Variational AI avoids the pipeline trap (with some help from Merck)

A few weeks ago, Variational AI announced a collaboration with Merck on a couple of small molecule programs. But what I found interesting about the announcement is that while the deal seems to be structured like a traditional partnership between a large pharma and a biotech startup, Variational is actually a purely digital. So this week, I want to explain why this partnership is interesting, and speculate about what this might look like from a technical perspective.

Variational AI provides a ligand-based generative AI service for small molecules: The pharma partner hands them a list of the hits and non-hits from their initial high-throughput screens. They hand back a list of compounds that should have a high(er) concentration of hits, and maybe with particular ADMET profiles, patentability, synthesizability, etc.

The pharma partner then synthesizes and tests these compounds, sends them through optimization, animal models, clinical trials, etc. Variational doesn’t (as far as I can tell) have anything to do with most of those steps, besides maybe a few more rounds of handing lists of molecules back and forth.

So a reasonable observer might expect that the pharma partner would pay Variational for their services, say thank you and move on. But that doesn’t seem to be what’s happening here.

Instead, the press release states that the deal includes milestone payments of up to USD$349 million. Yes, it mentions an upfront payment, but I would expect that to be a very small part of the $349 million, and significantly smaller than Merck would’ve payed if they didn’t also have the later milestone payments.

Milestone payments like this are much more common with co-development partnerships where a biotech startup does the bulk of the screening and optimization work (for a presumably larger upfront payment to cover that cost) then the pharma takes over for the clinical and/or commercial stages.

A deal like this is risky for Variational because they only get those milestone payments if Merck both runs the programs successfully, and decides to continue investing in the programs into the clinical and commercial stages. On the other hand, Variational has only raised $9 million so far, so even if they only get a fraction of the $349 million, their investors should still be very happy.

So maybe this model is an antidote to the usual forces that push biotech startups into starting internal drug development pipelines. They get to take on some of the risk of drug development, in return for some potential reward, while keeping their costs low and focusing on what they’re good at.

And speaking of what they’re good at, the real question is whether what they’re doing is anything special. Small-molecule generative chemistry is pretty widespread these days. Most small molecule startups have some kind of generative models. There are open source models you can download, and libraries that would enable a competent data scientist or computational chemist to put together their own. But, of course, some models are better than others, particularly if they’re trained on the right data.

Based on Variational’s website and their whitepaper, it seems like their main selling point is their approach to scaffold hopping - a way of generating molecules that you wouldn’t find through traditional cheminformatics.

A scaffold is basically a minimal molecule that you can use to create other molecules by adding atoms in different places and configurations. Chemists traditionally design new molecules by finding scaffolds that underly the hits from early screening, then generating new molecules from those scaffolds.

The idea of scaffold hopping is to find other molecules that have a similar three-dimensional shape to the hit molecules, but with different underlying scaffolds. There are published algorithms that do this and most small molecule startups are doing something along these lines. But again, some algorithms are better than others.

Merck has chemists who know much more than I do about scaffold hopping, and would have tried many of these algorithms themselves. So the fact that they decided to partner with Variational tells us that either their algorithm is significantly better than those other ones, or Variational is able to operationalize it in a way that’s hard for a large pharma, or someone in Variational’s BD team is really really good at their job.

I have no inside knowledge, so I won’t speculate on which one it is.

But regardless, I’ll be watching to see what comes out of this.