From Bulk Transcriptomics to Single-Cell: What Changes for Target ID

The question we hear most often from discovery teams making the move from bulk to single-cell workflows is: what do I need to change in how I think about results? The answer is more than most people expect. It is not simply a matter of using a different tool to get higher-resolution numbers on the same question. The questions themselves change shape, which means the hypotheses you come out with are structured differently, and the way you design follow-on experiments needs to adjust accordingly.

This article is intended for biologists and computational scientists who are already familiar with bulk transcriptomics and want a clear account of what the transition entails for target identification specifically.

What Bulk Data Gives You and What It Does Not

Bulk RNA-seq gives you the aggregate transcriptional output of a tissue sample. That output is a weighted average of every cell in the sample, where the weighting is the relative abundance of each cell type. It is a useful summary in many contexts. For differential expression between disease and healthy states, bulk data at large sample sizes is statistically powerful and gives you a robust picture of which genes change across conditions.

What it does not tell you is which cells are responsible for the change. A gene that is strongly upregulated in disease tissue in a bulk experiment could be: (a) moderately upregulated across many cell types, (b) strongly upregulated in one specific cell type that is relatively abundant, or (c) extremely highly upregulated in a rare cell subpopulation whose signal is diluted by the rest of the tissue. These three scenarios require entirely different drug discovery strategies. Bulk data cannot distinguish between them.

This is not a limitation of statistical power. It is a limitation of the measurement: the individual-cell resolution is destroyed at the point of lysis when preparing bulk samples. No amount of deeper sequencing of a bulk sample recovers per-cell information. You need to change the experimental design upstream.

The Unit of Analysis Shifts

In bulk transcriptomics, the unit of analysis is the sample. You have N samples from disease patients and M samples from healthy controls, and your analysis asks: which genes differ between these groups?

In single-cell transcriptomics, the unit of analysis is the cell. You have potentially hundreds of thousands of individual cell profiles, each labelled with a cell type (through annotation) and a condition (disease vs. healthy, treated vs. untreated). Your analysis can ask: which genes differ between cell types? Which cell types differ between disease and healthy conditions? Which genes within a specific cell type differ between conditions? These are finer-grained and more specific questions, and they open up hypothesis spaces that are genuinely inaccessible to bulk data.

The practical consequence is that your statistical model needs to account for the multi-level structure of the data: cells nested within donors, donors varying between individuals, cell types having baseline differences in gene expression. Ignoring this structure produces false positives and misattributed effects. The analytical approaches used in bulk (basic linear models, DESeq2 on pseudobulk) need to be used carefully and with awareness of the appropriate aggregation level for each question.

Cell-Type Annotation Is Now Part of Your Pipeline

Bulk data has no cell-type annotation step. You run the sample, get counts, normalise, and do differential expression. The cell-type composition of the sample is invisible to the analysis.

Single-cell data requires you to identify what each cell is before you can interpret what is happening. This annotation step is a significant addition to the analytical pipeline. For well-characterised tissues with published reference datasets (brain, liver, blood), annotation transfer from existing references is fairly reliable. For disease-altered tissues, novel cell states that do not map cleanly to reference types are common, and deciding how to handle them requires biological judgment.

Annotation errors propagate downstream. A cell type that is misidentified will produce spurious differential expression signals attributed to the wrong population, which will produce hypotheses that do not replicate. Getting this step right matters disproportionately to everything that follows.

Foundation model representations help here: when you project cells into a well-trained embedding space, cells with similar biology cluster together even without manual annotation, and their relationship to known reference types is captured in the embedding geometry. But the annotation step still requires biological review, especially for novel or transitional states that appear specifically in disease contexts.

The Target Hypothesis Changes Shape

A bulk-derived target hypothesis looks like: "GENE X is significantly upregulated (log2 FC 1.8, adjusted p < 0.001) in disease tissue versus healthy controls." It is a tissue-level statement. Its actionability depends entirely on follow-on work to determine the cellular context.

A single-cell-derived target hypothesis looks like: "GENE X is upregulated in disease-activated fibroblasts (log2 FC 3.2 in this cell type), with expression concentrated in this subpopulation and minimal baseline expression in neighbouring macrophage or T cell populations in the same tissue." This is a cell-type-addressed statement. You know where in the tissue to expect target expression, what populations you need to spare, and what your selectivity challenge is likely to be.

The downstream experimental plan is different in each case. The bulk-derived hypothesis sends you to cell lines to determine context. The single-cell-derived hypothesis tells you what cell line model to use, or tells you that primary cells are needed because no cell line faithfully represents the subpopulation of interest.

What Stays the Same

We should be clear about what the transition does not change. Single-cell data does not validate a target. It gives you a more specific and contextualised hypothesis. That hypothesis still needs functional evidence: what happens when you perturb this gene in the relevant cell type? Does it affect the pathological phenotype? Does it do so with acceptable effect on other cell types?

The tools for that functional work (CRISPR screens, cell line perturbation, primary cell culture, in vivo models) are not replaced by better transcriptomics. What changes is the quality of the input to those experiments: you are testing a hypothesis grounded in human biology at cell-type resolution, rather than a tissue-level correlation that may or may not be driven by the cells you can model in vitro.

It is also worth noting that bulk data at large sample sizes retains value for certain tasks. eQTL analysis, rare variant association, and large population cohort studies still benefit from the scale that bulk data affords. The choice is not either/or. For early-stage target identification in heterogeneous tissues, single-cell data is becoming the appropriate first choice for the reasons described above. For tasks requiring N in the hundreds or thousands and where cellular resolution is not the limiting question, bulk remains practical and powerful.

Practical Considerations for the Transition

Moving a discovery program to single-cell inputs involves decisions beyond the analytical framework. Sample acquisition is different: single-cell protocols typically require fresh tissue (or carefully managed frozen samples with appropriate fixation or hashing protocols) rather than snap-frozen bulk samples. For some tissue types, fresh collection constraints affect feasibility.

Data volumes are substantially larger: a 50,000-cell experiment generates significantly more raw data than a typical bulk run, and the downstream analysis requires more computational resources. Count matrices from single-cell experiments are also noisier per cell than bulk measurements, because each cell is sequenced at a fraction of the depth of a bulk sample. This technical noise requires appropriate handling in the statistical model, and it is distinct from biological signal.

The analytical skills required overlap partially with bulk workflows but diverge significantly at the cell-type annotation, clustering, and multi-level modelling steps. Teams making this transition typically need either to upskill computationally or to access specialist analysis capacity. The return on that investment, in terms of the quality and specificity of target hypotheses that come out the other side, is the core case for making it.