Operating Partner
SV Health Investors
It is time we talked about artificial intelligence (AI). I have always been a bit of a skeptic about how important AI will be for the biotechnology industry, but, given a few of the companies that have formed and the amount of money they have raised, it seems I should take another look, and perhaps may even change my mind. The generation of scientists who trained in the late eighties and nineties and onwards are very much more computer savvy than I am. I am barely (by their standards) computer-literate. I was asked the other day what ChatGPT was, and I had to go to Google to look it up. Don’t forget I am from the NEANMP (“no email and no mobile phones”) era, and had P&Q (peace and quiet) when there was no intrusive social media to contend with.
But I can observe from my usual perch halfway up the tree. I see that there are different types of AI companies being set up. Those that are mostly focused on AI technology and moving the state of the art forward in molecular dynamics and writing programs to identify better ways to predict protein–protein and protein–small molecule interactions. There are also companies that use these applications to design better antibodies or suggest what small molecules might be synthesized to agonize or antagonize particular enzyme–substrate partners or ligand-receptor interactions. They then make the molecules and test them with a view to obtaining better drugs that affect those interactions more quickly. The companies developing the software and the novel algorithms for selecting high-quality “binders” face competition from public domain open-source programs (e.g., Boltz-2 and others) that can do many of the things that the companies are also doing. Considerable amounts of money have been raised by these companies (e.g., Chai Discovery) and also by AI franchises that intend to become drug makers, such as Isomorphic Labs and Xaira Therapeutics.
Whatever type of AI company—if you are in the drug discovery business—the fun starts at the point where you have done things called “experiments” to test the output of your
AI-designed drugs in both in vitro and in vivo assays to find out if you are right in the real world. Sometimes these assays need to be developed first, which of course adds additional time to the process. AI can play a role there, too, because it can help to define what the appropriate assays may be and how to run them in multiplex forms, i.e., so that many suggested molecules can be run in parallel before doing more extensive experiments with a subset of the designs. But as anybody who has worked in the wet lab will tell you, cell-based assays are variable, and animal experiments even more so. Collecting the data from the actual experiments and refining things using AI tools is, of course, possible and is being done.
What it comes down to eventually is the quality of the data that is used for these AI approaches. The old adage of “crap in-crap out” still prevails. In the protein structure prediction space, there is enough data to be able to predict the three-dimensional structure of proteins at a resolution of about an angstrom, so predicted structures these days are pretty informative. AI is also being adapted by many groups to look (for example) at VHH designs that interact with proteins in an epitope-specific way, complementing data from phage display or yeast display libraries. Unfortunately, data sets coming from cellular data are much less robust than protein structures, and it will take a while before AI can deliver equivalent results in this arena.
The approach is not cheap either. You tend to think that it costs you nothing to do a search on your computer. This is true, but these generative AI programs are at a very high-volume scale and use very considerable amounts of computational power and electricity. The costs of AI infrastructure add up very quickly.
The other area where AI tools are being extensively used is to search for information in scientific publications. It is almost impossible these days to keep up with what is published. Most weekly issues of Nature are now around 400 pages, so there is an increasing reliance on using AI to abstract the data and information you are looking for. The rub is that AI looks at information in a relatively unbiased way. Scientists do not. The quality of the information, what journal it got published in, who the authors are, and where they work are all part of the diligence that scientists use to judge the results of the experiments that were done and the conclusions drawn. AI does not have that capability. You can test this for yourself by doing a Google Gemini search or equivalent on something you know about and see what you get. It is generally rather superficial and sometimes just wrong.
At the time when DNA sequencing was all the rage and Celera was in full flow, some of the comments at the time were that there was more heat generated by the computers used to put all the fragments that were sequenced together than progress towards a full human genome sequence. The “hype heat” in the AI space presently is way more than the heat generated by all the processors that are being used to generate the data that the biotech and pharma companies are now going to be using.
Have I changed my mind? Maybe a little bit, but I do know that at the end of the day, the experiments that are done to test the designed molecules will be most important (and always will be) when you are making drugs for people.
