When discussing RNA-Seq analysis, I sometimes encounter explanations that make perfect sense from a bioinformatics perspective but leave me unconvinced as a biologist. Why do discussions about the same data sometimes fail to connect? Exploring this question reveals more than just differences between statistical and biological perspectives. It highlights the difficulty of deciding what questions we should be asking in scientific research. And I believe this has important implications for how we interact with AI.
Are Statistically Sound Analyses and Biological Discoveries Addressing the Same Question?
Suppose we want to identify differentially expressed genes (DEGs) between healthy controls and patients. From a bioinformatics perspective, selecting appropriate filtering methods and statistical models, and ensuring statistical validity, are important considerations. But an experimental biologist may have other questions. Are the healthy controls truly homogeneous? Could examining gene expression profiles reveal previously unknown disease subtypes among patients who appear to have the same condition? Could technical noise or batch effects be present in the data? In other words, we are dealing with data in which we may not even know what the problems are until we examine them.
The disagreement here is not necessarily about whether a statistical model is correct. It is about what each person considers important in the analysis. Choosing appropriate filtering methods and statistical models to identify DEGs is certainly important. However, real datasets may contain biological heterogeneity or experimental and technical problems that were not anticipated when the analysis began. Biologists may want to explore the data with these possibilities in mind. Of course, bioinformaticians and biologists are not confined to separate perspectives in actual research. They often incorporate each other's viewpoints as they work. Nevertheless, their priorities can gradually diverge without either side realizing it, making communication difficult.
Not every researcher has the opportunity to spend sufficient time discussing an analysis with a bioinformatics specialist. Some work in laboratories without such specialists and must analyze data themselves, relying on textbooks and online resources. Others outsource their analyses to another laboratory or an external service provider, where opportunities to discuss research objectives and data characteristics in detail may be limited. As a result, researchers may feel compelled to accept established bioinformatics workflows without sufficiently examining whether those workflows are truly appropriate for their own research objectives.
We Do Research Because We Don't Yet Know What to Look For
There is another difficult aspect to this problem. To test a specific hypothesis statistically, we need to define what we want to evaluate. In exploratory biological research, however, we often do not yet know what we should be looking for. We discover unexpected expression patterns and try to understand what they mean. We formulate multiple hypotheses and investigate them through additional analyses and experiments. Sometimes, this process even changes the original research objective.
We conduct research precisely because we do not yet know what to look for, yet we are expected to decide what to look for before we can begin the analysis.
This is a fundamental challenge in exploratory scientific research. In an earlier article, Between Statistics and Biology: Thoughts After Reading Simply Statistics, I discussed the importance of formulating questions, examining data, and revising those questions, drawing on John Tukey's ideas about exploratory data analysis. Kelder et al. also emphasized the importance of finding the right questions, rather than simply seeking the right answers, in a 2010 article in PLOS Biology.
For example, visualizing gene expression profiles from a group of patients might reveal that a few patients differ markedly from the others. Whether those differences reflect previously unknown disease characteristics, sample quality issues, or batch effects cannot be determined without further investigation. We would need to examine sample information and quality, and perhaps reconsider our original analysis strategy. Rather than simply executing a predefined workflow correctly, we need to recognize new questions as we examine the data and reconsider what deserves investigation. This is also an essential part of biological research.
It is important to keep this distinction in mind when learning bioinformatics theory and analytical methods. Is the discussion about how to correctly solve a problem when we already know what we want to investigate? Or is it about how to discover new questions when we do not yet know what we should be investigating? Researchers who have ample opportunities to discuss these issues with bioinformatics specialists are fortunate. Those who do not must judge for themselves which type of problem the methods presented in textbooks and tutorials are primarily intended to address.
In fact, similar misunderstandings frequently occur in conversations with AI.
For researchers who cannot readily consult a bioinformatics specialist, AI offers something that textbooks and tutorials alone cannot: the opportunity for interactive discussion. Being able to ask questions immediately and discuss analytical methods suited to one's own research is a major advantage. However, if we ask a vague question such as "How should I analyze my data?", AI may fail to distinguish whether we want to know how to correctly solve a well-defined problem or how to discover something we do not yet know to look for. Instead, it may default to standard analytical workflows and statistical explanations that are extensively covered in the literature.
Moreover, unlike a textbook, AI responds to questions about our individual research projects. This can create the impression that it understands our research objectives, experimental background, and even our unspoken assumptions. In reality, AI may simply be filling in information we have never provided with assumptions of its own. It is hardly surprising that those assumptions sometimes miss the mark. Yet because AI produces fluent and persuasive explanations, researchers may not even realize that essential information has never been communicated to it.
