Report frames AI virtual cells as emerging life-science infrastructure

On September 21, 2026, MIT Technology Review China released a report that was initiated by Baiyao Technology and guided by the administration of the Zhongguancun Science Park. The document serves as the primary source for the analysis of artificial‑intelligence‑driven virtual cells as a potential new infrastructure for the life sciences.

The report defines an AI virtual cell as a multimodal, multi‑scale neural model that is capable of representing and simulating the states and transitions of a biological cell. This definition emphasizes that the model is not limited to a single molecular view but integrates diverse data modalities across spatial and temporal scales.

From Model Size to Generalisation

According to the report, the strategic challenge is shifting from merely increasing model size toward improving generalisation, task adaptation, and the quality of perturbation data. This shift reflects a broader consensus among the dozens of research institutions, technology firms and pharmaceutical groups interviewed for the study.

The ambition articulated in the report is to predict cellular responses to genetic perturbations, drug exposures or environmental changes. Unlike approaches that focus only on molecular structure, the virtual‑cell framework aims to capture downstream phenotypic outcomes.

A key observation of the report is that private datasets linking dose, timing, cellular context and phenotype are considered a competitive advantage. The authors argue that such richly annotated data are more difficult to reproduce than publicly available datasets, thereby influencing the reliability of model predictions.

The Dry‑Wet Loop as an Industrial Threshold

The report identifies a principal industrial threshold: a dry‑wet loop that connects data generation, AI‑driven prediction, experimental validation and the re‑injection of results into the model. This closed‑loop workflow is presented as a prerequisite for moving virtual‑cell technology from research prototypes to scalable applications.

The dry‑wet loop concept implies that each iteration refines both the data reservoir and the model parameters, potentially improving predictive fidelity over time. However, the report does not provide quantitative evidence of loop efficiency, leaving the speed and cost of iteration as open questions.

Potential Applications and Remaining Gaps

The envisioned uses of AI virtual cells span basic research, target identification, drug development and cellular engineering. In each domain, the ability to simulate cellular responses could accelerate hypothesis testing and reduce reliance on costly wet‑lab experiments.

Despite the breadth of anticipated applications, the report explicitly states that clinical validation remains to be demonstrated. This gap highlights the current limitation of virtual‑cell predictions when translated to human health outcomes.

• Fundamental biology research
• Target identification for drug discovery
• Pre‑clinical drug development pipelines
• Cellular engineering for biomanufacturing

The list above reflects the categories most frequently mentioned across the interviews, yet the report does not quantify the relative maturity of each category. Consequently, the readiness of virtual‑cell technology may vary widely between basic science and industrial drug pipelines.

One limitation noted in the report concerns the dependence on high‑quality perturbation data. Since private datasets are highlighted as advantageous, institutions lacking access to such data could encounter barriers to achieving comparable model performance.

Another open question relates to the scalability of the dry‑wet loop. The report describes the loop conceptually but does not detail the throughput of experimental validation steps, leaving uncertainty about whether the loop can operate at the scale required by large pharmaceutical programs.

Future research could explore whether improvements in data sharing frameworks might reduce the advantage of proprietary perturbation datasets, thereby democratizing access to virtual‑cell technology.

Overall, the report positions AI virtual cells as a prospective infrastructure that could reshape how the life sciences integrate computational prediction with experimental validation, provided that challenges around data quality, loop efficiency and clinical translation are addressed.