Synexmedia.com

March 2026

Introduction: A Revolution You Can Feel

Imagine you walk into a hospital with a cough that will not go away. After weeks of tests, a doctor sits you down and tells you that you have lung cancer. For most of the twentieth century, what happened next was remarkably predictable: you would be given a powerful chemical cocktail designed to kill fast-growing cells throughout your entire body. The drugs did not know the difference between a cancer cell in your lung and a perfectly healthy cell in your stomach lining, your hair follicles, or your bone marrow. They attacked everything. The nausea, the hair loss, the crushing fatigue—these were not unfortunate side effects. They were the intended mechanism of action, collateral damage in a war fought with a sledgehammer.

Now imagine a different scenario. The same diagnosis, the same hospital—but this time, a tissue sample from your tumour is sent to a laboratory where machines sequence every letter of its genetic code, measure which genes are switched on and which are silent, catalogue the proteins the cancer is manufacturing, and even analyse the chemical fuel it is burning to grow. All of that information—billions of data points—is then fed into an artificial intelligence system that cross-references your tumour’s unique biological fingerprint against vast databases of known cancer vulnerabilities. Within days, your oncologist receives a report identifying the precise molecular weakness driving your specific cancer, and prescribes a drug engineered to exploit that weakness and almost nothing else.

That second scenario is not science fiction. It is the emerging reality of precision oncology, and it represents one of the most profound transformations in the history of medicine. To put the scale of this change in perspective: as recently as the year 2000, oncologists choosing a treatment for most solid tumours had essentially no molecular information to guide them. The Human Genome Project was not yet complete. The cost of sequencing a single human genome was measured in the hundreds of millions of dollars. The idea that you could sequence a tumour’s DNA, measure its RNA, catalogue its proteins, and feed the results into an artificial intelligence system would have sounded like a plot from a science fiction novel.

Yet here we are, barely a quarter-century later, and all of those things are not only possible but increasingly routine at major cancer centres around the world. The story of how we got here—and where we are going—is a story about code. Not just the genetic code written in the spiralling double helix of DNA, but the computational code written by researchers building artificial intelligence systems capable of doing what no human mind can: making sense of the staggering complexity of cancer.

The Old Way: When One Size Was Supposed to Fit All

To appreciate the revolution underway, you first need to understand what came before it. For decades, cancer treatment was organised around a deceptively simple idea: cancers that looked alike under a microscope should be treated the same way. A pathologist would examine a tissue biopsy, identify the organ where the tumour had originated—lung, breast, colon—and classify it based on its appearance. Was it a squamous cell carcinoma? An adenocarcinoma? The name determined the treatment. Patients with the same classification received the same chemotherapy regimen, regardless of what was actually happening inside their tumours at the molecular level.

This approach is called cytotoxic chemotherapy, and the name tells you everything you need to know. “Cyto” means cell. “Toxic” means poisonous. These drugs are designed to be poisonous to cells—ideally cancer cells, but in practice, to all rapidly dividing cells in the body. The logic was brutal but straightforward: cancer cells divide faster than most normal cells, so if you flood the body with a chemical that disrupts cell division, you will kill more cancer cells than healthy ones. The margin, however, was often razor-thin.

The problem was that this population-based approach ignored a fundamental truth about cancer that scientists were only beginning to understand: no two cancers are the same. Two patients sitting in the same oncology ward, both diagnosed with Stage III non-small-cell lung cancer, might have tumours driven by completely different molecular engines. One patient’s cancer might be fuelled by a mutation in a gene called EGFR. The other’s might be driven by a rearrangement involving a gene called ALK. Under the old system, both would receive the same platinum-based chemotherapy. One might respond well. The other might not respond at all. And no one could predict in advance which patient would be which, because no one was looking at the molecular machinery under the hood.

The inconsistency of treatment responses was not a mystery—it was a clue. It told researchers that the real identity of a cancer was not defined by how it looked under a microscope, but by the molecular code driving its behaviour. To crack that code, medicine would need tools that did not yet exist.

The DNA Revolution: Reading Cancer’s Secret Instruction Manual

Every cell in your body contains a copy of your complete genetic code—roughly three billion pairs of chemical letters, called bases, strung along the famous double helix of deoxyribonucleic acid, or DNA. This code is the instruction manual for building and operating a human being. When cancer develops, something has gone wrong with those instructions. A letter has been changed, a section has been deleted, two chapters that should never have been joined have been fused together. These errors—called mutations—can switch on genes that tell cells to grow without stopping, or switch off genes that normally act as emergency brakes.

