Computational drug discovery often makes use of screening methods to prioritize virtual libraries, selecting from amongst all available ligands only those most likely to have good efficacy and ADMET-related properties. In recent years, the size of these virtual libraries has greatly expanded by including not just already-available ligands but also ligands expected to be readily synthesizable from already-available building blocks. Within Janssen, we are expanding this space even further by identifying readily synthesizable building blocks as starting points for virtual libraries. This presentation will discuss early work in this area, recent successful implementations, and future-looking efforts.

Generative models for structure-based molecular design hold significant promise for drug discovery, however, data sparsity and bias are two main roadblocks to the development of 3D-aware models. Here we present a first-in-kind protocol for bias control and data efficiency, combining large 2D medicinal chemistry data resources with datasets of 3D ligand-protein complexes. We will illustrate how this framework is used to deliver augmented interactive drug design to modelers and medicinal chemists at Astex.

Exscientia's motivation to produce better drugs faster has led to the pioneering approach of patient-first, AI-driven drug discovery, design, and development. This presentation will demonstrate how automation and multiple data types are incorporated in every step of our end-to-end pipeline and will showcase our variety of proprietary toolkits and how they are integrated into a design and optimisation workflow.

RNA-binding proteins (RBPs) are paramount effectors of gene expression, and their malfunction underlies the origin of many diseases. However, therapeutically targeting RBPs with small molecules has proven challenging due to highly polar orthosteric ribonucleic interactions and a lack of lipophilic cavities indicative of druggability. We present an integrated screening campaign integrating fragment-based differentiable design with molecular dynamics to discover a cryptic site enabling the discovery of a novel non-functional RBP binder. We demonstrate the physics-driven optimization of this hit molecule to induce the proximity of an E3 ligase, facilitating the proteosome-specific degradation of an RBP and restoring a tumor-suppressive miRNA.

High-throughput screening of RNA-targeting small molecules remains a challenge due to the lack of RNA-focused small molecule libraries. We applied AI-assisted technology to select a suitable small molecule library and carried out compound screening using ALIS. The increased hit rate was significantly higher compared to our previous campaigns, which demonstrates that AI-driven compound selection strategies can accelerate RNA-targeted small molecule drug discovery.

Atomic AI has developed PARSE, the Platform for AI-driven RNA Structure Exploration, which can locate 3D structures at unprecedented speed and accuracy in disease-relevant RNA targets. PARSE builds on our work featured on the cover of Science, and involves a tight integration of high-throughput wet-lab experiments and cutting-edge artificial intelligence capabilities. Through this data-driven approach, Atomic AI is enabling and pursuing drug discovery against undruggable targets.

Rapid identification of a first-in-class or best-in-class preclinical drug candidate against a new target, new mutant, new functional state was always an important objective for computational drug discovery. However, even more challenging is screening for candidates with particular fine-grained specificity, target and property profile which may include multiple desired and undesired activities and properties. To add insult to injury, the screening of billions of compounds for candidates with those complex profiles present an extra barrier.  Methods, models, implementations and applications will be discussed.  

DNA-encoded libraries (DELs) are used for rapid large-scale screening of small molecules against a protein target. At Anagenex, we employ machine learning techniques that incorporate nuances of chemical data to create improved deep learning these models. These models can be extended to screen external molecules and predict their binding affinity to a target of interest.

Drug-drug interactions (DDI) are key for safety treatments. Drug metabolizing enzymes (DME) and drug transporters strongly influence the drug disposition and can be involved in DDI of a large number of drugs. We will present an AI approach integrating structural bioinformatics and machine learning methodologies to predict interactions of drugs with DME and drug transporters. We focus on the DME: Phase I (cytochrome P450), Phase II conjugate enzymes, sulfotransferase and UDP-glucuronosyltransferase, and on the BCRP transporter. Such AI approaches can improve the prediction of DDI in clinical practice and drug development pipelines.

Drug discovery is challenging for many reasons. We developed a flexible toolkit combining best-in-class capabilities in physics-based simulations and machine learning to enable our discovery projects. First, we describe the QUAISAR platform, which combines quantum mechanics, molecular dynamics, and machine learning to predict structure-activity relationships. We then show an example of how molecular simulations on massive GPU resources predict complex biological motions and reveal strategies for small molecule allosteric modulation. 

ADME properties are important considerations for drug discovery. Machine learning models were developed for predicting the ADME properties of small molecules. These models existed for more than a decade and were used on various occasions in lead optimizations. We will discuss the development of these models and the use cases in Genentech.