Paper alert 🚨 @ MLGenX @iclr-conf.bsky.social 2026! PertOmni - CLIP-style multimodal representation-learning framework for contrastive alignment of perturbation readouts and textual embeddings.

Paper alert 🚨 @ MLGenX @iclr-conf.bsky.social 2026! PerturBERT openreview.net/forum?id=ZsD... - Most transformers in single-cell omics (e.g. scGPT) learn gene co-expression through pre-training on masked gene expression levels.

Caris is launching a WGS+ML blood test for early cancer detection. Impressive sensitivity: 56% stage 1 and 70% stage 2 - if it holds, it would blow competitors out of the water

Nano Banana 2 test. Humans are still safe 😌

"Constitution" is a bad name for model preference guidelines doc. It suggests, among others, public legitimacy, where there is none.

Gemini Deep Think 3 is the world's most capable model by many measures, huge amounts of progress on reasoning benchmarks and more. Available right now via the Gemini App for Ultra subscribers and in the API soon : ) x.com/OfficialLoga...

How to fake a robotics result: a short blog post listing many sins which annoy me (many of which I am guilty of from time to time, to be fair) open.substack.com/pub/itcanthi...

Haha @cpaxton.bsky.social is on fire: open.substack.com/pub/itcanthi... Some useful tips in there even for non-roboticists looking to make their paper artificially look good

And for a lot of Gen X journalists and academics, the answer to (a) — assuming existing skills and plans — is legit "no." AI can be useful, for sure, but the paths it differentially advantages are not the paths where they have accumulated momentum, expertise, and social capital. +

A quick story on how we matched genes across two datasets with different Ensembl versions. 1. There must be a tool out there. Ensembl ID History converter ofc! 2. Doesn't match Ensembl search outcomes due to a bug 3. Lesson: use this client instead github.com/Ensembl/ense... !

Also awesome for finding posters at a conference!

Predicting cell state in previously unseen conditions has typically required retraining for each new biological context. Today, Arc is releasing Stack, a foundation model that learns to simulate cell state under novel conditions directly at inference time, no fine-tuning required.

Introducing DroPE: Extending Context by Dropping Positional Embeddings We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solution’s simple: treat them as a temporary training scaffold, not a permanent necessity. arxiv.org/abs/2512.12167 pub.sakana.ai/DroPE

One very familiar pattern in AI and science right now is going from a lot of false starts on hard tasks (there have been near-misses where AI appears to solve an Erdos problem but just finds an old solution no one knew about) to actually doing the thing soon after. Three Erdos problems in 3 days.

We introduce epiplexity, a new measure of information that provides a foundation for how to select, generate, or transform data for learning systems. We have been working on this for almost 2 years, and I cannot contain my excitement! arxiv.org/abs/2601.03220 1/7

Finished the essay. moultano.wordpress.com/2025/12/30/c...

Autonomous RIVR delivery robots in Pittsburgh

A quick sunday rewrite of an old blog post about how one should evaluate the effectiveness of an empirical paper: open.substack.com/pub/emergere...

Good researchers obsess over evals The story of Olmo 3 (post-training), told through evals NeurIPS Talk tomorrow. Upper Level Room 2, 10:35AM. Slides: docs.google.com/presentation...

Elon’s power is that he offers a positive vision of the future. This attracts employees, funding, support. There’s a massive techno positive hole and he fills it.

Active learning with DrugReflector beats SotA in phenotypic hit-rate for virtual screening. Includes a sc perturbation dataset with 10 lines and 104 compounds. Out in @science.org now! Grateful to Cellarity and @fabiantheis.bsky.social for the opportunity to contribute to this outstanding project!

What if we did a single run and declared victory

Community notes when

Yeah this is my biggest “AGI hype is not real” is that almost no one at these companies behaves like it’s real

My skepticism of LLM-as-scientist stems from how imbalanced the literature is. Median paper is mildly negative result presented as positive, it's unclear how to RLHF on good hypothesis vs. bad hypothesis, etc. We barely know how to teach this skill, how can we RLHF it

For folks considering grad school in ML, my advice is to explore programs that mix ML with a domain interest. ML programs are wildly oversubscribed while a lot of the fun right now is in figuring out what you can do with it

A must-read before you jump on your first omics project - the top response here www.reddit.com/r/bioinforma...

