In RNA-Seq and microarray analysis, GO analysis and pathway analysis are often performed after extracting differentially expressed genes. For many years, these methods have been used as representative tools for biological interpretation at the end of omics data analysis.
However, this way of using GO and pathway analysis may rapidly become outdated.
This does not mean that GO analysis or pathway analysis will disappear. In papers and formal reports, tables of GO terms and pathways will likely continue to be used. However, the process of reading those tables one by one and building a biological story from them is likely to be replaced by AI-based summarization and comparison with the literature.
AI Will Start Reading GO and Pathway Analysis Results
Until now, researchers have reviewed tables of GO and pathway analysis results, selected terms that seemed important, and interpreted them by comparing them with papers and existing biological knowledge.
With AI, however, long lists of term names, gene lists, expression patterns, sample information, disease names, cell types, experimental conditions, and existing literature can be combined and summarized in a more readable form.
Reexamining Genes with Small but Shared Expression Changes
For example, consider the case shown in the video below. The genes analyzed here were not a set of strongly differentially expressed genes identified by conventional differential expression analysis.
These genes were examined in a comparison of edgeR, DESeq2, and the t-test in a paired design . They were identified as significant by the paired t-test, but not by edgeR or DESeq2. Although the magnitude of their expression changes was relatively small, examining their expression patterns across samples revealed several groups of genes that shared similar patterns of variation.
Such genes may be excluded if decisions are based solely on the magnitude of expression changes or on the results of a particular analysis method. However, if their shared expression patterns reflect biologically meaningful changes, they may contain information that should not simply be discarded. We therefore divided the genes into five clusters based on their expression patterns and entered the gene list for each cluster into ChatGPT to examine the functions shared by the genes in each cluster and the differences among the clusters.
In the video above, only cluster-specific gene lists are provided to AI. At present, however, it is often more useful to provide AI with the GO and pathway analysis results for each cluster as well, and to proceed with interpretation through an interactive dialogue. This makes it easier to organize the results in a way that reflects the biological context. In the future, AI may perform the necessary enrichment analysis directly from gene lists and support interpretation based on those results.
As this type of workflow becomes more common, the practice of manually reading GO and pathway analysis tables one by one will gradually become less central in day-to-day analysis.
Even So, GO and Pathway Analysis Will Remain
On the other hand, dialogue with AI has a reproducibility problem. Even if the same gene list is given to AI, the explanation may change depending on the AI model, the prompt, the information being referenced, and the way the question is asked. Even when using the same AI system, the exact same answer is not always returned.
For this reason, in papers and official reports, it is still necessary to record, in a reproducible form, which database was used, which statistical method was applied, and which GO terms or pathways were significant. Tables from GO analysis and pathway analysis will likely remain necessary as reproducible supporting evidence in papers and reports.
A Split May Emerge: Tables in Papers, AI in Practice
If we only read papers in the future, GO analysis and pathway analysis may appear to be used in the same way as before. However, in actual analysis practice, AI will likely be used more and more for deciding how to read those tables, which terms should be emphasized, and what kind of biological story can be built from the results.
In other words, for those learning RNA-Seq data analysis from now on, it will be important not only to learn how to read GO and pathway analysis tables, but also how to ask questions to AI and how to handle the answers it provides. AI-generated answers can contain errors. Even when an explanation sounds plausible, the supporting evidence may be weak when checked against the original data or the literature. Therefore, what matters is not to believe AI uncritically. What matters is to become familiar with using AI, and to check its answers by returning to the original data and the literature.
AI-Generated Interpretations Should Be Checked Against Both the Gene List and the Original Data
AI can summarize the results of GO analysis and pathway analysis and explain the biological meaning of a gene list. However, even when an explanation appears natural and convincing, it may not actually be supported by the genes in the input list. AI may also produce hallucinations, such as citing genes that are not included in the input list or presenting poorly supported interpretations in a convincing manner. In addition, AI does not automatically guarantee the validity of the input gene list itself or the quality of the original data from which that gene set was derived.
When using AI to interpret a gene list, at least two distinct levels of verification should therefore be considered. The first is to determine whether the explanation generated by the AI is supported by the input gene list and the results of enrichment analysis, and whether it contains any hallucinations. The second is to determine whether the gene list itself was appropriately extracted from the original expression data.
