Exascale search of molecules

Drug discovery is lengthy process which costs 2.6B per approved drug on average. The cost is so high because finding a small molecule that binds to a particular target is a perilous highly uncertain process that can’t be completed in a decade in many of the cases. And even when a molecule is found it might not be doing its job very well and might very likely fail in later stages of trials. Our goal is to find much better drug candidate molecules and do it very fast.

The number of possible drug-like molecules in the Universe is comparable to the number of atoms in the Universe.

All of the cures found in the history of mankind come from sampling a tiny fraction of this space. The rest of this space likely contains cures for most of the remaining diseases if knew how to search it better.

Our goal is to develop a search technique that could reach the exascale, or 10^18 molecules.

The search in this scale would take 1000 Billion CPU years for existing physics-based methods to simulate.

We define molecules as a set of standard building blocks (think of LEGO bricks) that could be recombined together in various ways. We define search in the molecular space as search in the space of possible configurations of these blocks. Similar to how AlphaZero which uses Monte Carlo Tree Search (MCTS) to master the game of Go, our algorithm called LambdaZero uses MCTS to navigate the space of molecules

Dopamine receptor D4 is an important protein in the human brain. Searching in the larger space of molecules* results in finding molecules with more negative binding energy. More negative binding energy means stronger association between the small molecule and the protein and therefore better binding and better properties as a drug.

*Ultra-large Scale Docking for Discovering New Chemotypes, Wu at al. Nature, 2019. 

Best molecule

200M simulations

Binding Energy: -75.6

Best molecule

 LambdaZero

 Binding Energy: -85.1

Applied to the dopamine receptor LambdaZero could identify a better molecule compared to one of the largest scale physics-based simulations* ever performed. 

*Ultra-large Scale Docking for Discovering New Chemotypes, Wu at al. Nature, 2019. 

Team

Maksym Korablyov

EngineerZero

Kostiantyn Lapchevskyi

Software Engineer

Sasha Luccioni

Project Manager

Sacha Leprêtre

HPC Specialist

Shenghao Liu

Software Engineer

John Bradshaw

Software Engineer

Shivam Patel

RL Engineer

Matt Kusner

Molecule Learning  Advisor

Marvin Segler

Molecule Learning Advisor

Will Hamilton

Graph Learning Advisor

Brooks Paige

Molecule Learning Advisor

Mike Tyers

Biomedical Assays

Jian Tang

Graph DL Advisor

Doina Precup

RL Advisor

Yoshua Bengio

Director, Advisor

We are urgently applying LambdaZero to help fight the COVID19 pandemic. Please, read the message from Yoshua Bengio.

Mila is launching several research projects aimed to leveraging ML to fight COVID-19, e.g., see the preliminary list there, https://mila.quebec/en/mila-and-its-partners-rally-the-scientific-community-to-develop-novel-data-driven-solutions-to-assist-with-covid-19-outbreak/.

Today I am writing to ask for volunteers to work on our project on Deep RL for accelerated discovery of SARS-CoV-2 antiviral molecules. Because time is of the essence to fight COVID-19, the focus of the proposed research is on speeding up the training and optimization in molecular space in order to come up with candidate molecules to be tested by our biomedical partners in biological and biomedical assays (and if these work, in clinical trials). We already got very promising results,v’est p allowing us to screen orders of magnitude more molecules than existing techniques and obtain stronger binding to protein targets. We now want to focus on protein targets for COVID-19.

We need people to help on several fronts for about 4 months, including:

  • Running experiments with our current code, exploring hyper-parameter space

  • Improving the efficiency of the current code, in particular for parallelization on cloud or clusters

  • Improving the RL strategies used

  • Improving the deep learning architectures used.

Here are some slide-like descriptions of the approach, drafted by Maksym Korablyov, who is an MIT grad student who has been working for about a year with me at Mila, has a lot of chemistry experience and leads the underlying science with me:

https://mila.quebec/en/exascale-search-of-molecules/

If you are interested in helping or knowing more, please fill this form: https://docs.google.com/forms/d/e/1FAIpQLSe1VjWc2WOAybytRbYe6MxiVxBwT4IySSqG2FkehxIOo9Oa7Q/viewform

Mila professors who already want to contribute include: Doina Precup, Pierre-Luc Bacon and Jian Tang (plus me of course). We need dedicated people on this project. We are also going to ask for funding to help the project (mostly for the fabrication of the molecules) so we should also be able to provide funding to alternative students who may be diverted from other previously planned projects while being involved in this one.

I hope that I don’t need to tell anyone how important it is that we get involved to the extent we can to fight this war, which hopefully will unite all of humanity against the virus and against the incompetent actions which could leave millions dead unnecessarily. Hopefully that fight and those deaths will also leave something positive in the minds of as many people as possible:  the profound realization that we are all in it together on this planet, for this fight as well as for many others.

– Yoshua

array(1) { ["wp-wpml_current_language"]=> string(2) "en" }

Mila goes virtual

Starting March 16, 2020, Mila shifts its activities to virtual platforms in order to minimize COVID-19 diffusion.

Read more