The technology that made it possible to read these mutations at scale is called next-generation sequencing, or NGS. Think of it as a high-speed scanner for DNA. Earlier sequencing methods could read a single gene at a time, slowly and expensively. NGS machines can read millions of DNA fragments simultaneously, assembling the complete sequence of hundreds or even thousands of genes in a matter of hours. The cost of sequencing an entire human genome dropped from roughly three billion dollars in 2003 to under a thousand dollars by the early 2020s, a pace of cost reduction that outstripped even the famous Moore’s Law governing computer chip performance.

With NGS, researchers could finally catalogue the mutations driving individual cancers. They discovered that the landscape was far more diverse than anyone had imagined. Cancers harboured single nucleotide variants, or SNVs—a single letter of genetic code changed to a different one, like a typo in a vast manuscript. They found multi-nucleotide variants, or MNVs, where two or more adjacent letters were changed simultaneously—a subtle but important distinction, because treating these as separate single-letter typos could lead to completely wrong predictions about what the mutation does. They found insertions and deletions, known collectively as indels, where stretches of genetic code were either added or removed. And they found gene fusions—dramatic structural rearrangements where two genes that should be on separate parts of the genome become physically joined, creating a hybrid gene that produces a rogue protein driving uncontrolled growth.

Each of these mutation types could, in principle, point to a specific vulnerability that a targeted drug could exploit. The BCR-ABL1 fusion, for example—found in chronic myeloid leukaemia—became the target of imatinib (marketed as Gleevec), one of the first great success stories of targeted therapy. But identifying these mutations across the full spectrum of cancer types, in real patients, in real time, required processing volumes of data that were rapidly exceeding human capacity.

Beyond DNA: The Multi-Omic Orchestra

Here is where the story gets genuinely exciting—and genuinely complex. DNA is the master blueprint, but it is not the whole story. To understand what a cancer is actually doing at any given moment, you need to look at multiple layers of biological information, each one adding a new dimension to the picture. Scientists call this approach multi-omics, and each layer, or “omic,” is named for the type of molecule it studies.

The first layer beyond DNA is the transcriptome—the complete set of RNA molecules in a cell at a given time. Think of DNA as a library of cookbooks. The transcriptome is the set of recipes that the cell has actually photocopied and brought into the kitchen. Just because a gene exists in the DNA does not mean it is being used. Transcriptomics, typically performed through RNA sequencing, reveals which genes are actively being read and translated into action. This matters enormously for cancer treatment, because a mutation in a gene that is not being expressed—not being read—is biologically irrelevant. Conversely, some cancers dramatically crank up the expression of certain genes, flooding the cell with growth-promoting signals even when the DNA sequence itself looks normal.

RNA sequencing is also critical for detecting fusion transcripts—the messenger RNA molecules produced by fused genes. Some fusions are notoriously difficult to detect by DNA sequencing alone, because the breakpoints—the exact spots where the two genes were spliced together—can occur in vast, hard-to-sequence stretches of non-coding DNA. But when the fused gene is read into RNA, the fusion becomes much easier to spot. Research using large real-world patient cohorts has shown that combining DNA and RNA sequencing can increase the detection of clinically significant driver fusions by roughly sixteen to twenty-one per cent compared to DNA sequencing alone, depending on whether the fusion falls within or outside of an established treatment indication. That represents a meaningful number of patients who would have been missed entirely by DNA-only testing.

The next layer is the proteome—the complete set of proteins in a cell. If DNA is the blueprint and RNA is the photocopy, then proteins are the actual machines, structures, and signals that do the work. Mass spectrometry, a technology that separates and identifies thousands of proteins by their molecular weight, allows researchers to catalogue these molecules with extraordinary precision. Crucially, proteins are subject to post-translational modifications—chemical tags that are added after the protein is built, which can dramatically alter its behaviour. A protein might be inactive until a phosphate group is attached to it, flipping it into an active state that drives cancer growth. These modifications are invisible to genomics and transcriptomics. Only proteomics can see them.