Do the extracted genes truly reflect differences between the samples? Does the list include variation from low-expression genes? Could the results have been influenced by outlier samples or batch effects? Do the results change substantially depending on the normalization or filtering conditions? When multiple statistical methods are used, are the same genes and expression patterns still supported?
If these points are not examined, asking AI to summarize the results may simply produce a plausible explanation tailored to the submitted gene list. At the same time, even when the input genes are relatively strongly supported statistically, the biological interpretation initially generated by the AI may still contain hallucinations or inappropriate generalizations.
What matters in omics data analysis in the AI era is not simply accepting the answer produced by AI. It is the ability to check an AI-generated interpretation against the input gene list, the results of enrichment analysis, the original expression data, PCA, clustering, heatmaps, sample information, and the expression pattern of each gene.
Conclusion
The era of treating GO and pathway analysis as the final goal is coming to an end. In the future, omics interpretation will increasingly be carried out together with AI, while checking the results against the original data.
Related Topics
- How to Learn RNA-Seq Data Analysis in the AI Era | From Using Tools to Judgment and Verification
- Is Network Analysis Necessary in RNA-Seq Analysis? | How an Unvalidated Network Diagram Can Become Scientific Knowledge
GeneAgent: An Attempt to Incorporate Self-Verification into Gene List Interpretation
When using AI to interpret a gene list, it is important to consider how the explanation generated by the AI can be verified. Even when an explanation appears natural and convincing, it may not actually be supported by the genes in the input list.
One attempt to address this problem is GeneAgent . GeneAgent is a research-oriented AI agent that first explains the representative functions of a gene set and then uses enrichment analysis, gene databases, and other resources to verify its own interpretation. By incorporating verification against external sources into the analysis process, it aims to reduce hallucinations generated by the AI itself. Rather than simply returning an AI-generated explanation, it is an interesting attempt to indicate the future direction of gene list analysis by verifying the explanation and revising it when necessary.
We therefore tested GeneAgent using Cluster 3 and Cluster 5 from the five clusters shown in the video above, because these two clusters fell within GeneAgent’s input limit of 400 genes.
For Cluster 5, the gene set was initially described as being related to cellular stress responses. After self-verification, however, the main interpretation was revised to one associated with translation and the formation of free 40S ribosomal subunits. This result demonstrates the potential value of using external databases and enrichment analysis to reevaluate an initially generated explanation.
At the same time, we also observed cases in which parts of the original explanation remained in the final text or an interpretation was generated without fully confirming its correspondence with the input gene list. Therefore, at present, even results generated by an AI with a self-verification function still need to be reviewed by a human.
For Cluster 3, GeneAgent initially described the gene set as being related to signal transduction, immune responses, apoptosis, the cell cycle, metabolism, DNA repair, and the cytoskeleton. However, some of the genes cited as evidence for this interpretation were not included in the Cluster 3 gene list that had been submitted.
The Cluster 3 and Cluster 5 analyzed above were derived from genes identified as significant by the t-test alone. We therefore next examined how GeneAgent’s self-verification would work when applied to DEGs supported by stronger statistical evidence. For this analysis, we extracted genes that were commonly identified as significant by edgeR, DESeq2, and the paired t-test, and analyzed one of the clusters obtained from their expression patterns.
This cluster was initially described as being related to signal transduction and transcriptional regulation. After self-verification, however, the interpretation was revised to one centered on cytosolic ribosomes and translation. In other words, even when a gene set was commonly supported by multiple statistical methods, the biological explanation initially generated by the AI was not necessarily appropriate.
Taken together, these results show that the first explanation produced when a gene list is submitted to a generative AI system such as ChatGPT should not be accepted simply because it appears natural and convincing. Even when a gene set is relatively strongly supported statistically, its biological interpretation should still be verified separately using enrichment analysis and external databases.
GeneAgent should therefore be regarded not as a fully developed gene list analysis tool, but as an attempt that points toward a future in which AI-based interpretation is combined with verification using external databases.
As enrichment analysis, gene database searches, literature searches, and comparisons among multiple gene lists become increasingly integrated through AI, the way gene lists are interpreted may change substantially. This is a promising and important field whose future development deserves close attention.
