Small molecule drug discovery is the iterative process of identifying starting compounds and improving them through multi-parameter optimization (MPO). How can we improve upon the current approach, which is still mostly driven by human experts with some assistance from computational approaches?  I’ll present examples of successful applications of machine learning (ML) and more recent deep learning (DL). These include prediction of DMPK and physicochemical properties, comparison of ML and DL prediction of log D, and a NNP (neural net potential) model to calculate DFT-level strain-energy for receptor-bound ligand conformation. 

In an evolving landscape of in silico chemical intelligence and machine learning, computer-aided synthesis can accelerate breakthroughs in drug discovery research. SYNTHIA™ retrosynthesis software is revolutionizing the way chemists design pathways to complex targets by harnessing the power of artificial intelligence with an expert-coded database of advanced organic synthesis rules to augment chemists’ expertise. Discover how this innovative cheminformatics tool is being used at the bench.

We derive machine learning (ML) models from over 50 fragment screening campaigns. Critically, our dataset includes true inactives as well as actives and our ML methodology produces interpretable models that we validate against expert annotations. We show that, given a high-quality training set, ML does not only generate models that separate binders from non-binders, but also accurately identifies which parts of a fragment drive its binding against the target.

Iktos has developed a retrosynthesis platform called Spaya, and a high-throughput API running on Spaya’s algorithmic engine. Integration of the Spaya API into a molecule generator leads to the generation of synthesizable molecules without penalizing the other objectives of the generator. Integration of synthetic accessibility is an enabling step to be able to take advantage of the full potential of generative models in real life, i.e.: obtaining synthesizable optimized molecules.

Modern cancer biology leans heavily on kinase inhibitors to probe the consequences of deactivating a particular kinase, but most commonly used chemical probes are not sufficiently target-selective for robust interpretation of the observed phenotypes. Today I will sketch out a path for rapidly designing high-quality selective kinase inhibitors; these inhibitors may be used as tools for discovery, and as starting points for drug development.

The lecture will focus on the development and application of generative models for creating novel compounds and for generating synthetic biological data with the desired properties.

Exscientia combines the strengths of AI compound design and human strategic thinking into the Centaur ChemistTM.  In this talk we outline the breadth of generative design techniques that are involved in our AI design system. Second, we show how these generative models can incorporate multimodal 2D and 3D data to enhance their efficiency. Third, we show how a range of Active Learning capabilities are used to optimally select compounds for enhancing information gain and thus the next cycle of design.

We have developed and applied machine learning software and models for various targets for rare, neglected, and common diseases which have enabled us to create a diverse portfolio of small molecules. Several of our recent published applications for infectious diseases and Alzheimer’s will be described along with our exploration of approaches for de novo molecule design. Our goal is to progress and de-risk assets that can then be out-licensed.

SPR is a key tool in drug discovery due to the information rich nature of the data it provides, supporting chemistry efforts in hit to lead development.  But analyzing 100s-1000s of multi-dimensional data points remains a key challenge, constrained by time and expertise.  We will present a novel, AI-based prototype that overcomes the analysis bottleneck to enable greater support of chemistry efforts in hit to lead development with SPR data.

A large panel of AI and 3D models helps to quickly repurpose drugs to new targets or find leads with specified activity profiles composed of new or known targets. Applications to targets related to SARS-CoV-2, flaviviruses, E. Histolytica and oncology are presented.

AI has become an indispensable tool for drug discovery, yet the technology is not widely utilized by wet-lab scientists. The core limitation of existing approaches is inability to leverage small datasets and poor interpretability. Koliber is developing an AI platform for peptide discovery that enables tuning of pre-trained models to small datasets, visualization of key features and sequence profiling to identify key sites and improved variants. The platform will be demonstrated using examples from immunology.

We focus on Cytochrome P450 (CYP) responsible for the metabolism of 90% drugs and on sulfotransferases (SULT), phase II conjugate drug metabolizing enzymes, acting on a large number of drugs, hormones and natural compounds. We established an original in silico approach integrating structure-based and machine learning modeling and developed a new software DrugME to predict CYP and SULT inhibitors. This approach allowed the identification of new drug inhibitors and substrates of CYP2C9.Such AI strategy will improve the prediction of drug-drug interactions in clinical practice and will be of high-value for early drug discovery.

In developing new molecules for the clinic, we cannot avoid structural alerts in all cases. In most cases, structural alerts are only a marker of risk, and are not actually bioactivated into reactive metabolites. Machine learning, a type of artificial intelligence, is giving us new ways to understand why and when structural alerts become toxic or not.