AQBioSim

SandboxAQ has built IDOLpro, a guided generative chemistry AI that can design novel optimal drug-like molecules. IDOLpro is an LQM trained on a combination of public data as well as physics-based simulations that creates novel drug molecules with both high binding affinity to a target protein and optimized synthetic accessibility. IDOLpro leverages AWS infrastructure, combining diffusion models with multi-objective optimization to guide molecule generation. Using benchmark datasets like CrossDocked and Binding MOAD, the model achieves 3.4x better binding affinity compared to state-of-the-art methods and has, for the first time, generated compounds that outperform even experimentally validated molecules. Each molecule can be generated and refined in minutes on a single GPU, compressing months of drug discovery and underscoring the efficiency and transformative potential of guided, cloud-based AI for accelerated drug design.
[AWS Blog] [Scientific Manuscript]

Structure-based methods have become a cornerstone of modern drug discovery, driving significant advancements in the field. Recent breakthroughs in free energy binding prediction and the integration of AI solutions are reshaping the landscape of medicinal chemistry. As the pharmaceutical industry evolves, the demand for innovative solutions to address long-standing challenges has never been greater. Large Quantitative Models (LQMs) are at the forefront of this transformation, enabling researchers to explore chemical space with unparalleled speed, accuracy, and efficiency. The power of SandboxAQ’s AQFEP, an innovative absolute free energy perturbation solution, lies in its ability to rapidly and cost-effectively explore enormous chemical spaces to select promising ligands (i.e. drugs) for further experimental investigation. Without a reference molecule requirement, AQFEP unlocks first in class molecules in an integrated workflow to efficiently screen large and diverse chemical libraries virtually, combining active learning with a rigorous physics-based scoring function. We validated the performance of the approach in the ranking of structurally related ligands, virtual screening hit rate enrichment, and active learning chemical space exploration. Notably, we have disclosed the largest reported collection of free energy simulations to date, further demonstrating the power of this approach [Scientific Manuscript]. AQFEP is already delivering exceptional results in challenging, historically "undruggable" targets, particularly in neurodegeneration and oncology, where it has shown remarkable success in hit identification and lead optimization.