⚡️ My PhD thesis is on arXiv! ⚡️ To quote my examiners it is "the textbook of neural differential equations" - across ordinary/controlled/stochastic diffeqs. w/ unpublished material: - generalised adjoint methods - symbolic regression - + more! arxiv.org/abs/2202.02435 v🧵 1/n

Having started a biotech co in the past two years, the recent Mithril Medicine news felt pretty unsurprising. When people ask me what surprised me most about being a biotech founder, I always say that I dramatically overestimated the number of people I would need to hire. AI has been transformational here, but so has the excessive build-out of CRO capabilities in the COVID ZIRP era making CRO prices dramatically cheaper than they would be otherwise. Operations, chemistry, tox, and most preclinical work can all be outsourced for 20-60% of the price of building the same capabilities in house. The core platform still needs hands-on internal work, but if you're building a very traditional asset-flip biotech startup, then there is no core platform to begin with. There is some sacrifice in terms of speed over the long-term, but the combo of AI+automation+CROs suddenly allows biotech startups to create value on previously-challenging VC fund timelines. My prediction is that a biotech startup model consisting of a tiny team of 1-5 people is going to become fairly standard in the near-term future.

mfw >biology is just a data problem >collect enough omics, simulate the whole cell >read about humanin >peptide encoded inside the mitochondrial 16S rRNA gene >the same sequence is also a structural part of the ribosome >24 aa if the cytosol translates it, 21 aa if the mitochondria do >because mitochondria run a different genetic code >and start translation with N-formylmethionine instead of plain methionine >nobody's fully sure which one actually makes it >also 13 nuclear copies that might make their own versions >gets secreted into the blood >hits cell-surface receptors >one of them is literally a mitochondria detector >also gets taken back up >binds Bax inside the cell to block apoptosis >discovered by screening surviving brain tissue from an alzheimer's patient >open my cell model schema >compartment: yes >molecule_type: rRNA, peptide, hormone >length: depends who's asking

Someone needs to raise $200M to launch cell-free (non-E. coli) GMP manufacture clearly the best way to make n=1 cancer treatments / any experimental low-batch-size stuff

Update: I've now *also* had multiple JAX-for-{fusion,molecular dynamics,...} companies ask me for referrals in the exact same way. I need to start hosting a job board haha

What the 'neopharma' discussion misses is that this is a class of companies, as "software" is a class. The VC pitch is the layer of detail below. Modality/in vivo data/etc. If you want to distinguish the serious/unserious, ask them what their GLP-tox strategy is.

The degree of target clustering in pharma right now is absolute madness. How is this not objectively idiotic? The pendulum will swing back to unique/novel MoAs eventually…it just has to. Right now you can get VALIDATED novel/unique MoAs for less than some of these me-80 drugs in some cases. The system seems broken.

As usual, doing my bit to talk about bio topics in front of my AI audience: COINS (home.treasury.gov/news/press…) a possible rule to restrict US investments/JVs with Chinese biotech, released yesterday. My guess is that this will backfire the same way that the GPU blockade did: it will just result in a stronger competitor. And China's homegrown industry is already arguably the strongest (?) in the world. I don't have a stat for that, but it's noticeable from the ground: e.g. Chinese CROs actually respond to their emails same-day instead of next-week; WuXi clearly the market-leading biologics CDMO; etc.

API for mRNA / nanoparticles here 👋🏽 w T cell activation data in vivo

AI+bio: the Escalante folks released a blog post yesterday, and as usual it's a banger, talking more about their multiplexed protein binding assay: blog.escalante.bio/endless-p… (I imagine this got lost amidst the securitypocalypse that landed yesterday.)

Great to see more API-for-biology companies out there (alongside Adaptyv, Biointron, Ginkgo, etc.) Warp looks chemistry-inflected (unlike a lot of the protein-focused ones we see). We still need more of this for the 'harder' properties: immunogenicity/T-cell activation, etc.

Excited to announce the follow-up to our NeurIPS 2026 Oral: Improved Training of Strong Stochastic Flow Maps! We improve the training of SSFMs by unlocking derivative-based losses! Joint work with @antoniofrancaib @rmincam @ChristianKoke @Niklas_TR and @AlexanderTong7

Lots of VCs kindof sound like this: "Your basketball team doesn't have a moat, any healthy human can throw the ball into the basket thingy. " People who have not run companies before don't understand how much the team matters more than some imagined structural moat.