Finally, there is the metabolome—the complete set of small molecules involved in the cell’s chemical reactions. Cancer cells are notorious for reprogramming their metabolism. One of the most famous examples is the Warburg effect, named after the German Nobel laureate Otto Warburg, who observed in the 1920s that cancer cells preferentially use a less efficient form of energy production called glycolysis, even when oxygen is abundant. Normal cells, when oxygen is available, use a highly efficient process called oxidative phosphorylation to extract maximum energy from glucose. Cancer cells, by contrast, convert glucose to lactate even in the presence of oxygen—a phenomenon known as aerobic glycolysis. Modern research has revealed that this metabolic reprogramming is not simply wasteful; it provides the cell with the chemical building blocks it needs to sustain rapid division, a fact that overturns Warburg’s original belief that the cancer cells’ mitochondria—their energy factories—were broken. The mitochondria work fine. The cancer is choosing to fuel itself differently. Metabolomics can detect these distinctive metabolic signatures, providing yet another window into the tumour’s biology.

Each of these layers—genomics, transcriptomics, proteomics, metabolomics—offers a partial view. The challenge, and the promise, lies in combining them into a single, unified picture. This is where artificial intelligence enters the story.

Enter Artificial Intelligence: The Pattern Hunter

A single tumour biopsy, once processed through the full battery of multi-omic technologies, can generate data measured in terabytes. Multiply that by thousands of patients, and you have a data landscape of almost incomprehensible scale and complexity. No team of human analysts, no matter how brilliant, can hold all of these variables in mind simultaneously and identify the subtle patterns that distinguish a cancer likely to respond to Drug A from one that will resist it. This is not a failure of human intelligence. It is a recognition that certain problems require a fundamentally different kind of computation.

Artificial intelligence—specifically, a subset called machine learning—provides that computation. At its core, machine learning is the science of building algorithms that improve their performance through experience. Instead of being explicitly programmed with rules (“if mutation X is present, prescribe drug Y”), these systems are trained on enormous datasets of real patient outcomes and learn to discover the rules themselves. The more data they process, the more nuanced and accurate their predictions become.

Within machine learning, a technique called deep learning has proven particularly powerful for biological data. Deep learning uses structures called neural networks—loosely inspired by the branching connections of neurons in the human brain—to process information through multiple successive layers. Each layer extracts progressively more abstract features from the raw data. The first layer of a neural network analysing gene expression data might learn to recognise simple patterns, like which genes tend to be expressed together. Deeper layers might learn to recognise complex biological pathway signatures that correlate with drug sensitivity or resistance. The network builds up an internal representation of the data that captures relationships far too intricate for any human to specify by hand.

To make this more concrete, imagine you are trying to predict whether a particular tumour will respond to an immunotherapy drug. The relevant signals might not be any single gene or protein in isolation. They might be a specific combination: a particular mutation in a DNA repair gene, plus elevated expression of a set of immune checkpoint genes, plus the presence of certain inflammatory proteins, plus a metabolic profile suggesting high glycolytic activity. A deep learning model can learn to recognise this combination as a unified pattern predictive of immunotherapy response, even if no human researcher has ever explicitly described it. This ability to discover complex, multi-layered patterns in high-dimensional data is what makes deep learning transformative for precision oncology.

Deep learning has already proven its worth in one of the most fundamental tasks in genomic analysis: variant calling—the process of accurately identifying which mutations are genuinely present in a tumour’s DNA versus which are sequencing errors. Tools like NeuSomatic, a convolutional neural network-based somatic mutation detector, and DeepSomatic, adapted from Google’s DeepVariant architecture for cancer-specific use, represent a new generation of variant callers that can outperform traditional statistical methods in certain contexts, particularly for low-frequency mutations that are easy to miss.

One of the most significant recent developments is the application of a specific neural network architecture called the transformer to multi-omic data fusion. Transformers were originally developed for natural language processing—they are the technology behind large language models—but their core innovation, a mechanism called self-attention, turns out to be extraordinarily well suited to biological data. Self-attention allows the model to weigh the importance of every piece of input data relative to every other piece, regardless of where they sit in the dataset. In a multi-omic context, this means a transformer can learn that a particular gene expression pattern in the transcriptome is only clinically significant when it co-occurs with a specific protein modification detected by proteomics. It can model cross-modal relationships—connections that span different data types—in ways that simpler models cannot.

Published research bears this out. Models like the Multimodal Co-Attention Transformer, developed for fusing whole-slide pathology images with genomic data, have demonstrated improved survival prediction in cancer patients. Another system, called SurvPath, presented at a major computer vision conference in 2024, integrates biological pathway information derived from transcriptomics with digital pathology image patches. A system called DeePathNet uses transformer architecture to integrate pathway-level information across proteomics, genomics, and transcriptomics simultaneously. These are not theoretical proposals. They are working systems with published results—though it is important to note that virtually all of them remain research-stage technologies validated on retrospective datasets, not yet deployed in routine clinical care.

