candidate molecules, their optimization objectives may not adequately account for selectivity. Instead, many approaches prioritize binding affinity to the intended target, potentially producing compounds that are potent but promiscuous.
The paper, titled "On the Selectivity of Generative Models in Structure-Based Drug Design," appears to analyze the extent to which current methods incorporate selectivity constraints. Given the abstract, it does not yet detail specific results, but it underscores a critical gap in the field: without explicit attention to selectivity, generative models may generate molecules that fail in later stages due to safety issues.
This work serves as a reminder that computational drug design must move beyond simple affinity metrics. Integrating selectivity into the generation process could lead to safer, more effective drug candidates and reduce the high failure rate seen in clinical trials.