What Is Included in the Scientific Knowledge AI Has Learned?
In the previous section, we considered how AI may fill gaps in the information provided by researchers with its own assumptions. But that is not the only problem. The scientific knowledge AI has learned or can access may itself contain gaps and biases.
First, not everything published in scientific papers is necessarily correct. Some studies involving omics data have problems with experimental design, data quality, or analytical methods that raise questions about the reliability of their conclusions. For example, a 2024 study published in Genome Biology examined 40 publicly available clinical RNA-Seq datasets and found substantial imbalances in sample quality between comparison groups in 14 of them, affecting the results of DEG analysis. The study examined a limited set of datasets, so its findings should not be generalized to all omics studies. Nevertheless, it demonstrates that even published scientific results can be affected by overlooked problems.
Furthermore, published papers do not tell us everything that happens in scientific research. Biological laboratories accumulate countless experiences involving experiments that failed to produce the expected results, observations that could not be reproduced, and problems discovered during data analysis. Not all of these experiences are published. Publication bias, whereby negative results or findings that do not reach statistical significance are less likely to be published, has long been recognized as a problem in scientific research.
Ending publication bias: A values-based approach to surface null and negative results
And unpublished knowledge is not limited to failed experiments. Experienced researchers sometimes notice that something looks unusual in a dataset or suspect an unexpected cause based on previous experience. They may be able to explain the reasons for such judgments if asked, but those observations and insights are not always recorded in writing. This kind of experiential knowledge may be difficult to acquire fully simply by learning from vast collections of published scientific literature.
Of course, this problem is not unique to AI. Human researchers cannot read every scientific paper or know every failure and experience accumulated in other laboratories. Indeed, AI has enormous potential precisely because it can connect knowledge across a much broader range of fields than any individual researcher could possibly master.
The important point is that biases in the scientific knowledge AI learns or accesses may cause its answers to be biased in ways that users do not readily recognize. AI may explain standard analytical methods in great detail while lacking sufficient information about the problems those methods have caused in actual research or the trial and error through which researchers have learned to use them. And because AI's explanations are logical and persuasive, researchers may find it difficult to recognize that such biases exist.
But AI development will not stop here. Even if the scientific knowledge available to AI contains gaps and biases, AI could make new contributions to research if it becomes better at questioning existing knowledge and assumptions, exploring alternative possibilities, and recognizing what remains unknown. The key to this progress is AI's reasoning ability.
Understanding Different Types of Reasoning to Move Beyond the Limits of Knowledge and Information
Scientific research involves several forms of reasoning. To explore their characteristics and how AI might develop in each area, I will organize them into five categories.
1. Deductive Reasoning — Deriving Logically Valid Conclusions from Given Premises
Mathematical proofs, reasoning about statistical models, and verifying computer programs are typical examples. In these areas, correct answers can often be defined relatively clearly, making it easier to evaluate AI's capabilities. In particular, when reasoning can be checked through mathematical verification or program execution, it becomes possible to introduce mechanisms for detecting and correcting errors.
Today's AI can solve advanced mathematical and programming problems, yet it may sometimes overlook relatively simple logical consequences. Its performance is not always consistent. Nevertheless, because the correctness of deductive reasoning is relatively easy to verify, substantial further progress can be expected in this area.
However, even logically flawless reasoning does not necessarily lead to conclusions that accurately describe reality if the starting premises are wrong. In other words, improvements in deductive reasoning do not automatically mean that AI will become capable of identifying biases in the scientific knowledge it has learned or can access.
2. Inductive Reasoning — Discovering Patterns in Data
Finding common patterns across large gene expression datasets or identifying trends shared by different experiments are typical examples of inductive reasoning. AI can process enormous amounts of data and identify complex patterns that might escape human attention. As its capacity to analyze larger and more diverse datasets and integrate information across different fields improves, further progress can be expected.
However, patterns discovered in one dataset do not necessarily hold for new samples or different experimental conditions. For example, a gene expression pattern that appears to be associated with a particular disease might actually reflect how samples were collected or a batch effect. Similarly, if successful studies are overrepresented in the scientific literature available to AI while failed attempts are underrepresented, the generalizations drawn from that literature may also be biased.
In other words, improvements in inductive reasoning do not automatically overcome biases in the underlying data or scientific knowledge.
3. Hypothesis Generation (Abductive Reasoning) — Exploring Possible Explanations for Observations
Consider a gene expression pattern observed in only three out of 60 patients. Many explanations are possible: a previously unknown disease subtype, differences in treatment history, variation in cell composition, or technical problems with the measurements. Abductive reasoning involves proposing possible explanations for observations like these. Because AI can connect vast amounts of knowledge from different scientific fields, it may suggest hypotheses that human researchers would not have considered.
