Integrating Multi-Modal Genomics for Target Hypothesis Generation
Single-cell RNA sequencing gives you a snapshot of transcriptional state. That is a rich readout, but it is one layer of biology. Gene expression is the output of regulatory decisions made upstream, encoded in chromatin accessibility, transcription factor binding, and the spatial organization of cells within tissue. A target hypothesis that rests on transcriptional evidence alone is missing context that often matters for predicting whether the target is mechanistically tractable.
Multi-modal single-cell genomics has made it possible to measure several of these layers simultaneously or in matched samples. The analytical challenge is integrating them coherently: extracting the signal from each modality, aligning cells across modalities, and combining the evidence into target hypotheses that are more precisely grounded than any single modality can provide.
What Each Modality Contributes
scRNA-seq measures mRNA abundance. It tells you which genes are being transcribed in each cell and at what level, which is directly relevant to asking whether a gene product is present and potentially active in a disease-relevant cell type. Most current target discovery work starts here.
scATAC-seq measures chromatin accessibility. An open chromatin region indicates that the DNA in that location is accessible to transcription factors. Peaks in accessibility near a gene's regulatory elements indicate active or poised regulation. For target discovery, ATAC data adds the regulatory dimension: not just whether a gene is expressed, but whether the regulatory apparatus is in the active state that could sustain expression under perturbation. A target whose locus is in an accessible chromatin region in the disease cell type is more likely to be actively regulated and thus amenable to modulation than one whose locus is in closed chromatin.
Spatial transcriptomics adds the tissue architecture dimension. Visium, Slide-seq, and related technologies measure gene expression while preserving the spatial location of cells within a tissue section. This matters because cells do not behave identically regardless of their neighbors. A disease-associated cell type that clusters near a particular tissue structure, or that shows distinctive expression only in specific tissue zones, provides spatial context that helps interpret the biology and can suggest whether a target's expression pattern is specific to the pathological tissue compartment.
The Integration Problem
Multi-modal data from the same samples enables paired analysis, where the same cell or the same nearby cells are profiled in multiple modalities. This is the most informative scenario, and technologies like 10x Genomics Multiome (simultaneous RNA and ATAC from the same nucleus) make paired data increasingly accessible.
More commonly, modalities come from separate studies. The same tissue type in the same disease has been profiled with RNA and ATAC by different groups, but not on the same cells. Integration in this case requires computational alignment: finding the correspondence between cell populations defined in one modality and cell populations defined in another, without direct pairing information.
Methods like Seurat v3 weighted nearest neighbor, MOFA+, and similar approaches perform this alignment by finding a shared low-dimensional space that captures variation common to both modalities. The alignment quality depends on how similar the datasets are in terms of tissue and processing. When the two datasets come from similar conditions and the cell types are broadly shared, alignment is typically reliable for major populations. For rare subtypes, the alignment is noisier because there are few cells to anchor the correspondence.
Using ATAC Data to Prioritize Target Hypotheses
One of the more practical applications of multi-modal integration in target discovery is using ATAC data to filter or rank gene candidates that emerge from RNA analysis. The reasoning is mechanistic: if a gene is differentially expressed in the disease cell population and its regulatory elements are also differentially accessible in the same population, the differential expression is more likely to reflect active regulatory change than stochastic dropout or measurement noise.
Concordance between RNA and ATAC evidence raises confidence in a candidate. Discordance, where a gene shows differential expression but no corresponding differential accessibility, does not necessarily mean the candidate is invalid, but it raises questions. The expression difference could reflect post-transcriptional regulation, differences in mRNA stability, or protein-level effects that the RNA snapshot does not directly measure.
We treat ATAC-RNA concordance as one component of a multi-evidence score rather than as a hard filter. A candidate that lacks ATAC support is not eliminated, but it carries a different evidence profile than one with concordant multi-modal support, and that distinction should be visible to the discovery team evaluating the hypothesis.
Spatial Context for Tissue-Level Target Questions
Spatial transcriptomics data is most informative for targets whose therapeutic relevance depends on tissue location. In a fibrotic tissue, for example, the cells most relevant to disease progression may be those at the interface between fibrotic and normal tissue, not cells evenly distributed throughout. Spatial data can identify whether a disease-associated cell population is concentrated in specific tissue zones, which gives additional context for thinking about delivery and local target engagement.
The current limitation of spatial data for target discovery is resolution. Technologies like Visium capture expression at a scale of roughly 50-100 micrometers per spot, which means each spot contains multiple cells. True single-cell resolution spatial data exists but is technically demanding and not yet widespread in the public literature. As resolution improves and more spatial single-cell datasets enter public repositories, the spatial layer will become more practically useful for target hypothesis generation.
At this stage, spatial data is most useful for confirming that a cell population identified from dissociated scRNA-seq samples genuinely exists in intact tissue and is located where the biology predicts it should be. The risk in dissociated single-cell analysis is that the dissociation process itself changes cell states, and cells that look disease-associated in dissociated data might reflect stress responses rather than genuine disease biology. Spatial data from matched samples provides an independent check on whether the cell population is real in its tissue context.
The Cumulative Evidence Standard
A target hypothesis built on multi-modal evidence is not necessarily more correct than one built on transcriptomics alone. It is more precisely characterized. The multi-modal layers answer different biological questions: RNA tells you what is being made, ATAC tells you whether the regulatory state is active, spatial tells you where in tissue the relevant biology is concentrated. These are additive sources of specificity.
This is also the limitation of multi-modal analysis as a filter: it is only as good as the availability of matched or alignable data. For indications where multi-modal public data is sparse, the analysis necessarily reduces to the modalities that are available. Pushing a discovery team toward a hypothesis that is well-supported by multi-modal evidence in a different indication when the relevant indication lacks multi-modal data is a category error.
The honest approach is to be transparent about which evidence types are present and which are absent for any given target hypothesis. A hypothesis with concordant RNA, ATAC, and spatial support carries more weight than one with RNA support only, but a RNA-only hypothesis in the right cell type with clear disease specificity is still a meaningful starting point for experimental follow-up. Multi-modal integration strengthens and contextualizes the biological rationale; it does not replace the experimental validation work that comes next.
