DE-Rigor-Agent screens differential-expression methods before running them, keeping assumption mismatches from inflating FDR
Synopsis
In bulk RNA-seq differential expression, general-purpose LLM agents tend to pick DESeq2 or edgeR without checking whether those methods' statistical assumptions fit the data at hand. DE-Rigor-Agent splits the analysis into input standardization, method filtering and method execution, and before running anything it evaluates six common methods against expert-curated method cards and dataset-specific diagnostics, recording why each is kept or excluded. Across 140 simulated datasets its filtering decisions largely matched the incompatibilities built into the simulations and stayed stable across random seeds, while two Codex baselines chose DESeq2 or edgeR for nearly all datasets.
Interpretation
Differential expression analysis gains a gate before execution: the agent first judges whether each of six common methods suits the current dataset, then decides which to run. General coding agents can plan an analysis, write code and inspect results, but successful code execution does not mean the statistical method was chosen well; the prior practice was to apply popular methods such as DESeq2 and edgeR without checking their distributional assumptions and input requirements. On 140 simulated datasets, two Codex baselines, one of them explicitly prompted to consider method assumptions, selected DESeq2 or edgeR for nearly all datasets, whereas DE-Rigor-Agent's exclusions largely matched the incompatibilities expected from the simulation settings.
The design fixes the analysis into three sequential stages, input standardization, method filtering and method execution, and every method must pass filtering before it runs. Filtering rests on expert knowledge written in natural language: method cards record each method's distributional assumptions, input requirements and supported study designs, and filtering principles add guidance for data characteristics such as sample size. The agent reads the expression matrix and sample metadata, generates and runs diagnostic code, and issues a keep-or-exclude decision with supporting evidence for each of the six candidate methods; retained methods are then handed to dedicated subagents that adapt expert-curated code templates.
Filtering decisions were largely stable across repeated runs, at the cost of a modest rise in per-dataset analysis cost. For 77 of 84 scenario-method combinations, at least 8 of 10 replicates produced the same decision, and 50 combinations were identical across all 10 replicates; inconsistencies clustered under mild deviations from the negative binomial assumption, where diagnostic evidence was less pronounced. Estimated from token usage at OpenAI's published prices, mean cost per dataset was $0.018 for DE-Rigor-Agent, $0.004 for vanilla Codex and $0.005 for guided Codex.
Perspective
The evaluation covers 140 simulated datasets across 14 scenarios involving small sample sizes, zero inflation, expression outliers, pre-normalized inputs and confounded batch effects, with six common differential expression methods as candidates; the authors state that further evaluation across additional simulation scenarios and real-world RNA-seq datasets would help establish the framework's generalizability.
Filtering depends on whether data diagnostics can detect assumption violations, and mild deviations are often hard to identify, so decisions vary across replicates; passing the filter does not guarantee FDR control or adequate power, as in the simulations NOISeq was retained at 3 samples per group with empirical FDR 0.10, and DESeq2 was retained in 9 of 10 runs at both 3 and 6 samples per group with empirical FDR 0.09 and 0.07.