Making Sense of the Data Deluge: How AI Fuses Multiple Omic Layers

Beyond transformers, researchers have developed a range of sophisticated computational frameworks for integrating multi-omic data. Understanding these methods matters because the way you combine information from different biological layers determines what patterns you can find and what clinical predictions you can make.

One influential approach is called Similarity Network Fusion, or SNF, originally published in the journal Nature Methods. The idea is elegant: for each data type—say, gene expression and DNA methylation—the algorithm builds a network in which patients who are similar to each other based on that data type are connected. It then fuses these separate networks into a single unified network that captures the combined information from all layers. Studies have shown that this fused network can identify cancer subtypes more accurately than any single data layer alone.

Another widely used framework is called MOFA, short for Multi-Omics Factor Analysis, and its scalable successor MOFA+. These tools use a statistical technique called Bayesian factor analysis to identify the hidden factors—the latent variables—that explain the shared and data-type-specific patterns of variation across multiple omic datasets. If you imagine each omic layer as a different camera angle on the same scene, MOFA is the algorithm that figures out what the underlying scene looks like by combining all the camera views.

Other established methods include JIVE, which separates the joint variation shared across data types from the variation unique to each one, and iCluster, which performs integrative clustering to identify patient subgroups defined by molecular patterns spanning multiple omic layers. All of these methods have been applied to large-scale cancer datasets and have contributed to our growing understanding of tumour biology.

Yet a critical gap remains between these research tools and clinical deployment. The vast majority of multi-omic integration frameworks have been validated retrospectively—meaning they were tested on historical data to see whether their outputs match known outcomes. Very few have been tested prospectively in clinical trials, which is the gold standard for proving that a computational tool actually improves patient care. Bridging that gap is one of the central challenges of the field.

Mapping the Web of Life: Graph Neural Networks

Cancer is not just a collection of broken genes. It is a disease of networks. Inside every cell, thousands of proteins interact with one another in elaborate webs of signalling and regulation. A growth signal received at the cell surface triggers a cascade of protein-to-protein interactions that eventually reaches the nucleus and switches on genes that promote cell division. In a cancer cell, one or more nodes in this network have gone haywire, sending constant growth signals even when no external trigger is present.

To model these networks, researchers have turned to a specialised class of artificial intelligence called graph neural networks, or GNNs. The mathematics of graph theory—the study of networks composed of nodes and edges—provides a natural framework for representing biological systems. Each protein becomes a node. Each known interaction between two proteins becomes an edge connecting them. A GNN can then learn to propagate information through this network, identifying which nodes are most critical to the cancer’s survival.

The practical implications are significant. If a GNN can identify a protein that sits at a critical junction in a cancer’s signalling network—a bottleneck through which multiple growth signals must pass—then targeting that protein with a drug could shut down multiple oncogenic pathways simultaneously. Several published models have demonstrated this capability. GraphSynergy uses graph convolutional networks on protein-protein interaction data to predict which combinations of drugs will act synergistically against cancer. DriverOmicsNet integrates multi-omic data with protein interaction networks to identify cancer driver genes. A comprehensive 2023 survey catalogued approximately ninety distinct applications of graph neural networks in cancer research, with drug target discovery and protein interaction network modelling among the most common use cases.

For the non-specialist, the key insight is this: traditional approaches to cancer drug development have often focused on individual molecules in isolation. Graph neural networks represent a fundamentally different philosophy—one that treats the entire system of molecular interactions as the object of study, and seeks to find the precise points where a well-aimed intervention can bring the whole dysfunctional network back under control.

The Testing Pipeline: How It Works in Practice

All of this might sound abstract, so let us trace how multi-omic AI-driven precision oncology works for an actual patient. A tumour biopsy—typically preserved in a wax-like substance called formalin-fixed, paraffin-embedded tissue, or FFPE—is sent to a genomic testing laboratory. Alternatively, a simple blood draw can provide cell-free DNA, or cfDNA—tiny fragments of tumour DNA that have been shed into the bloodstream, enabling what is known as a liquid biopsy.

