Autors :
Christophe Parsy2, Anna Kriukova1, Pascaline Jacquemard2, Maud Jusot1, Quentin Janet2, Nicolas Devaux1, Anthony Martinez2, Christopher Housseman1, Sautet Stephane2, Alexis Denis1, Quentin Perron1, Yann Lamotte2, Brice Hoffmann1
1 IKTOS Company, 2 ONCODESIGN SERVICES
Abstract :
In early stage hit-discovery or scaffold-hopping discovery development programs, the primary challenge is to find new chemical matter that is both novel and has exclusive Freedom to Operate (FTO), while also optimizing physico-chemical properties that are important for robust in vivo activity. This is a difficult multi-parameter optimization (MPO) problem, making the drug discovery process complex, expensive, and time-consuming. Traditional screening methods such as virtual screening of large chemical databases have limited success in solving the MPO problem due to exploring only a small section of the chemical space, which lacks diversity and novelty. To address these limitations, a 3D structure-based generative Artificial Intelligence (AI) pipeline was developed, which is focused on solving the MPO problem by identifying new, easily accessible molecules with drug-like characteristics and high activity on the protein targeted. The technology employs deep-learning based de novo design algorithms and is optimized through Reinforcement Learning, based on various factors such as diversity, novelty, quality, synthetic accessibility and 3D scores. This iterative AI-based chemical space exploration generates new chemical structures with optimal 3D scores and physico-chemical properties. Using this approach, novel active compounds for an oncology target (PIM1-kinase) have been successfully identified with high affinity, outside the chemical space of known PIM1-kinase inhibitors. Experimental results indicate good preliminary ADME properties for one of the novel PIM1 compounds, with an IC50<50nM. This represents a significant shift in tackling MPO in early stage drug discovery projects and increases the chances of success while drastically shortening the time for Hit finding.