AI in drug discovery 2026
Date & Time - March 9th-10th, 2026
Location - London, United Kingdom
Schrödinger is excited to be participating in the AI in drug discovery 2026 conference taking place on March 9th – 10th in London, United Kingdom. Join us for a presentation by Pieter H. Bos, Principal Scientist II at Schrödinger, titled “Overcoming Data Scarcity in Lead Optimization: A Physics-Guided Generative AI Workflow for Selective p38α/MK2 Molecular Glues.”
MAR 9 | 11:30am Overcoming Data Scarcity in Lead Optimization: A Physics-Guided Generative AI Workflow for Selective p38α/MK2 Molecular Glues
Speaker:
Pieter H. Bos, Principal Scientist II, Schrödinger
Abstract:
Generative AI in lead optimization is frequently bottlenecked by data scarcity. We present a workflow synergizing de novo design (AutoDesigner) with physics-based simulations (WaterMap/FEP+) to train generative models in low-data regimes. We applied this framework to the design of selective p38α/MK2 molecular glues, starting from a single reference compound with no established selectivity SAR. AutoDesigner first enumerated ~1 billion permutations to define the project’s physicochemical boundaries. By training on the AutoDesigner structural prior, a high-fidelity dataset of 1,500 active learning FEP+ affinities and WaterMap scores, the Generative AI learned to architect novel, potent drug-like molecules that displace high-energy hydration sites. The model learned to architect molecules that displace high-energy hydration sites, yielding 12,000 MPO-compliant candidates, a 56-fold increase over standard filtering. This generative cycle required just 4.5 hours, bypassing the 180,000 CPU hours needed for equivalent exhaustive enumeration. Experimental validation confirmed the method’s precision, yielding compounds with pIC50s up to 10.3 and 451-fold selectivity against p38α. This demonstrates that physics-empowered Generative AI can autonomously and efficiently solve complex specificity challenges without significant experimental data.
• Overcoming Data Scarcity: New workflow that integrates de novo design with physics-based simulations, enabling the training of generative models to initiate projects with only a single reference compound.
• Targeted Structural Optimization: Through active learning on FEP+ and WaterMap data, the generative AI learned to design molecules that displace high-energy hydration sites to enhance binding potency.
• Experimental validation: the workflow solved complex selectivity challenges, yielding compounds with high potency (pIC50 up to 10.3) and 451-fold selectivity against p38α.