From Differentially Expressed Genes to Disease Drivers

A transcriptomic experiment returns a list of thousands of significantly changed genes. Statistical thresholds are passed, fold-change cutoffs are met, and the adjusted p-values look clean. But that list, in itself, tells you almost nothing about which of those genes is causing the disease, which is responding to it, and which is simply a statistical artefact of how the data was processed. Moving from a differential expression result to a genuine understanding of disease mechanism requires a set of analytical and interpretive steps that most transcriptomic workflows do not include.

This is the central challenge of disease gene prioritisation: not identifying genes that are differentially expressed, but identifying the subset of those genes whose altered activity is mechanistically responsible for the disease phenotype. The distinction between a differentially expressed gene and a disease driver is not a minor technical caveat. It is the difference between a list and a hypothesis.

What Differential Expression Actually Tells You

Differential gene expression (DGE) analysis compares mRNA abundance between two conditions, typically disease versus healthy tissue, and identifies genes whose expression levels differ significantly. The tools used most commonly, DESeq2 and edgeR, use negative binomial models to account for the count nature of RNA-seq data and apply multiple testing correction to control false discovery rates.

The output, a ranked list of differentially expressed genes (DEGs), is the standard starting point for transcriptomic disease research. This analysis helps identify DEGs that may be linked to the investigated disease as genetic triggers, consequences, or indicators. That framing captures the problem precisely: the same list contains all three categories simultaneously, and the statistical method cannot distinguish between them.

A gene that is upregulated in disease tissue could be a driver whose increased activity promotes disease progression, a passenger responding to cellular stress without contributing to pathology, a compensatory response the cell is mounting against the disease, a secondary consequence of another gene's dysregulation, or a statistical false positive. The expression change is real; what it means biologically requires additional evidence.

The statistical problem is also more serious than is commonly appreciated. A 2022 paper in Genome Biology showed, through permutation analysis, that DESeq2 and edgeR produce exaggerated false positive rates when applied to human population-level RNA-seq datasets. Spurious DEGs identified in permuted datasets were systematically enriched in immune-related gene ontology terms, meaning false positives from these commonly used tools are not randomly distributed. They cluster in biologically plausible but incorrect categories. A researcher interpreting immune pathway enrichment from a transcriptomic dataset cannot assume the enrichment reflects real biology without additional validation.

The Driver-Passenger Problem

In cancer biology, the distinction between driver and passenger alterations is well-established. Driver mutations are those that confer a selective growth advantage and contribute causally to tumour development; passenger mutations are those that accumulate in the genome without contributing to pathology. The Pan-Cancer Analysis of Whole Genomes (PCAWG) project, one of the most comprehensive cancer genomics studies conducted, analysed the whole genomes of 2,658 people with 38 cancer types and identified mutations in 179 genes as drivers. Even in that dataset, driver mutations did not explain all cases, and some passenger variants were found to actively hinder cancer development rather than simply being neutral.

The same conceptual distinction applies to differential expression data in any disease context, not just cancer. Most differentially expressed genes in a disease state are passengers: they change because something upstream changed, not because they are causally responsible for disease. Identifying the drivers requires moving beyond the expression change itself.

The causal gene may not even be differentially expressed in the conventional sense. Causal mutations always cast a long transcriptional shadow over the rest of the data, and that shadow can be quantified even when the causal gene itself does not show significant expression changes. Network-based approaches that examine how dysregulation propagates through gene interaction graphs can identify upstream regulators that are themselves modestly or not differentially expressed but whose altered activity explains the expression changes observed across many downstream genes.

From DEGs to Mechanisms: The Analytical Gap

The standard analytical workflow after differential expression analysis, running pathway enrichment against Reactome, GO, or KEGG, narrows the gene list to annotated biological processes but does not resolve the driver-passenger question. A pathway appearing significantly enriched means that genes annotated to that pathway are over-represented in the DEG list. It does not mean the pathway is causally dysregulated, that the enrichment reflects disease biology rather than reactive gene expression, or that any of the enriched genes are suitable therapeutic targets.

Emerging analytical strategies, including network module analysis, knowledge graph integration, causal inference frameworks, and drug repurposing analyses, have been proposed to bridge the gap between DEG lists and mechanistic understanding. These approaches share a common logic: they use additional sources of biological knowledge, structural network information, genetic association data, and directional cause-and-effect relationships curated from the literature, to constrain interpretation of the expression data.

Network-based approaches are among the most practically useful. By mapping DEGs onto protein-protein interaction (PPI) networks, researchers can identify clusters of DEGs that are close to the disease-causal gene and determine pathways that could play central roles in disease pathogenesis. Proximity to known causal genes in the network is a stronger signal than expression magnitude alone, because it reflects the known topology of biological regulation rather than treating all genes as independent.

Causal Inference as a Framework

The most principled approach to moving from DEGs to disease drivers is causal inference: using statistical and computational methods to estimate which genes causally regulate the expression changes observed in a disease dataset, rather than simply identifying which genes are associated with disease.

One well-developed framework is Tumor-specific Causal Inference (TCI), which uses Bayesian causal learning to infer relationships between somatic genome alterations and differentially expressed genes within individual tumours. Applied to data from 5,097 tumours across 16 cancer types in TCGA, TCI derived tumour-specific causal models that distinguished candidate driver alterations from passengers by testing whether genomic changes causally explain the observed transcriptomic shifts in each tumour.

