DASH: A Pocket-Aware and Objective-Aware Framework for Million-Scale Structure-Based Molecular Generation

Abstract

Structure-based molecular diffusion models have shown considerable potential for de novo drug design. However, their practical use in million-scale candidate-library construction remains limited by fixed target-agnostic inference settings, insufficient support for objective-aware molecular prioritization, and limited integration of scalable property evaluation with standardized library production. Here, we present DASH, a pocket-aware and objective-aware framework for converting protein-conditioned diffusion outputs into property-tagged molecular libraries for computational prioritization. DASH combines pocket-complexity-aware sampling, configurable objective-aware molecular scoring, and a scalable production layer. The sampling strategy adjusts inference effort according to geometric and physicochemical features of the target binding pocket. The Objective-Aware Quality Module (OQM) filters and reranks generated molecules using configurable descriptors, desirability functions, and scoring profiles. The production layer supports million-scale execution through asynchronous GPU–CPU processing, streaming output, molecular scoring, and SDF annotation. We evaluated DASH through pocket-complexity analysis, OQM profile analysis, execution benchmarks, and multi-GPU/multinode scaling experiments. The results show that DASH can adapt inference effort across pockets, rerank generated libraries under different design objectives, and produce standardized molecular libraries for downstream analysis. An EGFR-oriented case study further demonstrates how DASH-generated libraries can support downstream computational prioritization and representative candidate selection. Together, these results demonstrate that DASH extends protein-conditioned diffusion models from raw molecular generation toward practical-scale, objective-aware candidate-library construction and computational hit prioritization.