Part V: Cellular Context
Central question: How do foundation model principles extend beyond one-dimensional sequence to embrace the full complexity of cellular and systems-level biology?
Prerequisites: Part IV (foundation model families). Parts II-III (7 Transformers and Attention, 8 Pretraining Strategies) for architectural background.
| Chapter | Topic | Key Concepts |
|---|---|---|
| 19 RNA Structure and Function | RNA Models | Secondary structure, splicing, RNA foundation models |
| 20 Single-Cell Models | Single-Cell Models | Sparsity, cell type annotation, perturbation prediction |
| 21 3D Genome Organization | 3D Genome | Chromatin loops, TADs, Hi-C prediction, spatial transcriptomics |
| 22 Graph and Network Models | Network Models | GNNs, PPI networks, regulatory networks, pathway integration |
| 23 Multi-Omics Integration | Multi-Omics Integration | Cross-modal learning, genotype-to-phenotype paths |
After completing Part V, you will understand:
- How RNA structure adds a second dimension to sequence modeling
- Why single-cell data requires different architectural adaptations
- How 3D genome organization connects distal regulatory elements
- When graph neural networks complement sequence models
- How to integrate multiple data modalities toward phenotype prediction
Biology operates across scales that sequence alone cannot capture. Cells of different types read the same genome differently, activating distinct gene programs that produce neurons, hepatocytes, and immune cells from identical DNA. Genes function not in isolation but within networks of regulation and interaction, where perturbing one node propagates effects throughout the system. The three-dimensional folding of chromatin brings distal elements into contact, creating regulatory logic invisible to models that treat genomes as one-dimensional strings. Sequence foundation models ask what a genome encodes; the models in this part ask what that sequence becomes in particular cellular contexts, interaction networks, and spatial architectures.
This transition from sequence-centric to systems-scale modeling demands new data modalities and new computational approaches. Single-cell transcriptomics reveals the cellular heterogeneity that bulk measurements average over. Hi-C and related methods expose the spatial organization that determines which enhancers contact which promoters. Protein interaction networks and gene regulatory graphs capture relational structure that no amount of sequence analysis can infer. Foundation model principles extend naturally to these modalities: learn representations from large-scale data, then transfer to specific prediction tasks.
RNA structure and function (19 RNA Structure and Function) extend sequence modeling beyond DNA, covering secondary structure prediction, splicing regulation, and the emerging frontier of RNA foundation models. Single-cell transcriptomics and epigenomics (20 Single-Cell Models) present distinct challenges of sparsity, noise, and scale that transformer architectures must adapt to capture. Three-dimensional genome organization (21 3D Genome Organization) adds spatial context, with models predicting chromatin contacts from sequence and connecting spatial structure to gene regulation. Graph neural networks (22 Graph and Network Models) represent biological entities and their interactions as structured graphs, integrating foundation model embeddings with relational reasoning. Multi-omics integration (23 Multi-Omics Integration) broadens the view further, jointly representing genomic, transcriptomic, proteomic, and clinical information to trace the path from genotype to phenotype.
- Part IV provides the sequence-level foundation models that feed into systems-scale approaches
- Parts II-III (7 Transformers and Attention) introduce the attention mechanisms adapted for these new modalities
- Part VI evaluation principles apply to multi-modal models with additional complexity
- Part VII clinical applications increasingly depend on systems-level integration