CadreAI: Agentic AI for Semantically Decoupled Protein–Drug Design and Optimization

De novo drug design must balance global drug properties with local structural determinants of target specificity. However, existing methods often reduce both objectives to local geometric optimization, relying on blind, reward-driven perturbations that can disrupt essential scaffolds and compromise overall drug-likeness. To resolve this destructive optimization trap, we propose CadreAI, a semantically decoupled agentic de novo drug design framework. CadreAI explicitly decomposes the entire drug design process into two independent subtasks: target-oriented semantic planning and plan-conditioned molecular generation. This strict separation mitigates potential task interference, allowing each module to focus exclusively on its dedicated objective without mutual disruption. Crucially, the strategic logic of how to formulate these target-oriented plans is driven by Simulation-Guided Insight Consolidation (SGIC). To eliminate trial-and-error guessing, this persistent memory mechanism translates transient numerical docking feedback into reusable textual heuristics, allowing the agent to continuously learn and logically deduce subsequent structural edits. Evaluated on the CrossDocked2020 and DrugGen benchmarks, molecules generated by CadreAI achieve superior docking performance with suitable drug-likeness and synthetic accessibility. Additional Boltz-2 evaluations provide orthogonal computational validation of predicted structural plausibility and binding affinity. These results demonstrate that semantically decoupled reasoning offers a highly controllable and practical route for in silico therapeutic design.