The laboratory then runs the sample through an FDA-approved comprehensive genomic profiling panel. Several such panels now exist, each approved through the rigorous Premarket Approval pathway. FoundationOne CDx, approved in November 2017, sequences 324 genes from tumour tissue and detects mutations, copy number changes, rearrangements, and two genomic signatures: microsatellite instability and tumour mutational burden. Guardant360 CDx, approved in August 2020, analyses plasma cfDNA across 55 genes—a liquid biopsy that requires only a blood sample. FoundationOne Liquid CDx, also approved in August 2020, provides a broader pan-tumour liquid biopsy. Tempus xT CDx, approved in April 2023, sequences 648 genes from tumour tissue alongside matched normal tissue from blood or saliva, detecting single nucleotide variants, multi-nucleotide variants, indels, and microsatellite instability. Illumina’s TruSight Oncology Comprehensive, approved in August 2024, analyses 517 genes and is notable for being the first distributable comprehensive genomic profiling kit—meaning it can be run at hospitals’ own laboratories, not just at a centralised facility. And Caris Life Sciences’ MI Cancer Seek, approved in late 2024, uses whole-exome and whole-transcriptome sequencing to cover 228 genes.

It is important to understand that DNA-based testing and RNA-based testing are often separate assays, even when offered by the same company. Tempus, for example, offers RNA sequencing through a separate product called xR, which operates as a laboratory-developed test under a different regulatory framework from the FDA-approved xT CDx. Clinicians commonly order both tests together, because the combination captures a wider range of clinically relevant alterations, but they are distinct products with distinct regulatory statuses.

The results are compiled into a report that translates the molecular findings into clinical language. For each identified alteration, the report indicates whether an FDA-approved targeted therapy exists, whether relevant clinical trials are recruiting, and what the published evidence says about likely sensitivity or resistance. The oncologist then uses this information, alongside the patient’s clinical history and preferences, to craft an individualised treatment plan.

It is worth pausing to appreciate the interpretive challenge this creates. A comprehensive genomic and transcriptomic report for a single patient might list dozens of molecular alterations, only some of which are clinically actionable. The oncologist must weigh the strength of evidence for each potential target, consider potential drug interactions, factor in the patient’s overall health and co-morbidities, and decide whether a targeted therapy, immunotherapy, conventional chemotherapy, or some combination offers the best chance of meaningful benefit. Many institutions convene molecular tumour boards—multidisciplinary panels of oncologists, pathologists, geneticists, and bioinformaticians—to review complex cases. AI-powered clinical decision support tools are increasingly being developed to assist with this synthesis, ranking treatment options and matching patients to trials, but the final decision remains in the hands of the treating physician and their patient.

How long does this all take? Turnaround time depends heavily on the test and the specimen. Tissue-based NGS testing typically takes about three to four weeks from biopsy to report. Liquid biopsy testing is faster: the NILE study, a prospective trial comparing plasma cfDNA profiling to tissue genotyping in newly diagnosed advanced non-small-cell lung cancer, reported a median plasma turnaround time of nine days versus fifteen days for tissue. That difference matters when treatment decisions are time-sensitive.

The Proof: Real Trials, Real Patients, Real Results

The most compelling evidence that precision oncology actually works comes not from laboratory demonstrations or retrospective analyses, but from prospective randomised clinical trials—studies in which patients are randomly assigned to receive molecularly guided treatment or standard care, and the outcomes are rigorously compared. Several landmark trials have now provided this evidence across multiple cancer types.

In breast cancer, the TAILORx trial, published in the New England Journal of Medicine in 2018, enrolled thousands of women with hormone receptor-positive, HER2-negative, node-negative early-stage breast cancer—one of the most common forms of the disease. Each patient’s tumour was analysed using a 21-gene expression assay that computes a “recurrence score,” a number that reflects how likely the cancer is to return. Patients with intermediate scores were randomly assigned to receive either endocrine therapy alone or endocrine therapy plus chemotherapy. The result was striking: endocrine therapy alone was noninferior to the combination, meaning that adding chemotherapy provided no meaningful additional benefit for this group. The National Cancer Institute estimated that approximately seventy per cent of women with this type of breast cancer could safely avoid chemotherapy based on their recurrence score—sparing them months of toxic treatment and its associated misery.

But what about patients whose cancer has already spread to nearby lymph nodes? The RxPONDER trial, published in the same journal in 2021, tackled exactly this question. It used the same 21-gene recurrence score in women with hormone receptor-positive, HER2-negative breast cancer involving one to three positive lymph nodes. The finding was nuanced and clinically critical: the benefit of adding chemotherapy depended on the patient’s menopausal status. Premenopausal women in the study showed a significant improvement in invasive disease-free survival when chemotherapy was added, but postmenopausal women with recurrence scores of zero to twenty-five showed no benefit at all. The same genomic score, applied to the same cancer type, produced different treatment recommendations depending on a biological variable—menopausal status—that the traditional one-size-fits-all approach would have ignored entirely.

