AI may reduce reliance on costly gene expression testing in some cancers, but it does not replace it outright
AI may reduce reliance on costly gene expression testing in some cancers, but it does not replace it outright
One of the biggest ambitions in modern oncology is to improve precision without endlessly increasing cost and complexity. That is why the idea of an AI tool for cancer gene expression profiling attracts so much attention. If artificial intelligence can pull biologically meaningful information from materials already used in routine care, such as pathology slides, it could reduce dependence on some of the most expensive parts of molecular testing.
The safest reading of the supplied evidence is this: AI is increasingly able to infer clinically relevant molecular or prognostic information from routine samples, which could reduce or complement the need for costly profiling in selected settings. But there is an important limit: the evidence supports partial substitution, inference, or triage far more clearly than it supports a full replacement of gene expression profiling.
Why gene expression profiling matters so much
Gene expression profiling became valuable because it helps clinicians understand how a tumour is behaving beneath the microscope. It can reveal biological activity, aggressiveness, recurrence risk, and sometimes clues about how a cancer may respond to treatment.
The problem is that these tests are not always simple. They can require specialized infrastructure, raise costs, take time, and remain unevenly available across different health systems. That creates a familiar tension in oncology: medicine wants more molecular precision, but real-world systems cannot always deliver that precision at scale for every patient.
This is where AI becomes so interesting. If an algorithm can detect patterns in a digital slide or another routine data source that correlate with molecular and prognostic information, then some of the practical value of a costly test might be reproduced — or at least approximated — in certain clinical situations.
What the studies show most directly
The supplied evidence supports that broader direction well. One of the strongest studies describes a multimodal deep-learning model in endometrial cancer that outperformed a costly current standard for recurrence-risk prediction.
That matters for two reasons. First, it shows that AI models can compete with already established and expensive approaches for clinically important outcomes. Second, it reinforces the idea that not all prognostic information must necessarily come through a traditional molecular route, provided another method can capture biologically meaningful signals with comparable or even better performance.
This does not mean molecular testing has suddenly become unnecessary. It means AI can, in some defined use cases, deliver similar or stronger performance for a specific clinical question.
When the image contains more than the eye can see
Another relevant study involved an image-based AI model in lung adenocarcinoma. The system inferred tumour cell differentiation trajectories and linked image-derived patterns to transcriptomic signatures.
That is particularly interesting because it narrows the gap between two layers of cancer assessment that have often been treated separately: the visible morphology on the slide and the molecular biology underneath it. The implication is powerful. Some of the information that clinicians seek through gene-expression-related methods may already be encoded, in some form, within routine histology — just at a level of complexity that the human eye alone cannot reliably extract.
If that principle continues to hold across more cancers and clinical settings, digital pathology with AI may evolve from a mainly morphological tool into a deeper form of biological inference.
A broader push towards alternative molecular routes
The evidence also includes earlier work on tumour-educated platelets, which helps place this story in a broader context. That work supports a wider movement in oncology: finding less invasive or more accessible ways of obtaining biologically useful information about cancer.
It is not directly about replacing gene expression profiling with slide-based AI, but it points in the same strategic direction. The field is increasingly asking whether some high-value molecular information can be captured through alternative methods that are simpler, less invasive, or more scalable.
AI in pathology fits naturally into that larger shift. It is not an isolated idea, but part of a broader effort to preserve biological sophistication while reducing cost and procedural burden.
What the headline gets right
The headline is right to suggest that AI could reduce reliance on costly methods in some settings. That is consistent with the supplied literature.
It is also right to place this development at the intersection of pathology, AI, and cost reduction. This is one of the most promising frontiers in cancer innovation: extracting more information from data already generated in routine care rather than automatically layering on another expensive test for every patient.
For health systems, that is highly appealing. If some molecular stratification can be anticipated, filtered, or complemented by AI, the benefits could include lower cost, faster turnaround, and better access.
Where the headline risks overstating the case
At the same time, it would go too far to read the headline literally as though AI were ready to replace cancer gene expression profiling broadly across cancers. The supplied evidence does not show that.
What it supports more clearly are specific use cases in specific diseases for specific outcomes, such as recurrence-risk prediction or tumour progression analysis. That is very different from saying cancer care can now move away from traditional gene expression profiling as a general rule.
It is also important to remember that inferring molecularly relevant patterns from an image is not the same as directly measuring gene expression. AI may detect correlated signals that are clinically useful and in some cases highly powerful, but that does not make the approach methodologically identical to molecular testing.
Where AI may enter first
The most realistic near-term role for these tools may lie in three areas:
- triage, to help identify which patients truly need full molecular testing;
- decision support, when expensive profiling is unavailable;
- and cost reduction, where AI already shows strong and repeatable performance for a clearly defined question.
That would already be meaningful. Rather than imagining an abrupt switch from old methods to new ones, it is more realistic to think in terms of integration. AI may reduce dependence on expensive testing in selected circumstances without making it obsolete.
Validation still matters enormously
Another key limit is external validation. AI models can perform impressively in the datasets they were trained on and then weaken when the context changes:
- the cancer type;
- the quality of the slide;
- the laboratory;
- the pathology workflow;
- the scanner;
- or the patient population.
This is a well-known problem in computational medicine. So even when early results are strong, consistency across multiple real-world settings remains essential before a model can be used widely.
Without that step, there is a risk of turning an exciting proof of concept into an unreliable clinical tool.
What this means for clinicians and patients
For clinicians, the most important message is that AI-guided digital pathology may become an added layer of clinical intelligence. It could help identify which tumours appear biologically more aggressive, which patients may need deeper molecular work-up, and where expensive testing can be better targeted.
For patients, that could eventually mean:
- faster access to risk stratification;
- less dependence on expensive testing in some circumstances;
- and more informed decision-making even in settings with limited molecular infrastructure.
But it does not yet mean traditional molecular profiling is disappearing.
The most balanced reading
The most responsible interpretation of the evidence is that AI is beginning to extract molecularly and prognostically relevant information from pathology slides and other routine sources in ways that may complement or, in selected contexts, reduce the need for costly gene expression profiling.
That is a real and promising advance. At the same time, it should not be overstated as though AI were already capable of broadly replacing direct gene expression measurement across cancers.
In short, the strongest story here is not that gene expression profiling has become obsolete. It is that oncology is entering a phase in which AI may allow routine pathology to carry some of the molecular value that once depended entirely on expensive specialized testing. In some settings, that could change practice considerably. But for now, the safer conclusion is not full replacement — it is intelligent complementarity with the potential to lower costs and widen access.