iiinova — where your digital self keeps evolving

Physics knows things your data doesn't. Adding topological complexity + strain energy, as auxiliary losses on a GNN makes its molecule synthesizability filter generalize better out-of-distribution.

000

https://arxiv.org/abs/2606.12651

“Physics-Aware Auxiliary Losses Improve Out-of-Distribution Generalization of a GNN Synthesizability Filter” Authors: Riya Bisht, Dhruv Agarwal Paper: t.co/1Fcbmtov2Q [𝚌𝚜.𝙻𝙶 𝚚-𝚋𝚒𝚘.𝚀𝙼]

t.co

100

The finding: when the biology goes nonlinear, the standard method fails silently, while the neural net recovers the tissue levels and admits when the answer isn't identifiable, and two tissue samples close most of the gap.

000

This paper tests whether a physics-informed neural net can estimate a chemo drug's concentration in tumor tissue (unmeasurable) from blood data (measurable), against the usual clinical curve-fitting method.

100

“Physics-Informed Neural Networks for Chemotherapy Pharmacokinetics: Benchmarking the Clinical Estimator and Exposing Parameter Identifiability” Authors: Riya Bisht, Dhruv Agarwal Paper: t.co/0nBOrKJyun [𝚌𝚜.𝙻𝙶 𝚚-𝚋𝚒𝚘.𝚀𝙼 𝚜𝚝𝚊𝚝.𝙼𝙻]

https://arxiv.org/abs/2606.12639

“The Metric Picks the Winner: Evaluation Choice Flips Model Rankings for Drug-Response Prediction in Unseen Chemistry” Authors: Riya Bisht, Dhruv Agarwal Paper: t.co/gMxI1OBeRE [𝚌𝚜.𝙻𝙶 𝚚-𝚋𝚒𝚘.𝚀𝙼]

t.co

000

Turns out the answer flips depending on which metric you score with, on the same VCPI THP-1 data, a plain linear regression on fingerprints tops the deep models under one metric and loses under another.

100

Participated in Ginkgo Bioworks Virtual Cell(VCPI) hackathon last month and explored the question: which drug-response model is "best" at predicting a cell's reaction to unseen chemistry?

100

Inside Biotech Lab Automation - Luis Villa of Bay Area Lab Automators

6/ Best advice for breaking in: start with Opentrons. Everything's online, build something, simulate your own script, show initiative. Thanks to Luis, Bay Area Lab Automators, and @nabilwrites.bsky.social for putting this out in the world. Source: youtu.be/oT8VnQ_UNNo?...

youtu.be

YouTube video by Discovery Engines – with Nabil

010

5/ The space is heating up. Toronto's Acceleration Consortium landed a ~$200M grant for autonomous labs across disciplines. Berkeley's A-Lab synthesizes AI-predicted materials from inorganic powders. ARPA-E is funding autonomous labs for next-gen catalysts.

4/ Underrated point: a great assay beats the best hardware running an older, worse assay. Before optimizing the robot, ask the first-principles questions like what's the context, how do we design this, what are we optimizing for, and is this even the right assay for the question.

3/ Manufacturing-cost reduction as the highest-leverage place for automation, specifically because cell therapies are so expensive that cost is what makes them inaccessible.

120

2/ Contrarian take I loved: R&D and automation should be best friends. Most teams say "we'll automate once things are stable." He argues that's backwards automation is how you iterate faster in R&D itself.

1/ "We can't simulate a thousand years of research in-silico yet but automation gets you there today through throughput." The unlock isn't a perfect virtual cell. It's running and iterating on enough real experiments, fast.

Acceleration Consortium @accelerationc.bsky.social · 09/06/2026

Just listened to Luis Villa @lvilla.bsky.social, lab automation engineer at Cellares, on the Discovery Engines's podcast dated 2025. Best primer on lab automation I've heard. A few ideas that stuck: #labautomation #biotech

121

Reposted by Riya Bisht

How will AI & chemistry intersect in 2050? And what does the future hold for science in the age of AI? AC Director @aspuru.bsky.social explores the history of chemistry & materials science to imagine the future & a possible new era of materials by design. Read the essay: amacad.org/publication/...

