Written by Arpita Kulkarni (Director of the Functional Genomics and Bioinformatics Core at HMS/BIDMC).
Across the life sciences, a quiet anxiety is growing: what is the role of the experimental biologist in the age of AI?
This sneak peek into the upcoming Multi-Omics Playbook offers a practical roadmap for how experimental biologists—especially in single-cell biology—can take leadership roles in shaping the AI systems of tomorrow. It is both a rallying call and a hands-on guide for those who want to stay at the forefront, and for cross-disciplinary teams ready to build something better.
1: How to Design Single Cell Data that Trains AI Discovery
Foundational models are rapidly transforming biomedical research, with pharma companies, academic labs, large-scale consortia like the Human Cell Atlas Consortium, and private funders like the Chan Zuckerberg Initiative, developing AI systems trained on large-scale, single-cell datasets. These models learn cross-context patterns—spanning cells, tissues, and diseases—to unearth core principles of biology and fuel new, testable therapeutic hypotheses to accelerate discovery. Yet a key question remains: what is the role of the biologist in this new landscape?
As a wet-lab scientist, one may wonder whether the rise of foundational models diminishes the value of traditional experimental skills. In truth, the opposite is often the case: AI systems need high-quality biological data to learn—and biologists are uniquely positioned to generate it.
What Makes Single-Cell Data Valuable for Training Foundation Models?
Not all single-cell data are equally useful for AI training. Foundational models learn best from datasets that are:
• Rich and diverse – including many cell types, perturbations, tissues, and conditions
• Clean and standardized – with minimal batch effects and high technical consistency
• Well-annotated – including metadata on sample origin, cell identity, and experimental context
• Dynamic – ideally including time series or perturbation data to model cellular transitions
• Open and FAIR – so model developers can legally and reproducibly use the data
By generating data that meets these criteria, biologists don’t just contribute to AI research, they become critical partners in shaping its future. Representative foundational models pre-trained on large-scale single-cell datasets include scFoundation, scGPT, Geneformer, and scBERT. In contrast, cross-modality models such as Cellarium, OpenCell and CZ CELLxGENE integrate multiomic or spatially-resolved single-cell data to support more complex biological queries.
Opportunities for Biologists as ‘Scientific Translators’
Wet-lab scientists can directly support foundational model development by:
• Designing datasets with AI in mind: Creating experiments that are scalable, diverse, and interpretable
• Embedding metadata and biological context: Ensuring the data is not just large, but meaningful
• Partnering with model developers: Acting as scientific translators to improve model training and evaluation
• Highlighting rare or underrepresented biology: From emerging model organisms to difficult tissues
This reframes the biologist’s role: no longer just data generators, we are now architects of biological knowledge, defining what questions models can learn to answer.
Call to Action
Rather than asking whether AI will replace biology, we should ask: How can biology shape the AI that shapes our future?
With thoughtful experimental design, close collaboration, and strategic positioning, biologists can help build foundation models that are not just powerful, but also biologically grounded, inclusive, and relevant. It’s time to lean into that opportunity.
2: How to Build a Foundational Model for Single-Cell Data: A Beginner’s Roadmap
Step 1: Understand What You Want to Build
A foundational model is like a “brain” trained on massive single-cell datasets that learns the patterns of gene expression, cell types, or responses (an overview of the model types summarized in the table below). It can then be adapted to answer all sorts of biological questions.
You don’t need to write code, but you do need to define:
• What question(s) should the model answer?
• What kind of data should it learn from (RNA-seq, ATAC-seq, multiomics)?
• Who are your end users (researchers, clinicians, pharma)?
Table: Types of Foundational Models in Single-Cell Biology
| Model | Description |
| Pretrained | Trained on massive, diverse single-cell datasets |
| Generalizable | Can be fine-tuned or transferred to new tasks |
| Multi-task | Used for annotation, imputation, prediction, harmonization i.e. for multiple types of tasks |
| Cross-species | May support multiple species e.g. both mouse and human |
| Unsupervised / Self-supervised | Learn gene and cell representations from unlabeled data by predicting, reconstructing, or contrasting features within the single-cell profiles |
Step 2: Assemble the Right Team
You’ll need:
• A computational biologist or data scientist
• A bioinformatician with experience in single cell RNA-seq or spatial transcriptomics data
• Typically, a machine learning (ML) engineer (responsible for developing and optimizing ML algorithms and models to apply deep learning techniques to a given dataset)
Your role as the biologist:
• Be the biological driver: provide context, questions, interpretation
• Stay curious and communicate your experimental insights and pain points
Step 3: Gather the Training Data
Use existing public datasets, representative examples include:
• Single Cell Expression Atlas
Alternatively, use your own single cell data produced for this purpose, or use one from a collaborator.
Tip: Ask your collaborator to help you download and standardize the data (they’ll use tools like Scanpy, Seurat, or scVI-tools)
Your job:
• Help select biologically diverse and well-annotated datasets
• Decide which tissue types, organisms, or disease states are important for question at hand
Step 4: Decide the Learning Task
Think of this as deciding what you want the model to learn.
Common goals:
• Cell type prediction across tissues/species
• Imputation: fill in missing gene expression values
• Batch correction & integration: align datasets from different experiments, platforms, or donors to remove technical variation and uncover shared biological signals
• Trajectory inference: map cell differentiation or disease progression
• Multimodal fusion: link RNA + ATAC or RNA + protein data
You can help define this based on your lab’s biology goals.
Step 5: Choose a Base Model (with Help)
Ask your team to explore open-source foundational models like:
scVI: Probabilistic models for single-cell omics data
scGPT (emerging): Transformer-based model learning gene relationships
totalVI: Model of CITE-seq, single cell RNA-seq and protein data
UniCell: Cross-species foundational model
Want to learn more? An extended version of this feature will be available in our upcoming Multi-Omics Playbook. Sign up to our newsletter for updates!
About the author: Arpita Kulkarni, PhD, is Director of the Functional Genomics and Bioinformatics Core at HMS/BIDMC. She is a cell and developmental biologist focused on bridging single cell biology and AI for therapeutic discovery.