A 2026 study published in Nature on Co-Scientist described an AI system designed to generate, critique, and refine scientific hypotheses. The study also showed that some AI-generated hypotheses were supported by biological experiments. This suggests that AI may contribute not only to explaining existing scientific knowledge but also to making new discoveries. As its ability to integrate knowledge and reason improves, its contributions to scientific research could become even greater.
However, many observations have more than one possible explanation, and determining which hypothesis is correct is not straightforward. Even if AI proposes a highly convincing hypothesis, that hypothesis may not accurately explain the underlying biological phenomenon. Furthermore, if existing scientific knowledge is biased, AI-generated hypotheses may favor familiar explanations and fail to explore less conventional possibilities sufficiently.
Even so, imagine AI generating new hypotheses from unexpected observations, connecting knowledge across disciplines, and selecting hypotheses worth testing. AI may eventually be able to carry out autonomously some of the intellectual work that researchers perform at the forefront of exploratory science today. Combined with automated experimentation, this could enable large numbers of hypotheses to be generated, tested, and refined through repeated cycles, on a massive scale and in parallel. In effect, this could lead to the industrialization of scientific research.
4. Causal Reasoning — Investigating the Causes of Observed Changes
Even if the expression of a particular gene is strongly associated with a disease, that does not necessarily mean the gene causes the disease. Its expression may have changed as a consequence of the disease, or another factor may be responsible for both. Causal reasoning involves examining such possibilities and attempting to identify the causes underlying observed relationships. If AI becomes better at constructing causal models, designing experiments, and integrating different types of data, it could make major contributions to identifying causal factors and testing hypotheses.
However, observational data alone cannot always distinguish between competing causal explanations. No matter how advanced AI's reasoning becomes, it cannot establish a causal relationship when the available data do not provide enough information to identify it. In some cases, additional evidence will be needed, such as experiments that manipulate gene function or observations under new experimental conditions.
Nevertheless, research could change dramatically if AI becomes capable of comparing alternative causal hypotheses and designing the specific experiments and analyses required to test them. Furthermore, if AI is integrated with automated experimental equipment and can autonomously organize what is known and unknown, decide what needs to be tested next, and repeatedly conduct experiments and analyses, fully autonomous research may become a realistic possibility.
5. Reconsidering the Research Question — What Should We Really Be Investigating?
However, there is a major difference between AI acquiring highly advanced hypothesis-generation capabilities, leading to the industrialization or automation of research, and AI becoming capable of deciding what should be investigated in the first place. Bridging that gap would represent a substantial leap. Battleday and Gershman also distinguish between solving predefined scientific problems and discovering which problems are worth investigating in their 2026 paper.
For example, suppose we initially intended to identify DEGs between patients and healthy controls. But after examining the data more closely, we notice several distinct expression patterns within the patient group. We might then conclude that exploring previously unknown disease subtypes would be more important than identifying DEGs shared by the entire patient group. This is not merely a change in analytical method. It means reconsidering the research question itself — what we should be trying to discover.
In situations like this, AI does not need to determine research objectives autonomously. Researchers can explain their thinking to AI and ask whether there are alternative possibilities or important issues they may have overlooked. Through that dialogue, they can reconsider their own research objectives. However, as discussed earlier, the scientific knowledge available to AI has gaps and biases, and researchers have not necessarily communicated all of their experience and concerns. This is precisely why it is valuable to develop the discussion by telling AI what seems questionable and what we consider important, rather than simply accepting its suggestions.
From this perspective, I do not think human researchers will disappear, even as research becomes increasingly industrialized and automated. Of course, we do not yet know how far AI will develop these capabilities. But expanding the scope of research automation is not the same as making human researchers unnecessary. Human researchers can contribute knowledge and experience to AI, while conversations with AI can also change how researchers themselves think. I believe this reciprocal process offers enormous possibilities for scientific research in the age of AI.
We have now examined five forms of reasoning, but the purpose was not simply to speculate about AI's future. This framework can also help us think about what researchers should consider, what information they should provide to AI, and what kinds of questions they should ask when using AI in research today.
What Should Researchers Tell AI, and What Questions Should They Ask?
1. Explain Your Research Objectives and What You Still Don't Know
If you ask, "Which method should I use for DEG analysis of RNA-Seq data?", AI will probably explain standard methods such as edgeR and DESeq2. However, the most appropriate analytical approach may differ depending on whether you want to identify expression changes shared across an entire disease group or characteristics found in only a few patients.
