←
CellType151 posts
CellType avatar

CellType

@celltypeinc

On-prem biological LLMs that simulate human biology and learn from your data. BioSimBench and BioSimGym (RL environments): celltype.com/b/x. YC W26.

◷ Joined February 2026▧ 61 Photos and videos
CellType avatar
CellType@celltypeinc·23h

Larger models score higher. Within one family, every step up in size gets closer to the measurement. This biology can be learned, and the curve is still rising.

Post attachment
CellType avatar
CellType@celltypeinc·23h

First results: frontier models get part of it right. The best scores 0.76, well short of how closely a repeat of the same experiment agrees. There is a long way to go, and this can be learned.

Post attachment
CellType avatar
CellType@celltypeinc·23h

So we turned the measurements into questions. Describe a printed tumor, change its neighborhood, and ask a model how the genes respond. Every answer is graded against the real experiment. The data are new. No model has seen them.

CellType avatar
CellType@celltypeinc·23h

To learn rules you need controlled experiments. Using the RASTRUM platform, the @InventiaLifeSci team builds and measures human tumor models across biological contexts: which cancer cells, which neighbors, and in what mix. Any tumor neighborhood can be built and measured.

0:12
CellType avatar
CellType@celltypeinc·23h

Nobody has written down this grammar yet. If AI could learn it, we could match a patient's tumor to a model of it, build that model, and ask what happens if we remove a cell type or add a drug, before running the experiment.

CellType avatar
CellType@celltypeinc·23h

In language, the same word means different things in different sentences. Context decides. Tumors work the same way. The same cancer cell behaves differently beside a fibroblast, beside a T cell, or beside both. Cells signal to each other, and the signals change the cell.

Post attachment
CellType avatar
CellType@celltypeinc·23h

Modern cancer therapies act on the tumor's immune environment. Some patients respond. Others, with the same diagnosis, do not, and we still cannot say why. The answer is in the neighborhood: cancer cells, fibroblasts, immune cells, and the matrix between them.

CellType avatar
CellType@celltypeinc·Nov 5

See the full BioSimBench leaderboard (twelve frontier models) and every task here: celltype.com/b/x

Can LLMs simulate biology? We tested frontier models on real lab experiments: predict how cells respond to drugs, gene knockdowns, and immune signals. Even the best get only part of the response right. Lots of room to improve. Introducing BioSimBench, built on @NVIDIAAI. 🧵

1:20
CellType avatar
CellType@celltypeinc·Nov 4

Why simulate? Drugs, genes, and cell contexts create far more experiments than any lab can run. A co-scientist reasons about mechanisms. A simulator answers a different question: in this cell, after this intervention, what will the experiment measure?

CellType avatar
CellType@celltypeinc·Nov 4

Built on NVIDIA. RAPIDS made the pipeline behind every task at least 4x faster than Scanpy, and up to 14.4x on single steps. Our models were trained on NVIDIA Blackwell GPUs. BioSimGym follows NeMo RL interfaces. @NVIDIAHealth @NVIDIAAI

Post attachment
CellType avatar
CellType@celltypeinc·Nov 4

Biological simulation can be learned. We trained a 27B model in BioSimGym, our RL environments for biology. It outscores the best frontier model on every BioSimBench task.

Post attachment
CellType avatar
CellType@celltypeinc·Nov 4

The task: given a cell and a drug, gene knockdown, or immune signal, predict which genes go up and down. Five tracks, one leaderboard.

0:10