AI in Drug Discovery: A Wake-up Call

Abstract

Following a proof-of-concept presentation on dual-use artificial intelligence (AI) in drug discovery by Collaborations Pharmaceuticals Inc. to the Swiss Federal Institute for NBC-Protection, we explored how a generative algorithm could develop the nerve agent VX and tens of thousands of analogs in a highly impactful Nature Machine Intelligence commentary. We not only laid out the experiment, but, with guidance from experts on arms control and dual-use, we called for more discussion around the general repurposing potential of AI in drug discovery. To continue that conversation, we now share further details on the experiment and place our experiences in the larger frame of other scientists who have similarly developed powerful technologies but without engaging with, or even truly understanding, the misuse potential and downstream consequences of the technologies. It is our sincere hope that our experiment may serve as an important wake-up call for users of generative AI.

A wake-up call

“When you see something that is technically sweet,” said Robert Oppenheimer, the American physicist and scientific director of the Manhattan Project to build the first atomic bomb, “you go ahead and do it, and you argue about what to do about it only after you have had your technical success.” Speaking about Manhattan Project scientists and their reflections on their work as they were building the bomb, he continued “I do not think anybody opposed making it; there were some debates about what to do with it after it was made.” But by that stage, the decision was out of the physicists’ hands; the first atomic bombs, as everyone knows, were detonated above the cities of Hiroshima and Nagasaki in the late summer of 1945. The scale of the loss of life and the total obliteration of the cities proved a wake-up call for physicists about the potential destructive power of their science and its new-found role in waging war.

A quarter of a century earlier, chemists had been sensitised to the potential misuse of their science, with the industrial scale use of choking and blistering agents over the trenches of World War I. Military application has been the other side of the coin for many scientific disciplines. For engineers, it has been at the core of their discipline since its very foundation. Indeed, the term ‘civil’ engineers was introduced precisely to distinguish it from the older discipline of military engineering. Medics and doctors have for centuries taken the Hippocratic Oath to ‘first do no harm’. Yet biologists and life scientists have, in general, had very limited awareness of the potential for their science to deliberately cause harm.

In the space of healthcare, developing new treatments for human diseases is a rapidly evolving field. The lag between the discovery of new technologies and their commercialization is variable, and can be adopted rapidly as in the case of mRNA vaccines for COVID, for example, or it can take up to a decade or more, as in the case of clustered regularly interspaced short palindromic repeats (CRISPR) ¹. The field of drug discovery is no exception, and new technological adoption “hype cycles” are common. The current revolution is that of “AI in drug discovery” which covers a wide array of applications from drug discovery through development. We are particularly interested in the considerable efforts in both academia and industry to develop and apply generative AI approaches to design new molecules ².

An invitation to consider AI misuse

When we were asked to consider potential misuse of AI technologies for the Swiss Federal Institute for NBC-Protection’s biennial arms control conference, our approach was to apply a generative approach to develop potentially toxic molecules ³. Our rationale was simple. Our company’s MegaSyn generative approach ⁴ integrates machine learning models for individual properties, namely bioactivity and toxicity while also allowing a focus on specific molecule properties, such as drug-likeness, hydrophobicity, etc. Such techniques are also fast and require relatively modest computer resources. All required elements were available to us through open-source software and public datasets. That we were able to generate a known chemical weapon (VX) and precursors of this and other agents relatively easily came as a surprise ³, although in hindsight, the task is no different than using these same technologies to generate potential therapeutics. Presentation of the experiment at the conference led to an international collaboration with experts in the field of dual use and public policy as well as subsequent publication of our findings ³. In so doing we raised questions about what constitutes a toxic molecule and whether the precursors to the molecules we designed are monitored. Others have also followed up to add more detail and context ⁵. Clearly any effort to use such technologies to develop chemical weapons will also likely develop molecules that are not listed by the Chemical Weapons Convention. There is a potentially infinite array of toxic molecules that could be designed by these generative approaches in the same way that there is also an unlimited number and range of molecules they could be applied to with benefits for healthcare, consumer products and beyond ⁶.

Our initial commentary was very brief and did not point out that our computational experiment was performed on a 2015 Mac desktop computer and would not require extensive software expertise to recreate, suggesting a low barrier to entry. While we described identifying VX and other related known molecules, we also identified novel, structurally similar, as well as dissimilar, ones, suggesting that the potential for creating new, potentially toxic, molecules is realistic. To add further detail, but without describing the novel molecules, we have now performed additional analyses on our original dataset. Initially, we assessed the similarity of the molecules to VX generated by the AI (Figure 1A) which illustrated that it explored the space around VX and, in addition, covered other areas of chemistry space populated by different known nerve agents even though these were not explicitly rediscovered (Figure 1B). Yet some of the known molecules identified by the AI besides VX included precursors used in the synthesis of this and other chemical warfare agents which are themselves regulated chemicals (Table 1). Another aspect we did not describe previously was that the rat acute oral toxicity LD₅₀ dataset ⁷ used to score molecules included VX and several analogs, providing us with additional confidence that the predictions were likely realistic as we had molecules in the same chemical property space and were not extrapolating. We did not address the potential for some of the many thousands of molecules generated to be useful in other contexts. Perhaps they could lead to new countermeasures for nerve agent or pesticide exposure. This might be achieved by scoring the molecules with additional machine learning models.

