As generative models grow more complex, discerning their actual contributions to drug design becomes challenging. This presentation assesses the true impact of these models, focusing on molecule synthesizability and 3D structural integration. We critically analyze limitations from small datasets and the models' tendency to infer patterns without genuine extrapolative power. Emphasizing the need for clarity in evaluation, we propose strategies for establishing meaningful benchmarks to ensure generative models deliver tangible improvements in drug discovery.

Structure-based generative modelling (SBGM) represents a change of paradigm in drug discovery, from virtually screening ultra-large chemical libraries to virtually growing molecules with desired physicochemical and ADMET properties directly inside the protein cavity. In this talk, the SBGM platform developed at Chemotargets (named D3) to generate novel synthetically feasible drug-like molecules for protein targets will be introduced. Both retrospective fragment-to-drug examples and prospective fragment-growing case studies that resulted in the identification of novel bioactive chemical matter will be presented.

Advanced AI/ML modeling makes it possible to efficiently and specifically discover new oral therapeutics for chronic disease. Montai’s CONECTA platform integrates proprietary bioassay data and machine-learning models built on repositories of chemistry and multi-factorial biological data on complex disease pathways and drug-like properties. The integration of multi-modal modeling into the candidate selection process efficiently identifies those with the highest probability of becoming successful drugs to address significant unmet needs in human health.

At Expert Systems, we firmly believe in temporal validation for all AIML models as a means to significantly reduce predictive errors and hallucinations. We will discuss our Target Druggability model, which, trained on data current as of 2017, correctly predicts nearly 80% of novel drug targets (2018-2024). We will also discuss using agentic LLMs that conduct automated reasoning to suggest novel therapeutic indications for existing intellectual property.

The exponential increase in biomedical data offers unprecedented opportunities for drug discovery, yet often overwhelms traditional data analysis methods. Here we introduce a framework for autonomous artificial intelligence (AI)-driven drug discovery that integrates knowledge graphs with large language models (LLMs) and which is capable of planning and executing automated drug discovery programs on a massive scale while providing details of its research strategy, progress, and all supporting data.

Computational methods are reshaping how we discover and prioritize drug targets. This talk introduces the key role of target discovery in drug development and reviews some of the computational approaches, from traditional methods to statistical machine learning models. A real-world case study to investigate novel targets for obesity will illustrate how integrated techniques can prioritize therapeutic targets effectively.

Endometriosis and alternative sources of non-hormonal contraception are neglected and challenging issues associated with women's health. Today, I will discuss how we leverage multiomics data and AI to identify novel targets implicated in endometriosis and how we contribute to the challenge of designing novel non-hormonal contraceptives using AI.

Addressing certain toxicities such as drug-induced seizures and steroidogenesis may require complex models to enable predictions. We now describe how we have generated large numbers of machine-learning models for either individual toxicity targets or large language models (ProtBERT, MolBART) to enable predictions for prospective testing.

Enchant is a multimodal transformer designed to break the data wall between lab and clinic by predicting clinical outcomes from preclinical data. Trained on diverse, heterogeneous datasets, it addresses the scarcity of clinical data by leveraging stage discovery data of various modalities. Enchant enables informed decision-making earlier in the drug development pipeline, accelerating and de-risking the path from discovery to the clinic.

We used a bioinformatics approach to analyze the immune repertoire of a llama immunized with a target antigen, prioritizing those antibodies inferred to be clonally expanded. We employed AlphaFold2 to dock these antibodies, selecting those predicted to bind an epitope of interest. This in silico strategy yielded a 60% hit rate in identifying binders. We found many binders that effectively inhibited binding of the target to its receptor.