Early in 2020, I got to spend three weeks in @NotionHQ's vibey, almost legendary Mission office before Covid sent us home for a year and a half. Now, six years later, the same office has become Rowspace's third HQ. The furniture is not as nice but the vibes are still immaculate. We're hiring product and research engineers, FDEs, and GTM here and in NYC to push the frontier on agent harnesses and context engineering for financial services and to partner with customers on deploying these to drive outsized returns. Come work with us!
https://t.co/gCfYvov26M
"The Bitter Lesson has fully arrived in sequence biology and protein structure. Evo 2, AlphaFold 2 and 3, ProGen3, RFdiffusion".
This sentence has some issues IMO. 1/
Excited to officially launch @Rowspace_AI — specialized intelligence for investors to make faster, sharper decisions by compounding their edge with AI. For the first time, they can turn their stores of messy, fragmented data into judgment and workflows that scale 🧵
@kenbwork Reminded me of this post by @iskander
Lots to be optimistic about but sometimes it’s hard not to eye roll at another post about “biology is the next software” etc.
I think this was part of the implicit premise of the first incarnation of Hammer Lab at Mount Sinai, about a dozen of us with math/CS backgrounds ditched tech for biomedicine.
And we got humbled hard: most of what we did flopped & techies don't understand experimental design.
@srikosuri@OmicsOmicsBlog@nimivashi15 Sorry, was referring to Luria-Delbruck. I also misremembered what I read. It was a nice section in David Mackay's ML book reanalyzing the experiment. He says how their original stats had very high variance. Sec. 35.2 (p. 446) https://t.co/djpCDRAoBH
@OmicsOmicsBlog@srikosuri@nimivashi15 I thought this one turned out to be an incorrect analysis that luckily produced the correct result. I think the experimental design was okay though.
2/ First, a new antibody discrete diffusion model. We specifically build on EvoDiff (@MSFTResearch) and D3PM to develop a diffusion-based tool for antibody sequence design. This complements Ginkgo’s scalable data product for antibody developability.
https://t.co/NoARksYH1w
Incredibly excited to share new results from Nabla Bio where we show we can design antibodies de novo for use in therapeutic discovery.
We introduce JAM, an AI system we’ve developed to design de novo antibodies with good affinities, early stage developability, and function. We’re seeing success across a range of soluble and hard-to-drug membrane proteins including a Claudin and a GPCR. We’ve extensively tested these designs in our wet lab and included detailed data and controls, providing the first clear demonstration of how de novo design could expand the scope and efficiency of therapeutic antibody discovery. 🧵
Technical report: https://t.co/W1RtFuTKiH
Blog: https://t.co/43KBQmZ629
A quick personal update: after 3.5 amazing years at Notion and some time off, I've started a company that's at the intersection of everything I learned at Stripe, Notion, and Google—AI, fintech, productivity, and search. Our small team is growing, and we're looking for founding engineers and designers. Please reach out if you or someone you know might be interested in something early stage and in person (in San Francisco)!
Thrilled to share our research on designing better 3' UTRs for mRNA therapeutics and vaccines!
By integrating high-throughput stability assays, machine learning models, and mRNA domain expertise, we engineered 3' UTRs that improve mRNA stability and protein production in vivo.
Excited to share our recent work at @Ginkgo . TLDR: high quality mRNA stability data -> performant ML models -> design 3’UTRs that increase stability -> in vivo validated ML 3’UTRs beat benchmark. https://t.co/z7PUAHaHHg
Nice work. A few important quotes
"While top convolutional neural network models performed slightly better than top LSTM models, transformers performed worse than both LSTMs and residual convolutional models" 1/
Really excited to share work from our team at @Ginkgo, led by @AlyssaKMorrow, on ML guided 3’ UTR design.
https://t.co/XQ997bw5wb
It's a great story of dataset generation, model training, generative design, in-vivo validation, and model release.
More details in thread!
https://t.co/QCtOIaybOj
Nice benchmark of single cell "foundation models" (scGPT, scFoundation) and GEARS (a GNN model) further hyped as "virtual cell models" against linear baselines on perturbation prediction.
Long-story short: they can't beat the linear baselines. 1/
I’m going to be judging at @Lux_Capital, @EvoscaleAI & @envedabio’s Bio x ML hackathon!
Don’t miss out. Apply by Sept 22: https://t.co/lO1Lzq43Yd - open globally.