> ## Content Index
> Fetch the complete content index at: https://aiadoption.org/llms.txt
> Use this file to discover other available public pages before exploring further.

# Virtual cell models emerged as a new biomedical AI frontier
- URL: https://aiadoption.org/ai-analysis/virtual-cell-models-emerged-as-a-new-biomedical-ai-frontier/
- Published: 2026-08-02T08:09:03.000Z
- Updated: 2026-08-02T08:09:03.000Z
- Description: The aspiration: predict cellular responses to drugs and genetic perturbations without wet-lab experiments. The current reality: experimental validation is still required, and smaller task-specific models continue to outperform on specific benchmarks.
- Author: Manu
- Tags: AI Analysis, AI in Healthcare, AI Research, Generative AI

If your map of biomedical AI in 2025 still organises around the established categories (protein structure prediction, drug-target interaction, medical imaging, clinical decision support), a new category has now formed at sufficient scale to track separately. Virtual cell models (AI systems that predict cellular states and responses to drugs, genetic perturbations, and environmental conditions) emerged as a distinct frontier in 2025\. The Arc Institute released Evo 2, a 40-billion-parameter DNA language model. DeepMind released AlphaGenome, a multimodal model. STATE arrived as a perturbation-response model. Research publications on virtual cell models rose from single digits in 2018 to 24 by 2025\. The aspiration: predict cellular responses without wet-lab experiments. The current capability: substantial but still requires experimental validation. The frontier is real; the timeline to broad utility is uncertain.

*Publications on virtual cell models, 2018–25:* 

### The publication count trajectory shows the formation pattern of a new research category. 

The 2025 inflection coincides with the release of foundational models that gave researchers new tools to apply to the cellular modelling problem. Evo 2 from the Arc Institute, AlphaGenome from DeepMind, and STATE from research teams working on perturbation-response models all became available in 2025\. The research community organised around these models faster than typical for new biomedical AI categories.

The technical aspiration of virtual cell models is substantial. The goal is to predict cellular responses to interventions (drug treatments, genetic modifications, environmental stresses) computationally, without requiring wet-lab experiments for each prediction. If achieved at high fidelity, this would substantially accelerate drug discovery (predicting drug efficacy before clinical trials), basic biology research (testing hypotheses about cellular function), and personalised medicine (predicting how an individual patient's cells will respond to specific treatments).

The current capability falls short of this aspiration in important ways. Current virtual cell and genomic foundation models still lag behind smaller, task-specific models on several benchmarks. The 40-billion-parameter Evo 2 was outperformed by the 200-million-parameter GPN-Star on multiple variant effect prediction tasks. The scaling assumption that drove Evo 2's parameter count did not produce proportional capability gains.

*Virtual cell model performance, variant effect prediction:* 

### The trajectory observation has two structural components.

**The first component:** the category is real and forming rapidly. The 2025 release of multiple foundation models, the doubling of research publications since 2022, the substantial industry interest from DeepMind, Arc Institute, and other research organisations all indicate genuine research momentum. The category will exist as a distinct field through 2026-2028.

**The second component:** the timeline to broad utility is uncertain. The aspiration of replacing wet-lab experiments with computational predictions requires reliability levels that current models do not achieve. Experimental validation remains required. The "virtual cell" framing implies a future capability that the 2025 models do not yet provide.

### Three structural drivers shape the development trajectory.

**The first driver:** data is structurally limiting. The Tahoe-100M dataset (over 50 cancer cell types exposed to more than 1,100 drugs) and the BaseData dataset (9.8 billion genes obtained through metagenomic mining) are the largest publicly available training datasets specifically for cellular modelling. Both were released in 2025\. But cellular state prediction requires data on cellular responses across many cell types, many perturbations, and many measurement modalities. The current data infrastructure covers a small fraction of this space.

**The second driver:** methodological direction is contested. Some research groups pursue large-scale general-purpose models (Evo 2, AlphaGenome); others pursue smaller specialised models (GPN-Star, STATE). The 2025 evidence suggests the smaller specialised approach is producing stronger benchmark results. But virtual cell modelling involves multiple subtasks (transcription regulation prediction, drug response prediction, cellular state prediction across modalities) where different methodological approaches may be optimal. The field has not yet converged on architecture or training-method consensus.

**The third driver:** evaluation infrastructure is immature. Standardised benchmarks for virtual cell model performance are emerging (the various benchmarks referenced in Ye et al. 2025) but the field does not yet have evaluation frameworks comparable to the maturity of ProteinGym for protein language models or FoldBench for cofolding models. Without strong evaluation, progress assessment is harder, comparison across methods is harder, and the trajectory to broadly useful models is slower.

### The trajectory implications follow.

**The first implication:** virtual cell models warrant separate strategic attention from established biomedical AI categories. The category is forming at speed; the strategic landscape (which research groups lead, which approaches dominate, which evaluation frameworks become standard) is being shaped in 2026-2027\. Organisations that wait to engage will enter as the strategic terrain is already set. Organisations that engage now have the opportunity to shape the direction.

**The second implication:** data infrastructure is the highest-leverage investment in the field. The 2025 dataset releases (Tahoe-100M, BaseData) demonstrate that data infrastructure investment produces field-wide acceleration. Organisations that fund similar dataset development, public release, or curation activities will catalyse field progress and benefit from the resulting capability advances. The leverage is large because the field is currently data-constrained.

**The third implication:** benchmark and evaluation infrastructure is similarly high-leverage. The field needs standardised evaluation frameworks comparable to ProteinGym for proteins. Organisations that develop and publish such frameworks will shape what "good performance" means in virtual cell modelling, with substantial influence on research direction.

**The trajectory:** through 2026-2028, virtual cell modelling will likely continue growing in research output (the 24 publications of 2025 could exceed 100 by 2028), see additional foundation model releases from major research organisations, and develop more standardised evaluation frameworks. It will most plausibly produce specific narrow applications with practical utility (drug response prediction in cancer cell lines, transcription factor binding prediction, particular variant effect predictions) before achieving the general "virtual cell" aspiration, with structural data infrastructure investment remaining the field-defining variable throughout.

> **The trajectory observation:** virtual cell models are the most likely candidate for the next major biomedical AI category to mature toward clinical utility. The timeline, though, is plausibly 3-7 years rather than 1-2 years. Strategic plans that engage with this longer trajectory will be operationally aligned; plans that expect rapid translation from 2025 model releases to clinical utility will be misaligned with the technical trajectory.

For organisations setting biomedical AI strategy in 2026, the virtual cell modelling category is the most important new development to track. The strategic decisions made in 2026-2027 (data infrastructure investment, methodological direction, research partnership choices) will shape competitive positioning when the technical capability matures in 2028-2032\. Plans calibrated against the visible trajectory will be aligned with the actual development direction.