Chaitanya K. Joshi

I recently joined The Pauling Principle podcast to discuss our research on RNA design, hot takes on architecture research in BioML, origins of life, and being in a wet lab!

Feels pretty surreal that I'm interesting enough to be on a podcast (?)

www.youtube.com/watch?v=ftrQ...

Four years of research on Geometric Deep Learning & Molecular Design, summarized in 3 questions:

- How expressive are our models? (Theory)

- Can we build foundation models for molecules? (Architecture)

- Can AI design RNA in the real world? (Application)

A Cambridge PhD thesis in three research questions Geometric Deep Learning for Molecular Modelling and Design: A personal scientific journey

New blog 💙: I reflect on why I worked on what I worked on...

I think a PhD is a very special time. You get to challenge yourself, push your boundaries, and grow. My thoughts go against the current AI/academia narrative online, so I hope you find it interesting.

chaitjo.substack.com/p/phd-thesis...

It relies on new algorithms for template search, invented by Kagglers such as G. John Rao. If no templates are found, the model can still work and works decently well without templates, too.

The main takeaway from the Kaggle competition was how powerful RNA template modeling can be — it was kind of shocking to us 🤯

RNAPro integrates 3D templates into Protenix/the AF3 architecture, and this moves the needle on automated RNA structure prediction!

Happy new year! A step change in RNA structure prediction, powered by top Kaggle-ers in a collaboration lead by Stanford University and NVIDIA

Happy to have played a small part in the new RNAPro model, significantly outperforming AlphaFold 3 as well as VFold (CASP winners)

The video was largely created using AI text-to-video, and I must say the results are pretty impressive! There's huge potential for this technology to make science more approachable, just like Google's NotebookLM.

YouTube video by Chaitanya K. Joshi AI vs. Humans: The "AlphaGo Moment" for RNA Design

Merry Christmas!🎄🌟

Sharing something interesting: Friso van de Stadt created a pretty neat explainer video for our gRNAde paper and "AlphaGo moment for RNA design" blogpost!

www.youtube.com/watch?v=wDeZ...

An AI Researcher in the Cathedral of Molecular Biology How we designed catalytic RNA functions, and what I learned holding a pipette at the MRC Laboratory of Molecular Biology.

I wrote some personal reflections about being physically embedded in a world-leading molecular biology lab @mrclmb.bsky.social, learning to communicate with experimentalists, back-breaking wet lab work, scientific rigour, and skin in the game!

💌: chaitjo.substack.com/p/an-ai-rese...

GitHub - chaitjo/geometric-rna-design: gRNAde is a Generative AI framework for inverse design of 3D RNA structure and function gRNAde is a Generative AI framework for inverse design of 3D RNA structure and function - chaitjo/geometric-rna-design

- How to set up the design prompt just right?
- How many sequences to sample and how to make them diverse?
- How exactly is filtering done step-by-step?
- How to analyze wet-lab results and create figures?

You can find out all of that in our GitHub codebase!

💻: github.com/chaitjo/geom...

With this release, we are fully openly releasing all the tacit knowledge for getting biomolecule design models to actually work in wet labs!

Seemingly small-but-critical details are often missing in flagship papers, and it took us a while to finetune our pipeline:

Many papers have built their RNA design pipelines using the gRNAde code & data from our ICLR 2025 spotlight paper - that's been exciting to see! Especially the attempt at biologically meaningful benchmarks 🤗

But how to take ML conference papers to wet-lab validated results?

Excited to release the fully open-source code for gRNAde - our wet-lab validated, generative AI framework for 3D RNA inverse design 🚀⭐️

I pride myself on open-science & this is probably the most intense release I've done!

LeMat-GenBench: A Unified Evaluation Framework for Crystal Generative Models Generative machine learning (ML) models hold great promise for accelerating materials discovery through the inverse design of inorganic crystals, enabling an unprecedented exploration of chemical spac...

Happy to have contributed to and now finally share LeMat-GenBench, a new open benchmark + leaderboard for generative crystalline materials models! ⚛️✨

It provides standardised metrics for validity, stability, & much more. Already includes results for 12 models!

