The accelerating pace of drug discovery demands integrating diverse computational methodologies to navigate increasingly complex chemical and biological spaces. We present a multi-tiered framework designed to accelerate the discovery pipeline, from initial target prioritization to the refinement of lead candidates. At SandboxAQ, we address the inherent trade-offs between computational throughput and molecular accuracy by combining high-capacity machine learning architectures with rigorous physics-based simulations.

Our approach leverages Large Quantitative Models (LQMs) to bridge gaps across the different stages of the discovery process. For target identification and selection, we utilize data-integrative models, including knowledge graphs (KGs) and omics-based analytics, to uncover novel disease-relevant nodes. In the hit-finding phase, we employ next-generation virtual screening protocols, incorporating AI-driven structure prediction, and ML-enhanced SBDD methods, to explore diverse chemical modalities and ultra-large libraries.

Furthermore, we demonstrate the transition to lead optimization through the application of advanced binding free energy perturbations (FEP) and alchemical methods. These techniques provide a granular understanding of the thermodynamic and kinetic profiles of ligand-protein interactions, particularly in challenging systems such as neurodegenerative disease targets and unconventional molecular frameworks.

We will share several case studies highlighting the successful application of this integrated platform across various therapeutic areas and modalities. These examples will illustrate how the synergy between generative AI, statistical learning, and physics-based modelling can significantly reduce cycle times and improve the efficiency towards clinical candidates.