When consulting AI about research, it is important to explain whether you want to solve a well-defined problem or explore your data to discover previously unrecognized phenomena or questions. Sometimes both are necessary. In that case, you might ask: "Please consider not only standard analytical workflows for achieving my original research objectives, but also approaches that could help uncover unexpected phenomena or previously unrecognized questions in the data."
For example, you could say: "I am interested not only in changes shared across the disease group, but also in expression patterns characteristic of a small number of patients. Please do not assume that the disease group is homogeneous, and compare possible analytical approaches." This encourages AI to consider the problem from a broader perspective. If your research objective is not yet clear, you could even say: "I don't yet know what would be most meaningful to investigate. Please suggest different research questions that could emerge from exploring the data."
2. Share Your Own Experience and Anything That Concerns You About the Data
The scientific literature available to AI does not contain all the experience accumulated in research laboratories. That is why information known to the researcher can be valuable. Examples include previous observations of unstable measurements under particular experimental conditions, results that changed depending on sample preparation methods, or unusual patterns noticed while visualizing the data.
For instance, you might say: "In a previous experiment using the same measurement method, I observed abnormal values under a particular condition. Please consider whether a similar problem might be affecting the current results." AI can then take that information into account.
Of course, a researcher's personal experience is not necessarily a reliable explanation in itself. The important point is to consider researchers' own experiences alongside published scientific knowledge, and to examine their validity through discussions with AI.
3. Work with AI to Develop Analytical Methods That Fit Your Research Objectives
Once you have explained your research objectives, what remains unknown, your experimental experience, and anything that concerns you about the data, you can work with AI to explore the analytical methods themselves.
Traditionally, researchers have generally selected methods suited to their objectives from those developed by bioinformatics specialists. However, if AI can serve as a research partner, there is no need to be confined to existing analytical methods. Depending on the research objectives and characteristics of the data, it may be possible to combine existing approaches, modify them, or develop entirely new analytical methods.
For example, you could ask: "Based on the research objectives, experimental experience, and concerns about the data that I have explained so far, please suggest possible analytical approaches without limiting yourself to established methods. Explain what each approach could reveal, what problems might arise, and how the reliability of the results could be evaluated."
By asking AI to write the necessary code, researchers may also be able to try analytical approaches that previously required specialized programming skills.
Of course, a newly proposed analytical method is not necessarily appropriate. Its statistical validity, the quality of the underlying data, and the reproducibility of its results must be carefully evaluated. But the guiding principle should be clear: Rather than adapting our research objectives to existing analytical methods, we should develop analytical methods that fit our research objectives. Working with AI could make this approach to research much more accessible.
4. Ask AI to Consider Possibilities That Challenge Your Hypothesis
One issue that deserves particular attention in conversations with AI is its tendency, in some circumstances, to agree too readily with the user. When a researcher is strongly committed to a particular hypothesis, AI may prioritize explanations that support it without adequately examining alternatives.
To address this, you can explicitly say: "You don't need to agree with me. Please actively consider ways in which my hypothesis might be wrong and whether alternative explanations could account for the observations."
You could go further: "Please identify the evidence supporting and contradicting each hypothesis, and suggest analyses or experiments that could distinguish between them." This can help move the discussion toward further investigation.
However, asking AI for counterarguments does not guarantee that its objections will be correct. The goal is neither to make AI agree with us nor to make it disagree, but to compare alternative possibilities based on evidence.
5. Ask AI to Reconsider the Research Question Itself
This is the suggestion I particularly want to emphasize in this article.
"Please also evaluate whether the research objectives and analytical strategy I have chosen are appropriate. If the characteristics of the data or existing scientific knowledge suggest other important research questions, please actively propose them."
Researchers themselves do not necessarily know from the outset what they should be investigating. Unexpected findings may emerge during analysis, making it necessary to reconsider the original approach. If AI can help examine not only how to solve a predefined problem but also whether the problem itself has been framed appropriately, it could become a valuable research partner.
When AI proposes a new research question, it is also important to ask: "Why do you consider this question important? What information supports your suggestion, and what remains uncertain?" We need to distinguish between suggestions supported by published evidence, hypotheses developed through reasoning, and speculation that lacks sufficient supporting evidence.
Rather than simply asking AI for the correct answer, we should consider what may be missing from its response and work with AI to explore what questions we should be asking.
This is not something we need to wait for future AI systems to achieve. It is something we can already practice in conversations with today's AI, and it has the potential to broaden researchers' thinking.
