In the early days of single-cell analysis, researchers treated cells like bags of RNA. Models like scVI, developed by Romain Lopez and Nir Yosef, revolutionized the field by using variational autoencoders to model the probabilistic nature of transcriptomic counts. However, biology does not operate in a transcriptomic vacuum. A cell's state is defined by a cascade of information—the "central dogma"—flowing from DNA accessibility to RNA expression and finally to protein synthesis.
The current frontier of research is moving away from these "siloed" models toward multimodal foundation models. These architectures aim to bridge the gap between different biological layers, integrating gene expression with chromatin accessibility (ATAC-seq), protein abundance, spatial coordinates, and even natural language descriptions.
Bridging the Central Dogma
The primary intuition behind this shift is that different "omics" layers provide complementary signals. While RNA tells us what a cell is doing now, DNA accessibility (epigenetics) might tell us what it is primed to do next. Christopher Blum and colleagues argue in Multi-modal Transfer Learning between Biological Foundation Models that treating DNA, RNA, and proteins as isolated data types ignores the biological reality of information flow. By aligning these modalities, models can learn representations that are more robust to the noise inherent in any single measurement.
For example, scMamba, introduced by Zhen Yuan and Jiajie Peng, uses a state-space architecture to integrate RNA and ATAC-seq data without the need for restrictive feature selection. This allows the model to capture the regulatory relationships between open chromatin and gene expression across hundreds of thousands of cells.
Integrating Expert Knowledge and Language
A significant bottleneck in single-cell research is the "interpretation gap." We can cluster cells based on their gene expression, but naming those clusters often requires manual, expert intervention. Researchers are now using Large Language Models (LLMs) to bridge this gap.
Suyuan Zhao and Zaiqing Nie developed LangCell, which aligns single-cell transcriptomes with natural language descriptions from biological ontologies. By projecting gene expression and text into a shared latent space, the model achieves zero-shot cell identity understanding. This means the model can identify a "memory B cell" in a new dataset even if it has never seen a labeled example of one, simply by understanding the textual definition of that cell type.
Similarly, GeneCompass, a foundation model by Ao Li and Handong Li, incorporates four types of prior biological knowledge—including gene regulatory networks and promoter sequences—to move beyond simple pattern matching toward a mechanistic understanding of gene regulation.
Scaling to the Atlas Level
As the scale of biological data grows to hundreds of millions of cells, as seen in the scCompass database, the computational challenge shifts from simple modeling to efficient integration.
Traditional contrastive learning, which compares every cell to every other cell, hits memory bottlenecks quickly. To solve this, Suyuan Zhao and Jiahuan Zhang introduced CellLM, which uses a divide-and-conquer contrastive learning approach. This allows the model to learn from massive datasets of both healthy and cancerous cells simultaneously.
Furthermore, frameworks like Moscot, developed by Dominik Klein and Giovanni Palla, use optimal transport theory to map these massive datasets across time and space. This enables researchers to track how a cell's multimodal state evolves during development or in response to a drug.
From Representation to Generation
The ultimate goal of these multimodal representations is not just to describe cells, but to predict how they will change. Yoel Shoshan and the team behind MAMMAL emphasize that drugs, targets, and cellular responses must be analyzed together. By learning the "molecular language" that connects chemical structures to transcriptomic changes, these models can predict drug sensitivity and synergy in silico.
Generative models like CellFlow and SCLD take this a step further. By learning the underlying "flow" or "diffusion" between biological states, they allow researchers to simulate how a cell population will respond to a specific genetic perturbation, effectively turning AI into a virtual laboratory for hypothesis testing.