AI/ML for Drug Discovery

Expanded insights into molecular pharmacology and diverse ways to modulate function of wild-type or mutated protein complexes have led to advanced and much more specific 3D/AI predictive models. We present a pipeline that identifies binding pockets, explores vast chemical spaces of synthesisable compounds by accurate GPU-accelerated docking, and annotates hits with thousands of properties for multi-parameter re-ranking. This procedure led to several successful drug leads currently in clinical trials.

The accelerating pace of drug discovery demands computational frameworks that can navigate increasingly complex chemical and biological spaces while balancing scale, speed, and molecular accuracy. We present an integrated multi-stage discovery framework that combines artificial intelligence, physics-based simulation, and multimodal biological data to support decision-making from target prioritisation through hit finding and lead optimisation.

We present an integrated computational framework for CMG helicase inhibitor discovery that combines generative AI, quantum optimisation, and large-scale molecular simulations. As a proof of concept, we apply this approach to CMG helicase, an emerging vulnerability in Ras-driven and extrachromosomal DNA-enriched cancers. Application to CMG helicase shows how combining physics-based simulations, AI-driven design, and quantum algorithms can accelerate the discovery of therapeutics targeting DNA replication in the most resistant cancers. This framework is designed to be generalisable across targets, providing a scalable strategy for efficient ligand discovery and optimisation.

Although recent advances in computational peptide design have demonstrated that passive permeability is achievable, simultaneous design of permeability and potency remains a major challenge. We present a physics-based generative-AI framework for the de novo design of cyclic peptides that exhibit both passive membrane permeability and high target affinity. Specifically, we describe the application of Menten AI’s design platform to generate cyclic peptides against a difficult protein–protein interaction target. The resulting molecules demonstrated robust permeability in PAMPA assays (P_app >10^-6) and nanomolar potency, highlighting the potential for AI-driven peptide design to accelerate the development of next-generation peptide therapeutics.

Recent advances in protein structure prediction have transformed our ability to model individual proteins, yet predicting the structures and binding affinities of protein–ligand co-complexes remains limited due to limited experimental data and current computational approaches. OpenBind seeks to enable a step-change in protein–ligand modelling by substantially expanding paired structure–affinity measurements. In this talk, I will discuss OpenBind and advances in computational methods for more accurate and reliable binding prediction.

OpenFold3 is a fully open-source co-folding model for predicting protein–ligand and biomolecular complex structures. Unlike closed systems, it releases code, trained weights, and the full training dataset under Apache 2.0, making it auditable and freely extensible. This talk presents OpenFold3's performance on discovery-relevant tasks, the case for open model development, and how academic and industry teams can adapt it to their pipelines.

We present an integrated platform combining AI with structural proteomics to inform the rational design of protein degraders. The talk highlights specific strategies for optimising lead compounds, featuring a deep dive into preclinical data from Protai's KAT6 degrader program.

The integration of AI/ML and advanced computational methods has become a cornerstone of modern drug discovery. Here, we present our evolving computational design toolbox for protein degraders, leveraging 3D-structure prediction and AI/ML to accelerate discovery and optimisation. We discuss practical opportunities and challenges when deploying these approaches for this emerging and mechanistically complex therapeutic modality.

AI-driven drug discovery is increasingly constrained by access to diverse, high-quality proprietary data. Co-folding is one example where public protein–ligand datasets remain limited, while valuable industrial-grade data is distributed across pharmaceutical organizations. This data bottleneck is being addressed by the AI Structural Biology Network, the largest pharma-to-pharma federated AI network hosted by Apheris. The federated OpenFold3 model showed strong improvements on interface-focused metrics and performance in biologically challenging settings.

A lot of the activity in AI for biology is clustered around designing binders. The model families we are using are trained on 2-4 orders of magnitude less data than even a modest LLM, and while they’ve made progress they often still struggle the further from their training distribution. Furthermore, a binder isn’t necessarily a hit let alone a lead. Big labs are betting on automated lab-in-the-loop to cross the data chasm. I’ll talk about what I think is a plausible recipe for true foundation models of biology.

Orakl bridges artificial and biological intelligence to derisk drug development, focusing on target identification and clinical trial design. Our platform is built on a proprietary longitudinal oncology dataset integrating six modalities of multimodal data linked to drug response, matched with patient-derived organoids. Our AI models achieve 89% predictive accuracy on clinical trials, enabling novel target discovery, responder subgroup identification, and combination therapy optimization—delivering measurable clinical risk reduction for pharma partners.