Let’s discuss the scaling law of virtual cells. A Nature Methods paper (nature.com/articles/s41592-0…) published yesterday is being interpreted by some as evidence that scaling laws do not hold for virtual cells. I read it in detail, and here are my 2 cents: It is a useful benchmark, but not a direct test of scaling laws in causal single-cell foundation models or perturbation-native virtual cells. The paper mainly studies PCA, scVI, Geneformer, and SCimilarity (which are relative small models) on observational atlas-style pretraining, with perturbation evaluation limited to a narrow Tahoe small-molecule/cancer-cell-line setting. These are important baselines on scFMs that focuses on learning cell embeddings, but they are not large-scale causal perturbation models (e.g, diffusion-based virtual cells, or other modern architectures designed natively for causal perturbation biology). The metrics also matter. Cell-type F1 and batch-integration AvgBIO are reasonable atlas/embedding metrics, but they are also tasks that can saturate quickly. They are not direct measures of causal perturbation prediction, target ranking, rare differential-expression tails, OOD genetic perturbations, or generalization across biological contexts. The “learning saturation point” in the paper is useful, but it is not really a scaling law. It asks: what is the smallest pretraining size that reaches within 95% of the best observed score on this benchmark? That is a helpful diagnostic, but it can be overinterpreted when the downstream task itself is saturated. The perturbation result is, IMO, limited: a few selected Tahoe-100M small molecules across several cancer cell lines, evaluated with genewise R²/MSE. The paper itself reports that a “no-change” baseline beats fine-tuned models for most drugs, which says as much about the evaluation regime as about model scaling. In fact, our scGPT work already showed three years ago that simply scaling the number of observational cells saturates quickly after a few million cells. So I agree with the warning: naive “more atlas cells = better virtual cell” is not enough. But that is not the real scaling question. In X-Cell, we study scaling across multiple axes: number of perturbation cells, number of biological contexts, perturbation diversity, and model parameters. On our Perturb-seq-scale data, we observe clear and encouraging scaling behavior. Similar trends are emerging from other perturbation-native virtual cell efforts as well. The important question is not: can more atlas cells improve cell-type F1? It is: with larger Perturb-seq datasets, larger models, better architectures, and harder OOD splits, can we predict causal cellular responses across genes, combinations, doses, cell states, and contexts? For X-Cell and the next generation of virtual-cell models, the goal is not just better embeddings. It is target ranking, rare DE tails, counterfactual biology, and prospective perturbation prediction. So my reading is: this paper is a useful caution against naive scaling, not evidence that scaling laws do not apply to virtual cells. The exciting regime is still open: scaling the right data, the right models, and the right objectives for causal cellular biology.

🇨🇦🇨🇦🇨🇦🔥🔥🔥

Jeff Bezos and Vik Bajaj give a first peek at Prometheus, an ARCH co-founded company to create an artificial general engineer. ARCH’s largest investment ever, from seed formation.

great thread about the most fundamental steps in single cell data analysis! 🙏

"I thought that by the end of the century was a stretch. Now I think it's too conservative." 🚀🔥🚀🔥

Congratulations ! @AmolAVerma & @DrFahadRazak This new initiative will provide huge opportunities for AI & Health in Canada! 🇨🇦

biology will deliver most of the returns in the stock market over the next 20 years

This is one of the most important parts of the Mythos 5 announcement: Drug design. Anthropic says its internal protein design experts used Mythos 5 to accelerate parts of the drug design process by around 10x. In one example, Mythos 5 used protein design and bioinformatics tools without (!) human assistance and matched or beat skilled human operators. It means the model could run parts of the early drug discovery loop itself: choosing binding sites, selecting tools, running protein design workflows, recovering from failures, and generating promising candidates. And here is what people still dont understand outside of our community: AI is moving from "assistant that explains science" to "agent that can actually execute parts of scientific work." The future of drug discovery may look less like one scientist manually testing ideas one by one, and more like thousands of AI-driven research loops running in parallel, with humans validating, interpreting, and deciding what actually moves forward. I keep thinking about what Demis Hassabis said: we are entering the golden age of science. There is probably no doubt about that anymore. Medicines are being developed faster and more precisely, and previously incurable diseases are becoming treatable.

As a frequent customer of air Canada, I have mixed feelings about this news 😂🇨🇦

If you are interested in knowing the latest AI in medical imaging research, @JunMa_AI4Health is a must follow!

Congrats @gelarehzadeh ! This is a great work about how spatial single cell omics can help cancer research!

Is biology fundamentally harder than vision or coding? Anthropic ran frontier models on one task: retrieve viral sequences from NCBI. Same query, three runs. Claude Sonnet 4 returned 106, 15, then 5 sequences. Ground truth: 266. One run estimated an Ebola outbreak origin as 1922. The fix wasn't a better model , but a thin deterministic wrapper (gget) hit ~100%. Bio databases were built for humans clicking browsers. Filtering logic lives in web UIs, metadata is inconsistent, identifiers drift between sources. No LLM fixes broken pipes. NCBI has 30 databases that need this treatment. That work hasn't started yet. We will soon release more results on how frontier agentic workflow can work with biological database including large-scale perturb-seq data at @Xaira_Thera . Stay tuned 🙏😁

Canada’s AI strategy: AI for all!

Re-tweeting. We need rapid policies that make USA companies more competitive and create a market mechanism that preserves patient options, while preventing harm to biotech innovation cycle.

Our workshop of Foundation models for medical vision at CVPR 2026 is happening on June 3rd! Four keynotes spanning oncology, pathology, virtual patients, and agentic systems, plus challenge winners breaking down their solutions. If you're at CVPR, this is the room to be in 🔥

This is just incredible! 🔥

Today at #ASCO26, more results about the newest clinical trial of daraxonrasib: In the Phase III RASolute-302 trial, a once-daily RAS(ON) inhibitor nearly doubled median overall survival (13.2 vs 6.7 months) and reduced the risk of death by ~60% versus chemotherapy in previously treated metastatic pancreatic cancer. KRAS was once considered “undruggable.” This is what persistence in science looks like. A landmark moment for targeted therapy and for patients who desperately need better options. #CancerResearch #PancreaticCancer

Advanced world knowledge: x.com/BoWang87/status/202707…

This is not looking good for Canada. The irony is that Canada helped pioneer breakthroughs behind AI and GLP-1s, two industries now creating trillions of dollars in value, yet much of that value is captured elsewhere. Canada produces world-class science, but we continue to struggle with commercialization at scale. Too little growth capital, insufficient compute infrastructure, and too few pathways to build globally competitive companies. We need to invest across the full stack: talent, capital, compute, regulation, and company-building, not just research grants. 🇨🇦

Opus 4.8 is almost as good as Claude Mythos in most of biology benchmarks! 🔥🔥