June 15, 2026
Cells, cash, and a huge flex
How to build a virtual cell and biology scaling laws
Biology’s AI moment? One bold bet has commenters ready to throw in cash and careers
TLDR: Markov Biosciences says bigger AI models trained on ordinary cell data can predict cell behavior better than more hand-built approaches, a potentially big shift for drug research. The early community mood is strikingly bullish, with at least one commenter ready to back the idea with both cash and career energy.
A San Francisco startup just dropped a very big claim: biology could be heading for its own AI-style breakout moment. Markov Biosciences says the path to a "virtual cell"—basically a computer model that can predict how cells behave—may not require endless expensive lab tweaking after all. Founder Adam Green argues the real bottleneck isn’t data, but computing power and asking the model the right question. In plain English: feed the machine lots of ordinary cell data, make it learn patterns at scale, and it may beat systems built with more hand-crafted scientific assumptions.
And the community reaction? Tiny thread, huge confidence. The standout response came from commenter d_silin, who skipped the cautious scientist routine and went straight to full-send venture energy: this is “one idea” they’d bet on with both money and skills. That’s the vibe here—not mild curiosity, but the kind of comment that reads like someone halfway through updating their résumé and wiring funds.
The underlying drama is classic tech-world culture war: careful, human-built models versus just scale it and let the machine learn. Green is basically reviving the old AI battle cry that brute-force learning can crush lovingly crafted expert rules. For some readers, that sounds thrilling; for others, it’s the kind of statement that makes traditional biology people reach for a stress ball. Even without a sprawling flame war, the mood is unmistakable: if this works, it’s huge. If it doesn’t, it’s one of the boldest swings in biotech right now.
Key Points
- •Markov Biosciences argues that virtual cell models can benefit from a scaling-based approach analogous to the "bitter lesson" in machine learning.
- •Adam Green says the limiting factors for useful virtual cells are compute and loss function design rather than the amount of available biological data.
- •Markov’s approach models single-cell RNA-seq as a ranking problem instead of using raw counts.
- •The article says observationally pre-trained virtual cell models improve monotonically on unseen perturbation prediction as model size increases.
- •The episode also covers a cancer-related antibody-drug conjugate case study and references Markov’s recent prediction work in that area.