A third breast cancer trial, MINDACT, took a different approach. It used a 70-gene signature called MammaPrint alongside a traditional clinical risk assessment to classify patients into four groups based on whether their genomic risk and clinical risk were each high or low. The most important finding concerned patients who were clinically high-risk—the ones who, under conventional guidelines, would have been told they needed chemotherapy—but genomically low-risk. These patients had a five-year distant metastasis-free survival rate of 94.7 per cent without chemotherapy. The genomic test identified a large group of women who would have been overtreated under the old rules.

The principle extends well beyond breast cancer. In colon cancer, the randomised DYNAMIC trial tested whether circulating tumour DNA—those tiny fragments of cancer DNA floating in the bloodstream—could guide decisions about who needs chemotherapy after surgery. In the ctDNA-guided arm, patients whose blood tests came back positive for tumour DNA after surgery received chemotherapy; those whose tests were negative were simply observed. The trial, published in the New England Journal of Medicine, reported two-year recurrence-free survival of 93.5 per cent in the ctDNA-guided group versus 92.4 per cent in the standard care group—meeting the statistical threshold for noninferiority. But the ctDNA-guided approach reduced chemotherapy use nearly in half, from twenty-eight per cent of patients down to fifteen per cent. Fewer patients received toxic drugs, and outcomes were just as good.

Perhaps the most futuristic-sounding example comes from melanoma. The KEYNOTE-942 trial tested an individualised neoantigen therapy—a personalised mRNA cancer vaccine, designated mRNA-4157 or V940, developed by Moderna and Merck—in patients with high-risk melanoma that had been surgically removed. Here is how it works: after the tumour is sequenced, a computer algorithm identifies up to thirty-four unique mutations (called neoantigens) specific to that patient’s cancer. An mRNA vaccine encoding those neoantigens is then manufactured and injected, training the patient’s own immune system to hunt down and destroy any remaining cancer cells carrying those specific mutations. In the trial, published in The Lancet in 2024, combining this personalised vaccine with the immunotherapy drug pembrolizumab reduced the risk of cancer recurrence or death by roughly forty-four per cent compared to pembrolizumab alone. The hazard ratio was 0.561—a substantial improvement. This is the first randomised evidence that a computationally designed, individually manufactured cancer vaccine can improve clinical outcomes.

The Hidden Information in Medical Images: Radiomics

There is one more data layer worth understanding: radiomics. Every cancer patient undergoes medical imaging—CT scans, MRIs, PET scans. To the human eye, these images show the tumour’s size, shape, and location. But buried in the pixel-level data of these images is a wealth of quantitative information that is invisible to the naked eye: texture patterns, intensity distributions, spatial relationships between regions of the tumour. These features, extracted computationally, can reveal information about the tumour’s internal heterogeneity, its blood supply, and even its likely genetic profile.

AI models trained on radiomic features can make predictions about treatment response without requiring a biopsy at all—a particularly valuable capability when a tumour is in a location that makes biopsy difficult or dangerous. Consider a tumour deep inside the brain or wrapped around a major blood vessel: taking a tissue sample could be life-threatening. But a non-invasive CT or MRI scan, when analysed by a trained AI model, can extract hundreds of quantitative features that serve as a surrogate for the molecular information a biopsy would have provided.

When radiomic data is integrated with genomic, transcriptomic, and proteomic data—a fusion sometimes called radio-genomics—the resulting models can achieve prediction accuracies that exceed any single data type alone. This is the central promise of multi-omic AI: not that any one data layer is sufficient, but that the integration of multiple layers captures the full biological reality of the disease.

The Black Box Problem: Can We Trust What the AI Says?

There is a question that haunts every application of AI in medicine: can the doctor understand why the model made a particular recommendation? A deep neural network that processes thousands of variables across multiple omic layers and produces the output “prescribe Drug X” is, in computational terms, a black box. The internal reasoning—the precise chain of mathematical operations that led to that recommendation—is opaque, even to the engineers who built the system.

This is not a theoretical concern. Clinicians need to be able to interrogate a recommendation before acting on it. If a model suggests an unusual therapy, the oncologist needs to understand the biological rationale, assess whether it makes sense for this specific patient, and explain the reasoning to the patient and their family. A recommendation without an explanation is a recommendation that cannot be trusted.