Causal reasoning approaches can also be applied to non-cancer disease contexts. In inflammatory bowel disease, combining quantitative pathway analysis with causal inference on transcriptome data from Crohn's disease and ulcerative colitis patients identified IFNG and GBP5 as subtype-specific regulators driving different pathogenesis, with directionality validated in an independent cohort. This type of analysis moves from the observation that immune genes are differentially expressed in IBD, a finding that has been replicated many times, to a specific mechanistic claim about which upstream regulators are causally responsible for the subtype differences.

Reverse causal reasoning extends this logic by using curated knowledge bases of biological cause-and-effect relationships, expressed in computable formats such as Biological Expression Language, to generate and evaluate mechanistic hypotheses from expression data. DEGs are mapped to causal networks and scored for their consistency with known regulatory relationships, yielding prioritised upstream hypotheses rather than gene lists. This approach provides an evidence-driven route to developing models of disease, drug action, and drug toxicity that are grounded in existing mechanistic knowledge rather than derived solely from statistical associations.

The Role of Genetic Evidence

The most robust way to identify a disease driver from transcriptomic data is to combine expression evidence with genetic evidence. A gene whose expression is altered in disease tissue and whose locus is associated with disease risk through GWAS, whose variant affects gene expression as a cis-eQTL in the relevant tissue, and which colocalises statistically with the disease GWAS signal, represents a convergence of evidence that statistical transcriptomics alone cannot provide.

This integration is the basis for post-GWAS gene prioritisation pipelines and for Mendelian randomization approaches that use genetic variants as instruments to estimate causal effects of gene expression on disease outcomes. The combination of transcriptomic and genetic evidence substantially reduces the probability that a nominally significant DEG is simply a passenger or a reactive change, because the genetic evidence is not subject to the same confounding and reverse causation that affect observational expression studies.

What Counts as a Disease Driver

A disease driver, as opposed to a differentially expressed gene or a disease-associated gene, should meet several criteria that expression data alone cannot satisfy. It should show altered activity in disease-relevant cell types at disease-relevant stages. Its alteration should be upstream of other observed changes, explainable as a causal regulator of the expression programme rather than a consequence of it. There should be functional evidence, from genetic perturbation, protein interaction studies, or experimental manipulation, that modulating the gene's activity affects the disease-relevant phenotype. And ideally, the causal claim should be supported by independent lines of evidence including genetic association, network topology, and directional regulatory data.

Meeting all of these criteria transforms a DEG from a candidate into a justified hypothesis that can be tested and, if it holds, acted upon therapeutically. The gap between a p-value in a differential expression analysis and that level of evidence is substantial, and the tools to close it, causal inference frameworks, network proximity analysis, genetic integration, and reverse causal reasoning, are now available and increasingly well-validated.

The starting point of transcriptomic disease research is the DEG list. The destination is a mechanistic account of which molecular events drive the disease and why. The distance between those two points is where most of the real work of computational disease biology sits.

---

References

1. Li Y et al. Exaggerated false positives by popular differential expression methods when analyzing human population samples. Genome Biology. 2022;23:79. https://doi.org/10.1186/s13059-022-02648-4

2. Zhang J, Zhang S. The discovery of mutated driver pathways in cancer: models and algorithms. IEEE/ACM Transactions on Computational Biology and Bioinformatics. 2016. https://arxiv.org/pdf/1604.01298

3. Gu J et al. Systematic discovery of the functional impact of somatic genome alterations in individual tumors through tumor-specific causal inference. PLOS Computational Biology. 2019;15(7):e1007088. https://pmc.ncbi.nlm.nih.gov/articles/PMC6650088/

4. Ma S et al. Tumour-specific causal inference discovers distinct disease mechanisms underlying cancer subtypes. Nature Communications. 2019. https://pmc.ncbi.nlm.nih.gov/articles/PMC6744493/

5. Xia J et al. IBD subtype-regulators IFNG and GBP5 identified by causal inference drive more intense innate immunity and inflammatory responses in CD than those in UC. Frontiers in Cell and Developmental Biology. 2022. https://pmc.ncbi.nlm.nih.gov/articles/PMC9020454/

6. Catlett NL et al. Reverse causal reasoning: applying qualitative causal knowledge to the interpretation of high-throughput data. BMC Bioinformatics. 2013;14:340. https://pmc.ncbi.nlm.nih.gov/articles/PMC4222496/

7. Flores M et al. Integrative transcriptomic and network-based analysis of neuromuscular diseases. Biomedicines. 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC12525345/

8. Pavlidis P, Noble WS. Beyond differential expression: the quest for causal mutations and effector molecules. BMC Genomics. 2012;13:356. https://pmc.ncbi.nlm.nih.gov/articles/PMC3444927/

9. Shen H et al. Differentially expressed genes: advances in analytical strategies, biological interpretation and clinical applications. Journal of Advanced Research. 2024. https://pmc.ncbi.nlm.nih.gov/articles/pmid/39647635

10. ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium. Pan-cancer analysis of whole genomes. Nature. 2020;578:82-93. Referenced via: https://oncologynews.com.au/?p=25410