In late 2021 Collaborations Pharmaceuticals Inc., were invited by The Swiss Federal Institute for NBC-Protection—Spiez Laboratory—as part of the ‘convergence’ conference series to present on potential misuses of artificial intelligence (AI) approaches. A quick thought experiment was proposed in which we used our preexisting machine learning models for bioactivity and rat acute LD<sub>50</sub> (in order to identify toxic molecules) along with a generative algorithm trained on the ChEMBL database to develop molecules with the aim of assessing if we could produce VX. In 6h we had produced 40,000 molecules including VX, related known analogs as well as thousands of other molecules that looked similar and were drug-like using the QED score. Many of these molecules scored better than VX. We were very surprised with the ease with which this was possible. We stopped well short of assessing molecule synthesizability, performing synthesis and testing for obvious reasons. The work was presented at the conference and subsequently we were introduced to Dr. Lentzos and Dr. Invernizzi who are experts on arms control and dual use. Together we collaborated to write up a commentary that was then submitted in October 2021 and ultimately published in March 2022. The article has drawn wide attention from the global media because it was published in the middle of the Russian war in which there is the potential for use of chemical weapons. We will present this thought experiment and the implications of it. We will address some of the many questions we have faced since, including whether we should have published the work in the first place and our motivations behind it. Our commentary calls for the need for more discussion around dual use in cheminformatics at conferences. These early efforts at using generative approaches indicated the potential scale of the issue with hundreds of companies around the world engaged in using similar approaches for drug discovery. We will offer potential solutions to limit the use of these developing technologies to prevent those with harmful intent from accessing data and models. We will also discuss the implications for open science and the need for teaching of ethics to those using AI. In the process we will propose that we have now created a ‘teachable moment’ for dual use and reiterated the need to be vigilant.

Making the world aware of the potential dark side of generative AI for drug discovery

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.