How Insilico Medicine Leverages Generative AI To Discover New Drugs

Insilico Medicine's generative AI platform has been leveraged by over 20 pharma companies and more than 15 external and 30 internal programs.

Most people are aware that generative AI like ChatGPT can be used to generate new text and DALL-E can be used to generate new images. But did you know that AI scientists at Insilico Medicine have been working with generative AI since 2018 to generate novel chemical compounds that could be developed into new medicines to treat diseases? Insilico Medicine's generative AI chemistry platform is one of the most exciting applications of AI in healthcare.

In 2020 scientists at Insilico Medicine launched the Chemistry42 platform. Chemistry42?is?an?automated machine learning platform for drug design capable of?finding?novel lead-like structures in?days. The platform connects state-of-the-art generative AI algorithms with medicinal and computational chemistry methodologies to generate novel molecular structures with optimized properties.

This week scientists at Insilico Medicine published a paper describing how they use generative AI to quickly design novel molecular structures that target proteins that play essential roles in disease progression.?The paper, entitled Chemistry42: An AI-Driven Platform for Molecular Design and Optimization was published this week in the Journal of Chemical Information and Modeling.

“Chemistry42 is an active learning system that relies on 42 generative algorithms that have been pre-trained to design drug-like molecular structures. They draw from a variety of molecule representations, base algorithms, and strategies to explore the chemical space thoroughly. The system benefits from ongoing partnerships with pharma companies whose feedback has strengthened and validated its performance and results over time.”?

Petrina Kamya, PhD, Head of AI Platforms, Insilico Medicine

Generative Models

The generative models include generative autoencoders, generative adversarial networks, flow-based approaches, evolutionary algorithms, and language models, among others.?A major advantage of the system is its customizable reward function. As the molecular structures are generated, they are dynamically assessed using the reward function and 3D physics-based modules. Each module scores the generated molecules and together with the generative algorithms, optimize those molecules that are most likely to succeed in terms of potency, metabolic stability, synthetic accessibility and more. The novel molecules are further ranked based on their ADME and selectivity profiles.?

Seminal Papers Published

Insilico has published a number of seminal papers in the field of generative chemistry. This paper focuses on two specific case studies. The first, published in Nature Biotechnology in 2018, relied on the earliest incarnation of Chemistry42, known as the GENTRL model. Trained on the ZINC dataset and then further focused on DDR1 inhibitors and a publicly available kinase inhibitor dataset, the system produced 40 structures. Of these, six were selected for synthesis and experimental validation. After 35 days, the compounds were synthesized and tested in vitro to inhibit DDR1 activity, and four were found to be active.?

“As generative AI is making headlines generating fancy text and images, some of the generative chemistry tools that were first proposed by our group in 2015 with the first theoretical publications in 2016 and experimental papers in 2018 have now reached industrial strength and deep market penetration. The many generative engines are now validated with over 40 generative algorithms integrated in Chemistry42. We also have clinical-stage therapeutic programs generated with the help of this platform. I am very happy to see the Application Paper , which outlines our path from theory into practice, that the many scientists using Chemistry42 can now cite when publishing their AI-generated molecules.”

Alex Zhavoronkov, PhD, Founder and CEO, Insilico Medicine

The second case study is from Insilico’s recent application of an AlphaFold predicted protein structure to the Chemistry42 platform to generate a novel inhibitor for CDK20, a promising drug target for hepatocellular carcinoma, the most common type of primary liver cancer. In total, 8,918 molecules were designed, and 54 that had unique scaffolds with diverse hinge binder profiles were prioritized. A hit molecule was identified, and two compounds displayed strong potency for the intended target in a second round. The research was done in partnership with AI experts from the University of Toronto’s Acceleration Consortium and published in the journal Chemical Science. The findings demonstrated how AI systems can work together to produce novel therapeutics where structural data is limited.?

Subscribe and Comment

I'm interested in your feedback - please leave your comments. To subscribe please click subscribe at the top of this article.

Copyright ? 2023 Margaretta Colangelo. All Rights Reserved.

This article was written by?Margaretta Colangelo. ?Margaretta is a leading AI analyst consulting at Insilico Medicine. She serves on the advisory board of the AI Precision Health Institute at the University of Hawai?i?Cancer Center.?Twitter?@realmargaretta