JamesKInner/Covalent-MAS

molecule generation, multi-objective optimization, and evidence-driven trajectory learning. Covalent-MAS is a modular architecture for iterative covalent drug design. It connects target preparation, covalent molecule generation, chemistry filters, structure-based evaluation, multi-objective optimization, and evidence-driven learning through stable Python interfaces. The architecture is model- and tool-agnostic. Generative models, property optimizers, docking engines, and external services can be integrated without changing the workflow contracts. This release includes an AutoDock Vina adapter and portable JSON/JSONL data structures for candidates, evaluations, optimization results, and design trajectories. Our covalent molecule generation pipeline improved generation quality and 3D grafting performance under the same project evaluation protocol: For molecule optimization: performance than GPT-5.6-Sol on our covalent optimization benchmark. performance in our evaluation setting. These values compare methods with the same task definitions, inputs, and evaluation protocol. Candidate structures still require downstream computational review and experimental validation. The workflow supports an iterative design loop: generative model. Each optimized molecule…