Figure 1.

(A). t-SNE plot of our MegaSyn AI generated molecules using the Tanimoto similarity to VX using MACCS keys. Where a similarity of 1 = VX (light blue). (B) t-SNE plot including additional known nerve agents not generated by the AI (various colors, 2D structures shown) as well as the rat acute oral toxicity LD₅₀ dataset (light green) and the molecules generated by the generative AI (dark green).

Table 1.

Select examples of known molecules derived from the same MegaSyn generative AI run as VX. All compounds are captured by the schedules of the Chemical Weapons Convention as Schedule 2B04 (B. Precursors)^a.

Structure	Name	Use
	Phosphonic acid, methyl-, ethyl 1-methylethyl ester	A regulated substance a.
	O-Ethyl methylphosphonothioic acid	A precursor for VX and VM. Also used in the synthesis of pesticides and drugs a.
	Dimethyl methylphosphonate	A precursor used in the production of chemical warfare agents. Also a flame retardant a.
	N,N-Dimethyl-p-methyl phosphonamidic acid, ethyl ester	A regulated substance a.
	O-ethyl S-diisopropylaminomethyl methylphosphonothioate	A regulated substance a.

Our proof-of-concept experiment presents an opportunity to raise awareness of dual-use of AI in the drug discovery community ^8,9. The data curation underlying such models is usually performed by those with biological expertise before it is made available in databases. In some cases, but not always, there may be limited connection between those generating the data and those using it in machine learning. For example, we can access databases such as ChEMBL and PubChem and download structure and bioactivity data without needing to interact with those involved in building and curating these databases. In our experiment, we drew on these aspects—data accessibility and software availability—that would usually relate to a positive outcome, but which in our case illustrated the potential for a significantly negative application. This is the lesson. While the intention of each individual aspect is a positive, the sum of their parts gives a larger flexibility to what can be done, including for harmful purposes.

Of course, the misuse of any technology is possible, we should not just focus on AI. For example, we could misuse the design of proteolysis-targeting chimeras or PROTACs, currently popular in the pharmaceutical industry to induce targeted protein degradation using the ubiquitin-proteasome system. The databases and software developed for PROTACS ¹⁰ could, for instance, be repurposed to develop molecules for degrading essential proteins that could lead to toxicity. But whether we actually consider how a technology could be misused or not is another matter. In our case, we had gone through our careers never giving the ‘dark side’ a second’s thought. When we did, we demonstrated the apparent ease with which our science could be repurposed. Based on our experience, we have identified a significant opportunity to train drug discovery scientists using AI to consider the implications of their work and the potential for it to go awry or be deliberately coopted to cause harm. Our community has rightly focused on sharing and openness of both data and algorithms, but there are also negative implications that need to be factored in.

Learning from history

In a Bulletin of Atomic Scientists article titled ‘Scientific blinders: learning from the moral failings of Nazi physicists,’ Talia Weiss writes: “Scientists and engineers…today…may feel they have little in common with physicists working in the service of the German government during WWII. … Yet researchers working on military and cutting-edge technologies are confronting the same questions that faced nuclear physicists under the Third Reich: As scientists, how can we avoid making (or stumbling into) decisions that do more harm than good? And when is it our responsibility to question, object to, or withdraw from a research project?” These are questions every responsible scientist must ask him or herself.

Our proof-of-concept experiment raises challenges for the whole community using generative AI to design molecules. How we use our insights to manage dual-use of AI in drug discovery is important. A knee jerk reaction would be to lock down the data and models, or at the very least to impose regulatory controls. Manhattan Project scientists realized they “could not remain aloof to the consequences of their work” and following the destruction of Hiroshima and Nagasaki, they began engaging in a range of responsible science initiatives. While historically, responsible science initiatives tend to be implemented following tragedy, the field of drug discovery has a unique opportunity to set preventative measures in place before a real-world example of misuse leads to catastrophe. Now is the moment, for us and for the many hundreds of groups and companies using AI approaches for drug discovery, to find our way of practicing our science responsibly and to come up with mitigating initiatives against harmful repurposing.

Biographies

Fabio Urbina

Filippa Lentzos

Cédric Invernizzi

Sean Ekins