Assumption checks move to the center of bulk RNA-seq differential expression
Bulk RNA-seq differential expression has always carried a hidden hazard: the method can be statistically elegant and still be wrong for the data in front of it. DE-Rigor-Agent was built around that problem. The new preprint presents an agentic workflow that does not simply pick a familiar package and run with it; it first standardizes inputs, then filters methods against dataset-specific diagnostics, and only then executes the methods that survive those checks.
The practical aim is straightforward. Popular differential expression tools are often treated as interchangeable, even though their assumptions differ sharply across sample size, count distribution, normalization state, and confounding structure. In a setting with sparse counts, zero inflation, expression outliers, pre-normalized matrices, or batch effects tangled with biology, a method that looks routine can inflate the false discovery rate or erase power altogether. DE-Rigor-Agent is designed to keep those mismatches from being silently baked into the analysis.
Method filtering before method execution
The workflow organizes analysis into three stages. First comes input standardization, which aligns the data into a form the downstream methods can interpret consistently. Second is method filtering, where expert-curated criteria and diagnostics are used to judge whether each of six commonly used differential expression methods is appropriate for the specific dataset. Third is execution, in which every retained method is run after the exclusions are documented.
That sequence matters because the agent is not just choosing among algorithms; it is explicitly reasoning about whether the algorithm is compatible with the data. In ordinary automation, a large language model may default to the most familiar tools in the field, even when the assumptions behind those tools are visibly strained. DE-Rigor-Agent was built to interrupt that reflex and replace it with an auditable decision chain.
One sentence captures the core advance: the agent decides what not to run before it decides what to run.
Simulated datasets exposed where method choice breaks down
Evaluation used 140 simulated datasets spanning 14 challenging scenarios. Those scenarios included small sample sizes, zero inflation, expression outliers, pre-normalized inputs, and confounded batch effects. The point of that design was not to stage a friendly benchmark. It was to create conditions in which method-dataset incompatibility would become visible, measurable, and hard to ignore.
Across those simulations, incompatible pairings could drive either inflated false discovery rates or loss of statistical power. In other words, the penalty for a poor method choice was not theoretical. It showed up in the two metrics analysts care about most when searching for true signal: whether the discoveries are real, and whether the workflow can find them at all.
A second blunt finding stood out: two general-purpose LLM agent baselines, including one explicitly prompted to consider statistical assumptions, selected DESeq2 or edgeR for nearly every dataset. That uniformity looks efficient, but it is precisely the kind of overgeneralization that can become dangerous when the data violate a package’s assumptions.
Filtering decisions tracked the simulated incompatibilities
DE-Rigor-Agent behaved differently. Its filtering choices closely matched the incompatibilities introduced in the simulations, and those decisions remained consistent across random seeds. Consistency matters in an automated scientific workflow because a tool that changes its mind from one run to the next is hard to trust, even when the final outputs appear polished.
The agent also did this without major computational penalty. The preprint reports little additional computational cost, which is important because assumption-aware analysis loses much of its value if it becomes too slow or cumbersome for routine use. Here, the workflow appears to preserve the speed benefits of automation while adding a layer of statistical discipline that standard agentic pipelines often lack.
The result is not a new differential expression engine. It is a control layer that decides whether the engine belongs on the road in the first place.
Why assumption awareness matters for automated genomics
The biological stakes of differential expression analysis are easy to underestimate because the output often arrives as a list of genes. Behind that list sits a chain of statistical decisions that can amplify noise into apparent discovery. When count distributions are irregular, sample sizes are small, or batch structure mimics biology, a default method can produce a convincing but unstable gene set. An assumption-aware agent confronts those failure modes directly by linking each analytic choice to the observed structure of the dataset.
That approach also changes the transparency of the workflow. Instead of delivering a black-box answer, the agent documents why each method is retained or excluded. For investigators, that audit trail can be as important as the final ranking of genes, because it turns a hidden statistical judgment into an inspectable step.
The preprint’s central claim is therefore narrower than a full automation fantasy and more credible because of it: DE-Rigor-Agent can assess method suitability in bulk RNA-seq differential expression more rigorously than baseline LLM agents that gravitate toward default tools.
Limits of a simulation-first demonstration
The evidence presented here comes from simulated datasets, not from a broad sweep of real-world studies with all the messiness that typically follows. Simulation is useful for stress-testing assumptions because it isolates failure modes one at a time, but it cannot capture every quirk in experimental design, sample preparation, or phenotype annotation. The preprint also focuses on six commonly used methods, so its reach is bounded by the methods built into the workflow.
Even so, the study addresses a real weakness in current automated analysis. LLM-based agents can be persuasive in style and still shallow in statistical judgment. DE-Rigor-Agent tries to close that gap by making method suitability a first-class decision rather than an afterthought.
For bulk RNA-seq analysis, the most consequential choice may be the one made before any differential expression table exists: whether the selected method is actually fit for the data that produced it.