Two case studies for AI methods in small-molecule drug discovery will be covered. First, as virtual libraries continue to grow, active learning paired with virtual screening is becoming an increasingly important technique. I’ll explain how a clever use of a traditional active learning technique allows for efficient application of virtual screening to unenumerated libraries. Second, I’ll talk about the practicalities and experience in further automating design decisions as covered in our Automated Chemical Design framework (Goldman et al., “Defining Levels of Automated Chemical Design”, 2022).

Efficient syntheses of complex small molecules involve speculative experimental approaches. The central challenge of such plans is that experimental evaluation of high-risk strategies is resource intensive, as it entails iterative attempts at unsuccessful strategies. This presentation describes a complementary strategy that combines creative human-generated synthetic plans with robust computational prediction of synthetic feasibility. This work defines how machine learning models can drive complex molecule synthesis.

State-of-the-art AI models such as large language models (LLM) have demonstrated impressive performance in chatbots and related human-interfacing tasks. In drug discovery, these models can be used for better understanding of protein structures, exemplified by recently reported folding algorithms. They can also be used to better describe therapeutic biomolecules of novel modalities, and can enable better hit-finding strategy and optimization options, which we will discuss in-depth in this talk.

In this talk, we explain how our engineering platform and our physics and informatics toolbox have enabled Exscientia to tackle challenging target product profiles, developing several candidates that have either entered the clinic or are in IND-enabling studies. This includes targeting PKC-theta, LSD1, and MALT1, solving complex design challenges for each that have eluded AI to date: kinase selectivity, brain penetration, and allosteric inhibition, respectively.

Molecular descriptors are algorithms that translate molecule structures into data types interpretable by computers. They are prerequisites to most ML/DL models for predicting molecular properties and are critical for their accuracies. BAKX Therapeutics developed a number of novel proprietary molecular descriptors incorporating deep learning and quantum chemistry, which, either standalone or combined, outperform traditional molecular descriptors at predicting ADMET properties and binding affinities.

The Recursion Operating System (OS) capitalizes on deep-learning neural network analysis of high-dimensional cellular images to enable phenotypic drug discovery. Advancement of Recursion’s phenomics platform has enabled a paradigm shift to inference-based searching of our growing map of human biology to initiate new research projects. Hit compounds identified through Recursion’s platform are subsequently optimized and progressed through virtuous chemistry SAR cycles using the same phenomics assays.

Using on-demand libraries, AI methods, structural diversity, and molecular modeling techniques, one can obtain enrichment factors that were a dream years ago. Our recipe starts with a diversity conformational search and follows with an adaptive learning procedure with virtual data augmentation from molecular modeling. Different flavors, such as when using generative modelling methods or analog/neighbour search, season compound generation, providing a diverse set of highly active molecules and an outstanding hit rate in just a few days.

Accurately predicted protein-ligand binding affinities can be a powerful tool for virtual screening of ligands. Here, I will present an approach incorporating protein and ligand dynamics into ML methods for predicting binding affinities. This approach leads to improved metrics, e.g., significant decreases in RMSE values. Additionally, I will show this method is useful for predicting affinities in cases where no structural information is available.

In recent years, we have seen significant advancements in the field of precision medicine. New genomic technologies hold great promise for the identification of actionable drug targets and associated biomarkers for several complex diseases, such as autoimmunity. Our approach uses single-cell RNAseq and machine learning to elucidate the precise cell types and gene expression programs involved in the progression of complex diseases in order to identify novel therapeutic targets.

In this talk, I will discuss new approaches that we are exploring in combining deep learning, omics, and functional data along with clinical information to predict new drug targets.

Cornerstone to the development of cancer and other age-dependent disease are the downstream influences of systemic metabolism and inflam-ageing. Relying on skin biopsy-derived fibroblasts, we developed multivariate phenotypic readouts that quantify metabolic and inflammatory rates that are specific to the individual biopsy donor at the single cell level. We found that these cell assay readouts correlate with pharmacological (rapamycin) or life-style (weight gain) interventions that had occurred in the individuals past and predict the onset age of cancers in the individuals future. Skin fibroblasts derived from cancer-free individuals from Li Fraumeni patients that were subject to years-long clinical surveillance, predicted the specific (personalized) onset age of cancers that developed years post-biopsy.

Algen is a platform therapeutics and drug discovery company using the world’s leading CRISPR and AI to uncover disease-driving RNA messages to find treatments for cancer, inflammation, and other diseases. Spun out from Nobel Laureate Professor Jennifer Doudna's Lab, Algen aims to develop the world’s smartest drug discovery decision platform and data universe to create next-generation therapeutics.