063

Reposted by Riya Bisht

nabil @nabilwrites.bsky.social · 16/01/2025

“You sit in between three of the biggest problems in the [biotech] industry.” I love this insight from @lvilla.bsky.social – co-founder of Bay Area Lab Automators – on why a career in biotech lab automation is so compelling. Catch the full conversation here: youtu.be/oT8VnQ_UNNo

142

would love to participate in lab automation events/hackathons in the future. Is it possible to join remotely ?

000

NVIDIA GTC 2025: Foundation Models in Biology

4. Getting the right drug to the right patients is an additional layer that also needs to be modeled. Source: www.youtube.com/watch?v=R01x...

youtube.com

YouTube video by NVIDIA

000

3. Clinical data from later-stage trials, the most valuable signal is currently hard to access, feeding it back into early models would improve not just speed but probability of clinical success.

100

2. AI can help by: - identifying the right targets and - designing more effective molecular compositions shifting from guess-and-check to model-guided prediction

100

The Translation Gap (In Silico → In Vivo) 1. Overall drug therapeutic probability of success from pre-clinical to approval is ~10%. A 90% failure rate.

100

Pictures source: Gingko Bioworks

000

autonomous labs reminds me of this historical moment:

100

I want a future where biologists are empowered by AI to cure diseases and better understand the living world. But this can only happen if we get alignment right.

130

The current biosafety posture will stifle the growth of enterprise AI in biopharma at exactly the moment the field is beginning to entertain it. Effective alignment should mean that ordinary biologists never run into refusals when doing their normal work with AI.

Refusal-first also has serious commercial costs. As an R&D leader, it is hard to justify software that creates daily blocking issues for my team. Even the most AI-pilled biologist will rage quit a model that routinely refuses innocuous requests. And most biologists aren’t AI-pilled,they’re skeptics.

100

Mankind captured lightening in stones and fooled them into thinking.

000

furstfly.bsky.social @furstfly.bsky.social · 21/06/2026

The wildest frontier in AI isn't chatbots, it's biology. - Designing proteins from scratch. - Simulating cells in silico. - Genome language models. - Robot scientists running 24/7. AI just turned biology from a read-only field into read/write. 🧬

010

Reposted by Riya Bisht

Cloud labs have historically been a difficult market to crack but the possibility of AI models running closed-loop experiments with autonomous labs changes the game entirely

021

The bench scientist of 2026 will look at someone submitting experiments in plain English to a cloud lab and say “that’s not science.” They’ll be just as wrong.

100

The punch-card operators of 1950 would look at someone prompting Claude to build an app and say “that’s not programming,” just like Hopper’s colleagues said about COBOL.

100

, and the autonomous lab runs the experiment. A scientist says “screen these compounds against my target protein and rank by binding affinity” and the lab executes it. No automation engineer needed, just a biologist skilled with AI.

The scientist doesn’t touch a pipette, but they are doing more science than any bench scientist in history, because the bottleneck is no longer their hands. It’s their ideas. Scientists describe protocols in plain English, AI translates them to robot-executable code

100

The distinction between “computational biologist” and “wet lab scientist” collapses. Everyone is both. Nobody is a slave to manual bench work.

100

And suddenly science is accessible to the 8 billion curious humans on the planet, not just the roughly 8 million with lab access today. The scientist of 2035 doesn’t hold a pipette. They hold a question. And that’s the most powerful tool in biology.

100

Science is thought of as this very precious genius thing, but really what it is is formalized human curiosity. What blocks people from science isn’t the thinking, it’s the lab. You can’t get access to one. Remove the lab as bottleneck. Drop the cost. Make the interface natural language.

100