DIMENSION

That's a wrap! Today, we're (@_DimensionCap) releasing the third volume of our Research Magazine. Before anything else, I want to shout out my tremendous research colleagues Bauer (@bauer_lesavage) and Frank (@ChemVagabond), who were both core contributors to this volume. We're also releasing a digital version: dimensioncap.com/research-ma… ~~~ Compared to previous magazines, our third work is more heavily focused on machine learning. That's by design. Dimension invests in all companies, small and large, spanning the continuum from compute to the life sciences. At our core, we love technical teams working on the hardest, most important problems. That means data center infrastructure, new networking protocols, new silicon, inference optimization, physical AI, lean/tech-enabled biotechs, data generation, and more. As such, we challenged ourselves to go deeper into the compute stack than ever before, and how it connects to various branches of science. This magazine is a small sliver of the proverbial 'research exhaust' that our team generates over the course of a few quarters. ~~~ One piece I want to highlight is Without Form and Void, which was my attempt at unifying my background in fluid mechanics, my admiration of clever inductive biases in ML, and a major industry chokepoint seldom talked about in CS circles, which is bioprocess scaling. It's also the first piece I've tried writing with this much history (back to Galileo). ~~~ We hope you like Vol. 3 and we can't wait to show you all what's in store in Vol. 4!

We have a new essay out about how our growing understanding of cell membrane dynamics is leading to better medicines. This work was led by my fantastic colleague Frank (@ChemVagabond), who commands the topic so well given his time at AbbVie and Moderna doing relevant work. ESSAY LINK: research.dimensioncap.com/p/… ~~~ I really enjoyed Frank's usage of HER2, perhaps the most well-known cancer target, to tell this story. I was surprised to learn how the rolling advent of new measurement technologies keeps kicking up more therapeutic innovation. With each new way to study cell surface dynamics, there's a new little widget of the biology that drug developers can cleverly exploit to fix something about previous-generation drugs, or to make new drugs altogether. During our behind-the-scenes conversations, I kept thinking of this Mark Twain quote: "It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t so." With each new tool, it's important to reconsider whether prior knowledge continues to get supported, or if it begins to erode. When I analyze preclinical data or start to understand a therapeutic hypothesis built on years of evidence, it's good to continue to ask this question. For example, the shared prior might be, "this is a bad internalization target", when in reality, maybe it's actually a great internalization target, and the previous set of tools landed at the wrong conclusion because of this limitation. Nonetheless, I enjoyed learning from Frank about these assay techniques and the backend approaches to integrating their signals in a clever way. It's timely because new biology is incredibly important for building better medicines.

It is almost 2027 and I cannot believe we are still talking about AI drugs, this time revived by Shkreli's comments on the live All-In podcast. There is literally no such thing as an AI drug. Never was. So playing the game of tracking when one will be clinically approved, how many are created a year, what the cost-delta is between them versus 'regular drugs', all of these debates begin from a false premise and are thus not useful to talk about IMO. Don't get me wrong. ML methods are extremely powerful and have already clearly and unequivocally contributed to many successful drugs. Like that's not even a debate. Every now-genetically-validated target discovered by sequencing? Thank god we had CNNs and other older ML methods to convert the fluorophore data into useful signal. Every small molecule developed using older ML-based classifiers during virtual screening. That's ML too. Of course, these are the 'boring' ML methods that people don't even count anymore because they're way past their hype phase and have become silent infrastructure underpinning the whole industry. In ten years, today's hype methods will do the same. They'll lose their sheen and people will goalpost move, or they won't even acknowledge them as AI/ML because they'll be so busy saying whatever current thing from 2035 is fake and will never contribute. All sufficiently useful technology moves through these cycles. Once it works, it won't be called AI anymore. It will just be math. Or statistics. Or something else. Any drug program from here on out will use some form of ML, hype or not, so arguing about some false binary whether a molecule/program is 51% AI is just ludicrous to me IMO. This isn't even about Martin at all. Like I get his frustration, and I think it behooves anyone to be skeptical when magical elixers are passing hands. It's just that I think this discussion has run its course and I don't think it's a good use of time.