🔗 Paper: arxiv.org/abs/2512.04562
1/4

Thank you to everyone who made the inaugural Virtual Cell Challenge a success.

Over 5,000 participants from 114 countries competed to build AI models that predict cellular responses to genetic perturbations. Today we're announcing the winners and reflecting on what we learned.

I think the term ‘Virtual cell’ will have the same trajectory as ‘AGI’ or ‘Foundation models’: Initially opposed by rigorous scientists, while the Bay Area and Demis Hassabis are the only ones comfortable using it

→ becoming a mainstream term in academia soon, in few years (Overton window)

Thankful for the tireless work and motivation from yourself and the Eterna team!

An AI researcher interested in biochemistry modeling successfully improved his RNA language model through participation in the Eterna pseudoknot design competition. Congratulations, Chaitanya! 🧬🧪 #RNAsky

The polymerase ribozyme results are pretty cool too. 😎

Enumerating possible pseudoknots that a sequence can form with nearest-neighbor models is an NP-hard problem. Even evaluating these structures is challenging, let alone designing them. So it’s great to see data-based models starting to crack the RNA structural design problem! 🧬🧪

An AlphaGo moment for RNA design How our AI system, gRNAde, matched human experts at the complex game of RNA folding—and why it matters.

5/ Want the story behind the science? 🏰

I wrote a blog post about our "AlphaGo Moment" and what it was like being an AI researcher embedded at the legendary @mrclmb.bsky.social (holding a pipette for the first time!)

Read it on Substack: chaitjo.substack.com/p/alphago-mo...

Generative inverse design of RNA structure and function with gRNAde The design of RNA molecules with bespoke three-dimensional structures and functions is a central goal in synthetic biology and biotechnology. However, progress has been limited by the challenges of de...

4/ This was a massive team effort bridging AI and biology, from one end of Cambridge to another 🤗🚲

Thanks to Edo Gianni* @edogia.bsky.social, Sam Kwok*, @simonmathis.bsky.social, Pietro Liò, and @philholliger.bsky.social for this journey!

📄 Preprint: tinyurl.com/gRNAde-paper

3/ The Mechanistic Insight 👽

What did gRNAde learn? Humans stick close to nature, but gRNAde changes ~70% vs. WT sequence.

And it "sees" invisible 3D constraints that rational methods miss, allowing us to make "generative jumps" to new functional islands in sequence space 🏝️

2/ The Function Challenge ⚙️

But can we move from static shape to dynamic RNA machines?

We show how with a complex RNA Polymerase Ribozyme.

Rational design failed (3% success). gRNAde succeeded (31.5% active), discovering improved variants 15-20 mutations away from nature.

1/ The Structure Challenge 🧩

We entered gRNAde into Eterna OpenKnot: a CASP-style blinded, wet-lab competition by @eternagame.org

The result? Parity with the world's best humans, and big gains over Rosetta, RFdiffusion.

We can automate expert intuition for complex RNA folding!

Introducing gRNAde: our own little "AlphaGo Moment" for RNA design! 🧬🚀

📝: tinyurl.com/gRNAde-paper

Unlike proteins, RNA design has long relied on "wisdom of the crowd" (human experts) or the slow crawl of directed evolution — gRNAde changes that! 🧵👇

Wow thank you, Jamie, for sharing! 🤗

To make future progress it’s worth revisiting the past. From Olke Uhlenbeck, “Keeping RNA Happy”
pmc.ncbi.nlm.nih.gov/articles/PMC...

Beyond structure-based biomolecule design Dynamics, black-box data, and the antedisciplinary frontier of biomolecule design

🚀🧬 Beyond Structure-based Biomolecule Design

Its an important moment for structure-based biomolecule design: models starting to work and action shifting from academia to industry.

So what are the next scientific problems academia could be thinking about?

chaitjo.substack.com/p/beyond-str...

And at big industrial labs with large budgets, the scientists are just able to do a lot more intuition-building. They are able to try ideas a lot faster, and get a 'feel' for what ideas will work much faster as a result.

Why do 'frontier' labs train the best models?

I think its because training deep learning models is less like science/engineering, and more like cooking. It takes some time to develop the intuitions around learning dynamics of big models.