MIT Technology Review China Releases AI Virtual Cell (AIVC) Report

Over the past decade, AI tools represented by AlphaFold have solved the problem of designing "new parts"—predicting protein structures, properties, and interactions, and improving the discovery efficiency of candidate molecules. But the report points out that this type of AI drug development path makes it difficult to determine whether a target will bring a truly effective response in cells of a specific disease. AIVC advances the object of study from the molecular level to the cellular level, treating the cell as a dynamic system jointly constituted by gene regulatory networks, signaling pathways, metabolic processes, and phenotypes, with the prediction logic evolving from "target hit" to "cellular response."

The real-world pressure behind this shift comes from the "double-ten rule" in drug development: an average of ten years and more than one billion US dollars, with more than 90% of failures occurring at the clinical translation stage. The report believes that AIVC is expected in the future to compress experimental demand and costs, and to provide computational support for directions such as cell therapy and gene therapy. The farther goal is to extend from cellular response to systemic effects, answering the most expensive question in R&D: whether changes at the molecular level can truly translate into patient efficacy.

AIVC did not appear out of thin air. On the life sciences side, the Human Genome Project completed from 1990 to 2003 opened the era of large-scale datafication of life information, single-cell transcriptome sequencing around 2009 provided the data foundation at single-cell resolution, and from 2016 onward perturbation-response datasets began to be established; on the AI side, the Transformer architecture was proposed in 2017, and in 2020 AlphaFold 2 proved at the algorithmic level that deep learning can learn complex laws of life. The two threads converged, and the industry therefore regards AIVC as "the next AlphaFold moment."

The report breaks down the construction of AIVC into four layers: data, models, autonomous intelligence, and the dry-wet closed loop. At the data layer, four types of data need to work together: static cell data (mainly single-cell transcriptomes), perturbation-response data, clinical data, and biological prior knowledge that provides structured constraints and interpretability. Before data enters model training, it must undergo four processing steps: denoising, deduplication, standardization, and cross-modal alignment, among which cross-modal alignment is called by the report the most core and most difficult step.

The model layer is undergoing iteration from static representation to state transition and continuous evolution. Early self-supervised pretrained models represented by Transformer answer "what a cell is," with representative achievements including the Broad Institute's Geneformer; perturbation models introduce pre- and post-intervention control data to answer "how cells change," with representative achievements including Stanford University's GEARS and Arc Institute's STATE; the field is currently exploring dynamic modeling, using world models to predict the future state of the system in latent space to answer "how cells evolve," with representative achievements including Baiyao Technology's CellOS, Tongji University's AlphaCell, GenBio AI's AIDO Cell, and Arc Institute's Stack, among others.

It is worth noting that the report explicitly cautions: AIVC model parameter scale has jumped from the early million level to the hundreds of millions level, with some achievements exceeding the billion level, but unlike natural language models, simply expanding scale does not necessarily bring better zero-shot generalization capability; training objectives, perturbation coverage, cellular context, and task type are equally critical. Above the model, there is also a need for a scheduling layer that breaks down scientific research questions into executable tasks, namely autonomous intelligence, in which general large models are responsible for understanding questions, planning, and scheduling tools, AIVC models are responsible for predicting cell states and perturbation responses, and agents connect the two and interface with wet-lab validation.