Comparing Selectivity-Aware Generative AI and Library Screening in a Virtual DMT Cycle

Amit Kadan, Erika Lloyd, Andrew Wildman, Leo Zhang, Steffen Ridderbusch

Keywords: Generative AI, De novo molecular design, Multi-objective optimization, Drug discovery, Target affinity, Synthetic paths, Selectivity, Experimental feedback

TL;DR: Simulated a design make test cycle with experimental feedback, optimizing for potency and selectivity.

Abstract:

The rise of generative AI for $\textit{de novo}$ molecular design has fueled optimism regarding our ability to rapidly design compounds that yield favorable experimental outcomes. To align computational predictions with practical drug discovery requirements, generative workflows must expand beyond optimizing solely for primary target affinity. Generative models should optimize selectivity, which ensures efficacy and safety; prioritize synthetic accessibility, which ensures that compounds can be created in the lab; and directly integrate experimental data, which ensures the properties that guide the generative model reflect reality.
In this work, we present ALPAQAFlow, a generative framework that optimizes compounds by balancing on-target affinity and off-target selectivity, aiming to widen the therapeutic window. Our framework also enhances predictive accuracy by conditioning on experimental values, grounding heuristic scores with empirical data. Lastly, ALPAQAFlow extends recent advancements in reaction-pathway-based generation to ensure the resulting molecules are explicitly synthesizable. We put our platform to the test in a simulated drug discovery campaign, using Boltz-2 as a proxy for experimental affinity. Starting with a commercial library, we evaluate a selection of molecules with our proxy, then use the yielded IC$_{50}$ values to train surrogate models. These models are used to guide ALPAQAFlow and to select highly potent and selective compounds. We show that generated molecules exhibit improved binding profiles relative to the best commercial compounds. This work demonstrates how generative AI can effectively balance complex design parameters $\textit{in silico}$, aligning computational outputs more closely with the requirements of downstream experimental validation.

Submission Number: 105