The '4th Wave' of AI Drug Discovery is Here, According to This Report
I’ve come across a really well-put report on the evolving state of artificial intelligence (AI) in drug discovery in 2024 and beyond, authored by Chris Bradbury from Stanford University Graduate School of Business.
Essentially, the author breaks down AI progress in drug discovery into 4 historical stages, or "waves", with the following characteristics on the slide:
The report delivers several notable messages about a new wave of companies that build new tech stacks in AI and biology, allowing for holistic drug discovery strategies.
It is somewhat similar to what I was talking about during my 2023 keynote speech in Brussels, alongside Alexander De Croo, Prime Minister of Belgium:
the key transformative promise of modern AI systems in drug discovery is not about better high throughput screening (although AI can improve it)
It is not about better virtual screening techniques like docking or ligand-based virtual screening (although AI can improve it)
It is not even about better molecular dynamics simulations (although I saw studies where deep learning was reported to improve it)
Rather, it is all of the above at once of course, but most importantly this:
The real value of applying AI in drug discovery, IMHO, is changing the drug discovery strategy and workflow in the first place, transitioning from target-based approaches towards holistic biology-first approaches where you model the entirety of omics at scale, creating a sort of “world model” of cells and even multicellular systems.
Importantly, you do it not with just one specific cell, but with numerous patient samples, both healthy and unhealthy, to create a global and dynamic map of what is happening in biology when a disease unfolds. And you mind data context, and metadata a lot.
Also, you have to mind really well the sample preparation and conditions, as I wrote previously on this topic that Even 10 Minutes Matter in Cancer Research. In brief, the cold ischemia time (CIT), or the time it takes to preserve tumor tissues after surgical removal is really affecting many molecular states, like gene expression levels etc. So if you "wait for too long", the sample would be giving misleading information for AI model to digest.
Anyway.
Once you build sufficiently accurate global models representing biology at sufficiently deep level, the rest of target-specific problems above will be solved as a consequence, or it will be easier to solve them as sub-problems of the bigger problem that has been already solved.
For myself, I call this new AI-driven approach “rational phenotypic-based drug discovery (RPDD)” if it makes sense to you. Some call it "biology-first." Whatever.
IMO, the main value of AI comes from solving two major bottlenecks:
Finding the “right human body” to intervene in the first place, based on unique body attributes and via novel biomarkers (genetics or biochemistry based etc). I think it is called “patient stratification.” The important aspect of clinical research, that can be massively enabled by holistic AI-based models of biщlogy is so called reverse translation in clinical trials. One popular business case is drug repurposing. And we do see quite successful repurposing examples in the past.
Predicting toxicity and systems effects of a future drug in the human body (systems interactions, polypharmacology, off-target effects, kinetics, bioavailability, BBB crossing, etc).
What happened with many existing AI companies that started hyping AI-driven drug discovery years ago, is they focused on narrow problems (e.g. virtual screening) and so they did not really change the game.
This newsletter reaches over 7.4K industry leaders from organizations like NVIDIA, Microsoft, Novo Nordisk, Pfizer, Novartis, FDA, the US Department of State, Illumina Ventures, and hundreds more.
Interested in sponsoring?
Contact us at info@biopharmatrend.com
Now, according to the new report "The New AI Regimen Emerging in Biology and Chemistry" I mentioned above, there is a wave of next-generation companies like NOETIK, Nabla Bio, BioMap Xaira Therapeutics Iambic Therapeutics and others, that are building a completely different tech-stack, both from the AI standpoint (e.g. unsupervised learning, foundation models, novel multimodal systems etc) and from biology data standpoint (building lab automation for biology at scale, own biobanks with human specimens, etc).
I would argue somewhat with the company classification provided in the report, though.
For instance, I would certainly put Berg Health (now BPGbio, Inc. ), and Insilico Medicine in Wave 4 companies, similarly to Recursion and some others.
Berg Health started with a biobank long ago and they were modeling omics at scale for years and have pretty cool clinical assets.
As well, Insilico Medicine has created one of the most generalizable AI models out there, and also built a fully-automated wet lab facility in China. I am not even mentioning the rapid pace of nominating candidates by them, although they all are small molecules.
Also, Enamine is not an AI company at all, and while they are suppliers of huge chemical data for many AI startups to model chemical spaces, they look like an outlier here. There are other datapoints I would claffify differently.
But this doesn’t really diminish the report value and doesn’t change the key idea of the report which is enlightening: the first AI wave was mostly hype, but we are on the cusp of a more nuanced and more transformative industry upgrade, and you better watch those new players and what they do.
Anyway, kudos to Dylan Reid for pointing out this report to me!
I think it is quite an interesting read and anyone in this space should go through slides—they are super straightforward and to the point (well, Stanford Business School…) and I am a bit envious it is not me who created this super valuable resource :-)
Some additional reading before you go: