Generative AI & Predictive Modeling

Chemical space contains 10⁶⁰ synthesizable, drug-like molecules. This enormous space is too large to search exhaustively with VHTS. Generative AI promises to find optimized molecules with fewer queries and less data. Rather than zig-zagging across the breadth of the landscape to find the peak of a mountain, generative AI can walk directly uphill. However, many popular techniques, such as Bayesian optimization and reinforcement learning, cannot efficiently navigate latent representations of chemical space. We show that gradient-based optimization is significantly more effective. Combining this with a novel, domain-specific architecture and carefully curated datasets, we demonstrate efficient generative AI for drug discovery.

Commercially available virtual, synthesis-on-demand chemical catalogs are expanding rapidly with trillions of compounds and increasingly complex chemistry, providing value throughout all pre-clinical drug development stages. However, their exponential growth poses significant challenges for traditional search-and-score methods to efficiently explore catalogs. Here, we present and experimentally validate a generative AI that efficiently designs catalog compounds with desired properties.

We present Exscientia's approach to drug discovery and the important role of generative small molecule design in the process. The discovery of a novel LSD1 inhibitor and a successful proof-of-concept study leading to the discovery of novel, bispecific anti-malaria agents are presented as applications of generative design.

Therapeutic antibody discovery for oncology is challenging due to an inability to control the epitope binding site as well as toxicity resulting from healthy tissue reactivity. These problems are overcome with machine learning design of peptides that match the sequence and structure of the target epitope. Antibody libraries are screened against the engineered peptides, resulting in efficient discovery of on-epitope binders. The antibodies are then masked with these peptides to improve their off-tissue safety profiles in the case of oncology targets.

Most out-of-the-box generative chemistry models struggle to encode the chemical properties that medicinal chemists favor during drug discovery campaigns. By fine-tuning molecular language models for specific chemical spaces, such as the kinase inhibitor chemical space, we can align generated molecules more closely with chemists' preferences.

Our team is developing a suite of LLM experts, each focused on different tasks and activities related to early drug discovery. These include PharosGPT (targets, diseases, ligands), litGPT (learn from papers), ChEMBLGPT (compounds and bioactivities) and ActivityGPT (predict bioactivity endpoints) provide LLM-based support for our projects. DrugInteLLM is the human-facing orchestra conductor that oversees our GPT-based platform for target and ligand discovery.