Active learning cycles in drug discovery are boosting hit finding in terms of true positive rates and generating diversity by means of screening ultra large libraries or using generative modelling. We are exploring to further boost their performance by adding a learning  step based on affinity predictions through Protein Language models, significantly speeding up the process and allowing to implement it in parallel to dozens of targets.

The use of observed or predicted protein structures provides new targets, i.e. active sites, for drug development. AI/ML methods further support drug discovery and/or virtual screening but provide limited understanding of why large protein structures are preserved through evolution in support of small active sites. The ability to decipher this enigma presents new opportunities for identifying novel targets that more specifically impact biological function.

Next-generation sequencing (NGS) accelerates lead generation and optimization in antibody discovery by enabling high-throughput profiling of immune repertoires. It identifies diverse candidate sequences, tracks clonal evolution, and informs affinity maturation. Advanced clustering and machine learning techniques refine downselection, preserving diversity while reducing redundancy. Integrating NGS with computational pipelines enhances the efficiency of hit-to-lead workflows, improving the selection of high-affinity, developable therapeutic antibodies.

Target evaluation is a critical step in the drug discovery process that can be advanced through a combination of physics-based and machine-learning approaches. The use of large language models for rapid literature review to identify potential drug targets for a given indication will be discussed. Next, computational hot-spot mapping using protein structures and AlphaFold models to assess binding site druggability will be described. Finally, the machine learning-enabled prediction of antibody-antigen binding for the assessment of antibody targets will be presented.

Accurately modeling antibody–antigen and TCR–MHC interactions remains a major challenge for current structure prediction methods. We present an integrated approach that augments stochastic sampling from AlphaFold-type models with physics-based refinement using Fast Fourier Transform (FFT) docking, combined with machine learning architectures. This hybrid strategy enables accurate modeling of immune recognition interfaces and is particularly well-suited to biologics discovery. We demonstrate the utility of this method through blind predictions in the recent CASP/CAPRI protein interaction experiment, achieving high-precision models for antibody–antigen and TCR–MHC complexes. 

Using our generative design system JAM, we de novo design hundreds of VHH antibodies against the GPCRs CXCR4 and CXCR7 with picomolar to low nanomolar affinities, high selectivity, and strong early-stage developability. Many act as potent antagonists, and strikingly some are CXCR7 agonists—the first antibody agonists for this receptor and the first computationally designed GPCR antibody agonists—one matching the potency of CXCR7’s natural ligand SDF1a.

Epilepsy encompasses a set of complex, multifaceted disorders presenting a large panel of disease symptoms. We developed and validated a computational pipeline to measure behavioral phenotypes of an epilepsy animal model in an unbiased manner. We leveraged open-source ML algorithms to automatically process long-term video data and uncover behavioral fingerprints. We analyzed the usage of behavioral fingerprints across treatment groups and explored the associations between behavioral transitions and seizure events.

We integrate predictive and generative AI, computational chemistry, and computational combinatorial approaches to design and optimize small molecules targeting resistant Ras-driven tumors with complex resistance mechanisms. Through this multidisciplinary approach, we aim to accelerate the design of next-generation inhibitors and the molecular expansion of existing libraries, followed by rigorous computational optimization and experimental testing to enhance efficacy, reduce toxicity, and overcome therapeutic resistance.

Few organizations apply a systematic approach to elucidating disease biology, discovering and characterizing novel targets. Systematically applying target discovery and validation strategies to a comprehensive corpus of scientific knowledge can yield novel targets and accelerate target discovery and validation.

In this study, we employed scRNA-seq to uncover the complex phenotypic landscape of early T-cell precursor acute lymphoblastic leukemia (ETP-ALL). The computational analyses of gene programs revealed intricate interplay between oncogenic states and immune evasion programs that drives the complex phenotypic landscape in ETP-ALL. The cellular states and transitions identified not only improves our comprehension of disease pathogenesis but also aids in the identification of potential therapeutic targets.