ちなみに ・・・ Chris de Graaf (Sosei Heptares) Retweeted: ※Morgan Thomas (University of Cambridge) Twitter: "Pleased to present some of my work in collaboration with Sosei Heptares at AIChem20posters". (9/18) "P39: Towards Integrating Deep Generative Models with Structure-Based Design" ・Assessing the difference between ligand-based (QSAR) or structure-based (docking) scoring functions to guide a deep generative model via reinforcement learning.
Event: ※3rd RSC-BMCS / RSC-CICAG Artificial Intelligence in Chemistry (AIChem20): Monday-Tuesday, 28th-29th September 2020. Poster Presentation 39: " Towards Integrating Deep Generative Models with Structure-Based Design": Morgan Thomas, University of Cambridge / Sosei Heptares, UK . Posters Displays: ・All posters will be displayed on Twitter, AIChem20posters from 18th September until shortly after the event.
↓ ※Artificial Intelligence in Chemistry (AIChem20) Poster Presentation 39: "Towards Integrating Deep Generative Models with Structure-Based Design" : (Morgan Thomas, , Rob Smith, Noel O’Boyle, Chris De Graaf & Andreas Bender). Introduction: Deep generative models are a promising application of advances in deep learning to de novo molecule generation. These models are typically trained on a library of exemplar molecules to learn the underlying chemistry and subsequently trained to bias molecule generation towards a desirable property space via an external scoring function and optimization algorithm. The vast majority of models only use simple molecular descriptors or ligand based QSAR models. This approach has several limitations: 1) application is restricted to data rich areas。. 2) ligand-based methods bias molecule generation to non-novel chemical space. 3) ligand-based methods evidence.
REINVENT Implementation: Here, we assess the ability of molecular docking - a structure-based technique - to direct de novo molecule generation. As proof of concept, we use the REINVENT framework and Glide docking software to compare de novo molecule generation of Dopamine Receptor D2 (DRD2) predicted active molecules. In particular, G Protein-Coupled receptors (GPCRs) - such as DRD2 - are an abundant class of protein targets that could benefit greatly from increased structure-based design input. Due to the limited availability of X-Ray crystal structures and hence, increased value in drug design projects. ・・・ Conclucion: ・Structure-based de novo molecule generation evidences less failure relative to ligand-based. ・More complex scoring functions (docking score) can be optimized by reinforcement learning - this translates to relevant chemistry. ・Ligand and structure-based approaches generate complementary chemistry but structure-based results in more diversity. ・Crucial residue interactions are learnt by the structure-based approach. ・Structure-based approaches appear as fit for purpose as ligand-based.