Alexis Molina

PhD @ AITHYRA | AI for Cell and Tissue Bio x Drug Discovery

also prev: @alexismolinamr

[Alexis Molina @alexismolmr]

New preprint: Do more complex single cell perturbation models generalize better to unseen perturbations and cellular contexts? Not necessarily.

We find that generalization depends more on aligning model inputs with the response being predicted.

https://t.co/3JbwgSj0fo

1/7

[Loïc A. Royer 💻🔬⚗️ @loicaroyer]

What if any dataset, any size, any number of dimensions, opened in a browser tab from a link? 🔬🧪💻

Luxar is out today: write it in Python, share it as a link, explore it in any browser. Open source. 🧵 @biohub

[Jayden Teoh @jayden_teoh_]

I've seen "the Bitter Lesson" stifle academic research many times, including my own. Too often, promising ideas are dismissed as "not bitter lesson–pilled enough". This is, of course, not Sutton's fault, but a misinterpretation of his original message.

Yes, it would be nice to have very clean ideas that scale with computation instead of relying on human inductive biases. But it is important to note that research is circumstantial; it happens within the constraints of our time.

Time and time again, we've seen ideas that, in hindsight, look "not bitter lesson–pilled", be eventually replaced by cleaner, more general approaches. But that doesn't mean those earlier ideas were misguided. Often, they were exactly what allowed us to make progress at the time. And I wish we, as a research community, did not dismiss those ideas/works at the moments they transpired.

RL post-training in its current form, in my opinion, is also not particularly elegant. Much of its success relies on extensive environment construction and data engineering to patch missing capabilities in frontier LLMs (or, in other words, "benchmaxxing"). Yet it's incredibly timely! It takes advantage of a time where building these data flywheels is increasingly feasible for well-resourced labs. We'll eventually see cleaner, more general approaches that replace parts of this machinery.

RLMs are another great example. Giving agents a REPL environment to programmatically manage their own context was a good step forward. I would argue it has also changed the way labs trained their model. And because of that, agents become better in managing their own context and that enables "cleaner" ideas like Context Language Models.

This applies more broadly. We're beginning to see a shift toward methods that "go where the agent goes", and this building on the foundations laid by foundation model pre-training research, be it ugly or not.

If I were to add to Sutton's tweet, I think we need more clever methods that don't necessarily have to be "bitter lesson–pilled", but challenge existing conventions and unlock new capabilities that overcome today's bottlenecks. An idea doesn't necessarily have to survive scaling forever to be an important contribution to AI progress.

[Richard Sutton @RichardSSutton]

Brute force computation is not enough for intelligence. Clever new methods are needed. The science of AI is far from finished. IMHO

[Francesco Ruscio @francescoshox]

Very happy to share the first paper of my PhD: Normalizing Flow Maps! (NFM)⚡️

We make likelihood evaluation in flow maps fast and tractable by replacing the DiT backbone with a normalizing flow🔄.

A big thanks to @tim_hitge, @mmbronstein, @bose_joey and the rest of the team!

[Rohit Singh @rohitsingh8080]

Over the course of a few weeks, preprints from Xinyi Zhang, ours, and @cbg_ethz labs independently noticed something structural about current perturb-seq datasets.

We all saw similar sparks but followed them in different directions. I'll talk about ours here.

[Alexis Molina @alexismolmr]

The broader lesson:

Generalization depends more on matching the model’s information to the response component being predicted than on increasing model complexity.

Work guided by Xinyi Zhang at AITHYRA.

Feedback is very welcome!

7/7

[Alexis Molina @alexismolmr]

New preprint: Do more complex single cell perturbation models generalize better to unseen perturbations and cellular contexts? Not necessarily.

We find that generalization depends more on aligning model inputs with the response being predicted.

https://t.co/3JbwgSj0fo

1/7

[Alexis Molina @alexismolmr]

We instead use gene coessentiality information, which is better aligned with perturbation-response variation.

Simple linear and MLP models using this prior outperform state-of-the-art architectures across unseen perturbations and unseen cell lines.

6/7

[Stephanie Wankowicz @stephanie_mul]

BioxAI is scale-pilled. AF2 worked because the target (3D structure) + metrics were well-defined, not because of data scale. Most biology problems have undefined targets. Scaling can multiply ambiguity. Define the target & metrics, then think about scale. https://t.co/YxSQj3YsSi

[Ihtesham Ali @ihteshamali]

Jennifer Doudna won the Nobel Prize for gene editing and went on Bloomberg to say the chatbots everyone is betting on cannot innovate at all. Every promise Silicon Valley is making about AI curing disease just hit the one person qualified to check it.

She has spent her whole career inside the actual frontier of curing disease.

So when she talks about what AI can and cannot do in biology, she is not guessing. She is reporting from inside the lab.

Her words were blunt. She is not seeing chatbots innovate. They summarize data. They write reports. They do not come up with a brand new idea nobody has ever had.

Then the interviewer pushed. So you're saying AI can't innovate?

Doudna did not flinch. She does not know if it can't. She just does not see it doing it right now.

This lands harder when you remember who is making the opposite case. Sam Altman says AI will eliminate disease within five years. Larry Ellison says AI will cure cancer in a 48 hour window.

An OpenAI executive even floated that the company should get a cut of sales on any drug discovered through ChatGPT. Doudna answered that in two words. Good luck.

Even the cancer specialists Altman is selling to keep warning that cancer is not one disease but hundreds, each needing its own cure, and that compute does not skip the years of lab work.

Her reason is simpler. Biology is hard. You cannot simulate your way to an understanding of the human body.

The people promising cures are the ones selling the tool.

The person who actually won a Nobel building them is telling you it has not happened yet.

Source: Bloomberg Originals

Watch the full video on their official channel.

[Martin Pacesa @MartinPacesa]

Shows very nicely why the current “virtual cell” efforts are just a waste of compute. People throwing data at models hoping that they learn principles without the proper context, rather than putting effort in to actually capture the underlying biochemistry/biology.

[Rafael Irizarry @rafalab]

A flaw in Person-Δ may be overstating progress in single-cell perturbation prediction models. Pearson warned about this in the 19th century: reusing the same controls induces spurious correlation. Split the controls, and much of the claimed prediction power fades. Link below 👇