We propose a transformer graph generative AI model for de novo drug design suitable for clinical oncology applications. The model overcomes drawbacks of current models and delivers a prospective tool for addressing resistance to treatment. Compared to existing models, the model demonstrates superior performance in exploring the chemical space and generating diverse molecular structures, thus underlining its potential to advance drug discovery efforts.

Drug development has a central dogma: disease —> target —> drug, but most of pharma discovery focuses on the target and drug aspects, particularly using AI/ML methods for virtual screening and drug design. To overcome the high failure rate in drug development, it is critical to also focus on understanding the disease, including phenotypes and endotypes. Examples from hypertensive disorders of pregnancy (preeclampsia) and hypertension, in general, will be presented.

Everybody agrees that finding small molecules that can bind specifically to selected 2-3 targets (not to hundreds of anti-targets) is the ultimate drug design goal for complex diseases. While this has been limited to serendipity or to some flavour of (mostly inefficient) merging/fusing/linking approach, now AI is allowing us to dream of finding true small molecule polypharmacology. We will introduce our efforts in this line and recent implementation studies, demonstrating that such a dream is possible today.

Recent advances in machine learning software and hardware are poised to transform the “data-rich and algorithm-poor” status quo into a new era of data-driven discovery and development. What hinders digital transformation stems from most chemical information residing on high-dimensional manifolds rather than in a low-dimensional Euclidean space. We have developed manifold learning concepts and methods to capture quantum information of a molecule for predictive and generative deep learning.

Although great success has been achieved targeting GPCRs, lack of efficacy continues to be the main reason for failure in Phase II and Phase III clinical trials. To overcome issues with hard-to-target GPCRs, we have leveraged an AI-based approach to develop a new class of GPCR molecules, the allosteric uncouplers. Such allosteric uncouplers have shown efficacy in both cell-based models and in animal disease models, and thus, could potentially be used to develop future drugs.

Predicting the binding modes of ligands bound to proteins is at the heart of structure-based design. At Astex, for any drug discovery project, we typically collect many 100s of experimental (X-ray or Cryo-EM) structures of protein-ligand complexes. Here we will discuss how we use these rich datasets of structures, combined with machine-learning approaches, to make high-quality predictions of the binding modes of newly designed compounds.

We present the discovery, biochemical, structural, and pharmacological characterisation of the noncovalent WRN helicase inhibitor, HRO761. The simultaneous optimisation of permeability and lipE of a beyond-rule-of-5 lead was achieved by chameleonic transformations identified with the help of a physic-based digital assay. Despite an MW of 702 Da, HRO761 has high permeability, low lipophilicity, and low clearance, leading to excellent PK properties.

The integration of Machine-Learning (ML) and Deep-Learning (DL) in drug discovery has garnered significant attention for its potential to accelerate early discovery. ML/DL models can be leveraged in the costly hit-to-lead and lead-optimisation phases of early research. We will explore DL approaches to help molecular design, and the move towards generative methods. Underlying issues such as bias in data and usefulness of generative methods, especially in 3D will be discussed.

Leveraging protein structural data for structure-based drug design (SBDD) of small molecules provides critical insights that are otherwise unavailable. In this talk, I present approaches to integrating protein structural data into generative molecular design, comparing several methods and discussing recent trends in the use of AI/ML for generative SBDD. Several conceptual and practical challenges in current methods for designing drug-like molecules are highlighted. Finally, a successful application of protein-guided generative molecular design to identify novel adenosine A2A receptor ligands is showcased, emphasising how the choice of protein structure influences the chemistry explored.

In this talk, I will discuss methods we have been developing to aid in the selection of macrocyclic peptides from mRNA Display screens for therapeutic applications. These methods involve the use of large language models and other machine learning techniques and provide a complement to traditional counts-based next-generation sequencing approaches.

Our team has developed a retrosynthesis algorithm that combines AI/ML with curated chemical knowledge to address drug discovery challenges. It integrates hybrid retrosynthesis models, feasibility estimations, and path optimization to predict synthetic routes rapidly. Supporting diverse molecular inputs and customizable constraints, it evaluates pathways by steps, cost, and difficulty. Achieving parity with experts in 68% of cases, the advanced methodology accelerates discovery, balancing speed, accuracy, and adaptability.