I think scientists thought people could tell apart the serious science from the bad fluff and ideological work that we all mostly ignore. We were not ready for people to start conflating all of them together

The more rigorous peer review happens in conversations and reading groups after the paper is out with reputational costs for publishing bad work

There are people, in tech (and now in the government!), who will mislead you about what current AI models are capable of. If we don't call them out, they'll drag us all down.

Oops I read my parrot a math textbook and now it keeps squawking out the answer to unseen math competitions

Excited to share that I started my summer at @genentech.bsky.social BRAID Perturbation team in SF with Alex Wu! It's my first time on the West Coast - If you are around and would like to talk about ML and/or biology, hit me up! Looking fwd to the AI x Bio Unconference tomorrow 🚀

Analyzing your single-cell data by mapping to a reference atlas? Then how do you know the mapping actually worked, and you’re not analyzing mapping-induced artifacts? We developed mapQC, a mapping evaluation tool www.biorxiv.org/content/10.1... from the ‪@fabiantheis lab. Let’s dive in🧵

From cell lines to full embryos, drug treatments to genetic perturbations, neuron engineering to virtual organoid screens — odds are there’s something in it for you! Built on flow matching, CellFlow can help guide your next phenotypic screen: biorxiv.org/content/10.1101/2025.04.11.648220v1

Just a gentle reminder that deceptive hyping in scientific publications (which includes preprints) is actually antithetical to the core mission of the scientific process. We can stay grounded, truthful, humble while being ambitious. Revealing caveats, pitfalls & limitations speeds up progress.

Our paper benchmarking feature selection for scRNA-seq integration and reference usage is out now www.nature.com/articles/s41...! Keep reading for more about how we did the study and what we found out 🧵 👇 1/16

Exciting dataset! If you're looking for a complimentary scRNA-seq drug perturbations in healthy/primary tissue (PBMCs), check out our dataset with ~36% # drugs of Tahoe. proceedings.neurips.cc/paper_files/...

The fact that we seem to be marching straight towards another cold war, where AI is the defining technology, is hard to emotionally accept and even harder to deeply accept how hard some of the next few years can become.

Last moments of closed-source AI 🪦 : Hugging Face is openly reproducing the pipeline of 🐳 DeepSeek-R1. Open data, open training. open models, open collaboration. 🫵 Let's go! github.com/huggingface/...

Trying to identify preclinical models that resemble clinical tumors you work on? Check out our MOBER, now out in @science.org Advances! www.science.org/doi/10.1126/... . There's also a web app to explore the results mober.pythonanywhere.com

When a paper is published, any data must be easily and completely accessible, or the publication is a sham and should be retracted editorially.

1/7 Planning to build a single-cell atlas? Or wondering how atlases can be useful to your research? Read our guide on single-cell atlases www.nature.com/articles/s41... published in Nature Methods, by @lisasikkema.bsky.social, @khrovatin.bsky.social, Malte Luecken, @fabiantheis.bsky.social et al.

1/🚀 Excited to share RegVelo, our new cell model combining RNA velocity with gene regulatory network (GRN) dynamics to model cellular changes and predict in silico perturbations. Here's how it works and why it matters! 🧵👇 biorxiv.org/content/10.1101/2024.12.11.627935v1

4️⃣ “A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types” @chatgtp.bsky.social neurips.cc/virtual/2024...

What are the most interesting datasets and benchmark-related work for ML in drug discovery at NeurIPS? We’ll be at the conference doing short interviews with researchers and handing out some Polaris merch! Here’s who we have on the shortlist. 🧵

I’m told burnout comes less from having too much to do but rather feeling like what you have to do is out of your control (and/or unpleasant work). So, working out substantial changes in what you’re obligated to be doing is the best way out. Thus making space for something new/interesting!

Public service announcement: Conference deadlines & other artificial deadlines are not excuses to put out half-baked work. If u want to put out half-baked work, u shud clearly state in the paper that results are preliminary & discuss areas of concern that might need careful verification. 1/

Here's a Cheminformatics Starter pack! Let me know if you would like to be added! And more importantly, do encourage people to join the blue skies :D

Three BioML starter packs now! Pack 1: go.bsky.app/2VWBcCd Pack 2: go.bsky.app/Bw84Hmc Pack 3: go.bsky.app/NAKYUok DM if you want to be included (or nominate people who should be!)