BIOAI READER GUIDE
How should we compare virtual cells, AI models, and companies?
Compare single-cell models, structure prediction, and protein design by inputs, outputs, tasks, and validation conditions. Distinguish model papers from company capabilities.
BioAI Daily · Reader guide · updated and sources checked 2026-10-08
First define what is being predicted
AI virtual cells are a research direction that uses biological data to build computational representations of cells and predict their states and changes. Single-cell foundation models can address some of these tasks. Molecular structure prediction and protein design address other levels of biology. A company is an R&D organization; comparing it also requires specific products, data, and public evidence. [Cell · 2024 · Research vision and evaluation framework] [Nature Methods · 2024 · Single-cell foundation model paper]
This comparison uses original paper versions to explain task boundaries. It does not rank the strongest models or extend the experimental results of one version to the capabilities of an entire company.
Four kinds of objects, four ways to read the evidence
| Object and category | Main inputs and outputs | Evidence to inspect | Limits of comparison |
|---|---|---|---|
| scGPT · Single-cell foundation model | Learns cell and gene representations from single-cell omics data. Specific outputs depend on the downstream task configuration. | The original paper’s tasks, fine-tuning methods, test sets, and baselines. | Cell representation quality and prediction of unseen perturbations require separate validation. [Nature Methods · 2024 · Single-cell foundation model paper] |
| AlphaFold 3 · Structure prediction model | Predicts complex structures containing proteins, nucleic acids, or small molecules from molecular composition and related inputs. | Independent structure tests, interface accuracy, and confidence; consider the limitations stated in the original paper. | Structure prediction does not directly establish cell responses, drug efficacy, or patient benefit. [Nature · 2024 · Biomolecular complex structure prediction paper] |
| RFdiffusion · Protein design method | Generates protein structures under design conditions. Sequence design, screening, and experiments follow. | The number of generated candidates, number experimentally tested, hit criteria, and independent validation. | Structure or binding validation supports only the designs and conditions tested. [Nature · 2023 · Protein design method and experimental paper] |
| Isomorphic Labs · Company | Official announcements describe its partners and drug discovery projects. A company name is not an executable model version. | Contract or collaboration announcements, corresponding research results, and subsequent disclosures for specific projects. | The existence of a collaboration, platform performance, and candidate-drug progress each require their own evidence. [Isomorphic Labs · 2024-01-07 · Company collaboration announcement] |
Which conditions should a virtual-cell project specify?
Our reading suggestion is to first specify the measured inputs and whether the target is a cell state, a population-average change, or a post-perturbation distribution. Then check whether training and testing share donors, cell lines, drugs, or targets. Randomly splitting cells from one dataset and testing in an entirely unseen biological context address different validation questions.
With only unpaired CTRL and PERT populations, we can compare state distributions and average responses. Without tracking or paired measurements, these data cannot identify each cell’s true trajectory. RNA-level predictions also cannot automatically be interpreted as changes in protein abundance or modification. These limits follow from the measurements and experimental design; a model name does not remove them.
Record the paper’s tasks, data splits, simple baselines, evaluation metrics, and failure cases before deciding whether its results apply to your research. The virtual-cell vision article itself discusses requirements for data, evaluation, and biological accuracy. It does not establish that any model has fully reproduced a cell. [Cell · 2024 · Research vision and evaluation framework]
From a model paper to a company profile
Topic pages link related source records, while company pages link official sources and reviewed events. Automated topic classification helps discovery; it does not establish an affiliation or authorization between paper authors and a company. When comparing companies, look for specific model versions, project stages, and traceable collaboration events in their profiles.
Original sources
- How to build the virtual cell with artificial intelligence: Priorities and opportunities
Cell · 2024 · Research vision and evaluation framework - scGPT: toward building a foundation model for single-cell multi-omics using generative AI
Nature Methods · 2024 · Single-cell foundation model paper - Accurate structure prediction of biomolecular interactions with AlphaFold 3
Nature · 2024 · Biomolecular complex structure prediction paper - De novo design of protein structure and function with RFdiffusion
Nature · 2023 · Protein design method and experimental paper - Isomorphic Labs kicks off 2024 with two pharmaceutical collaborations
Isomorphic Labs · 2024-01-07 · Company collaboration announcement