Researchers have developed a set of tools to address this challenge. SHAP, introduced at a major AI conference in 2017, provides a mathematically grounded way to measure how much each input feature contributed to a model’s output. If a model recommends immunotherapy, SHAP can reveal that the recommendation was driven primarily by high tumour mutational burden and elevated expression of the PD-L1 gene, for example. Integrated Gradients, introduced the same year, uses a different mathematical approach to assign attribution scores to input features in deep neural networks. LIME, published in 2016, takes a model-agnostic approach: it creates a simplified, interpretable model that approximates the behaviour of the complex model in the neighbourhood of a specific prediction.

These tools do not solve the interpretability problem completely, but they represent a meaningful step toward making AI-driven oncology recommendations auditable and trustworthy. Regulatory bodies are increasingly attentive to this issue. In January 2026, the U.S. Food and Drug Administration finalised updated guidance on Clinical Decision Support Software, clarifying when AI-based clinical decision tools require regulatory oversight as medical devices and when they are exempt—a distinction that hinges in part on whether the software allows the clinician to independently review the basis for the recommendation.

Where We Are Now—and the Honest Limitations

It would be irresponsible to tell this story without being candid about where the limitations lie. The AI models described in this article—the transformers, the graph neural networks, the multi-omic fusion architectures—are, with very few exceptions, research-stage technologies. They have been validated on retrospective datasets, meaning they have been tested on historical patient data to see whether their predictions match what actually happened. Many have shown impressive results in these settings. But the vast majority have not yet been tested in prospective clinical trials—the gold standard in medicine—where predictions are made before treatment and then verified against real outcomes. Reporting standards are tightening: frameworks called TRIPOD+AI and PROBAST+AI have been published specifically to improve how AI prediction models are reported and assessed for risk of bias.

The gap between a promising research model and a clinically deployed, regulatory-approved tool is enormous. It involves not just demonstrating accuracy, but proving safety, reliability, fairness across demographic groups, and interpretability.

There are also serious data equity challenges. The genomic datasets used to train these models are disproportionately drawn from patients of European ancestry at major academic medical centres in wealthy countries. This underrepresentation is not a hypothetical concern—it is documented across the field’s foundational resources. The Cancer Genome Atlas, the single most important public cancer genomics dataset, has known diversity gaps. AACR Project GENIE, a major real-world cancer genomics registry encompassing over one hundred thousand cases, has explicitly documented that many of its participating centres perform tumour-only sequencing—which can confound the distinction between inherited genetic variants and cancer-specific mutations—and that its patient population does not proportionally represent racial and ethnic minority groups relative to disease burden. Because genomic alteration frequencies and molecular signatures can differ by ancestry, models trained predominantly on one population may miss critical patterns in another.

Furthermore, cytotoxic chemotherapy—the “old way”—has not been abandoned. For the majority of cancer types, it remains the backbone of treatment. Targeted therapies are available for a growing but still limited number of molecular targets. Many cancers harbour no known actionable mutation at all, or develop resistance to targeted therapies over time. The vision of precision oncology is not to replace chemotherapy overnight, but to steadily expand the toolkit of targeted options while using AI to match the right patient to the right treatment at the right time.

Despite these caveats, the trajectory is unmistakable. The cost of genomic sequencing continues to fall. FDA approvals for comprehensive genomic profiling panels continue to expand. Major national initiatives—including the Clinical Proteomic Tumour Analysis Consortium, which systematically integrates mass spectrometry-based proteomics with genomic data—are building the large, multi-omic datasets needed to train the next generation of models. And each year brings new publications demonstrating that multi-omic AI integration outperforms single-layer analysis in predicting treatment response, survival outcomes, and resistance mechanisms.

Guarding the Data: Privacy, Consent, and International Governance

Precision oncology runs on data, and data about your genome is among the most sensitive information that exists. It can reveal not only your cancer vulnerabilities but your predispositions to other diseases, your ancestry, and your family relationships. Handling this information responsibly is not just an ethical imperative—it is a legal one.

In the United States, the National Institutes of Health Genomic Data Sharing Policy, effective since January 25, 2015, sets expectations for how genomic research data should be responsibly shared. Public datasets like The Cancer Genome Atlas operate under a two-tier system: some data is open access, available to any researcher, while more sensitive information requires controlled access through a formal application process. Internationally, the Global Alliance for Genomics and Health—a consortium of over five hundred organisations—has established a framework of foundational principles for responsible sharing of genomic and health-related data, grounded in human rights, privacy, non-discrimination, and procedural fairness. In the European Union, genetic data is classified as a sensitive “special category” under the General Data Protection Regulation, imposing additional constraints on how it can be processed.

In Canada, Health Canada regulates medical devices—including genomic testing platforms—through the Medical Devices Active Licence Listing, the official database for all licensed Class II through IV devices offered for sale in the country. The regulatory landscape for AI-based medical software is evolving rapidly worldwide, with the EU’s AI Act introducing new obligations that complement existing medical device regulations.

The Future: From Reactive to Proactive

The ultimate aspiration of AI-driven precision oncology is not just to match patients to existing drugs more effectively—it is to anticipate what the cancer will do next. Tumours evolve. They accumulate new mutations, activate alternative signalling pathways, and develop resistance to therapies that initially seemed to be working. Today, resistance is typically detected only after it has already manifested clinically—the tumour starts growing again on a follow-up scan. By then, the cancer may have changed so much that the next treatment choice is essentially a guess.

Multi-omic AI models have the potential to change this. By analysing the full molecular landscape of a tumour at diagnosis, these models could predict which resistance pathways are most likely to be activated and recommend combination therapies designed to block those escape routes before the cancer has a chance to use them. Imagine a graph neural network that analyses the tumour’s protein interaction network and identifies not only the primary oncogenic driver, but also the three most likely alternative pathways the cancer could reroute through if the primary driver is blocked. The oncologist could then prescribe a combination of drugs targeting all four nodes simultaneously, closing off the escape routes before the tumour even knows they exist.

This is the shift from reactive medicine—waiting for the cancer to outsmart the drug and then scrambling for an alternative—to proactive medicine, where the treatment strategy is designed from the outset to stay one step ahead. It is a vision that depends entirely on the kind of integrated, multi-layered analysis that only AI can provide at the necessary scale and speed.

Meanwhile, individualised neoantigen therapies like the mRNA-4157 vaccine tested in KEYNOTE-942 are moving through larger registrational trials. If phase 3 results confirm the phase 2b findings, computationally designed personalised cancer vaccines could become a guideline-supported treatment modality across multiple cancer types—a development that would have been unthinkable a decade ago.

We are not there yet. But the foundational technologies—the sequencing platforms, the multi-omic profiling pipelines, the transformer and graph neural network architectures, the growing clinical datasets, the regulatory frameworks—are all converging. The question is no longer whether AI will transform cancer treatment. It is how quickly.

Behind The Story

This article was produced through a multi-stage fact-checking process involving two distinct source documents—a technical analysis of AI and multi-omics in precision oncology, and a comprehensive executive summary of AI-powered precision medicine in oncology—both verified claim by claim against primary sources including FDA databases, peer-reviewed journal publications, official trial registries, and government policy documents. Every substantive factual claim in this article has been verified to be accurate.

The first source document contained one clear factual error and one misleadingly framed statistic. The error involved the conflation of two distinct Tempus commercial products—xT CDx (a DNA-only, FDA-approved companion diagnostic) and xR (a separate RNA sequencing laboratory-developed test)—into a single platform. The original claimed that xT CDx performs both DNA sequencing of 648 genes and whole-transcriptome RNA sequencing. In reality, these are distinct assays with different regulatory statuses. This article corrects this error. The misleading statistic was a claim that combining DNA and RNA sequencing identifies twenty-one per cent more patients with driver fusions “eligible for FDA-approved targeted therapies.” While the twenty-one per cent figure traces to a real peer-reviewed study, the improvement drops to sixteen per cent within FDA-approved treatment indications specifically, and the study was authored by Tempus employees. This article presents the range with appropriate context.

The second source document was substantially more accurate. Of twenty-three specific claims verified against primary sources, twenty-one were fully confirmed. Only two issues were identified. First, the document stated the NILE study reported median plasma genotyping turnaround time of approximately seven days; the published figure is nine days. Second, its description of Caris MI Cancer Seek omitted the FDA-approved ERBB2 copy number amplification indication for breast cancer. All other claims—every clinical trial, every FDA PMA number and approval date, every gene panel count, every hazard ratio, every survival percentage, every policy effective date, every computational tool—checked out against original sources.

The clinical trial data presented in this article draws from five landmark randomised trials and one major prospective study: TAILORx (New England Journal of Medicine, 2018), RxPONDER (New England Journal of Medicine, 2021), MINDACT (New England Journal of Medicine, 2016, with Lancet Oncology long-term update), the DYNAMIC trial (New England Journal of Medicine, 2022), KEYNOTE-942 (The Lancet, 2024), and the NILE study (Clinical Cancer Research, 2019). Each trial was verified against its primary publication and, where available, its ClinicalTrials.gov registration.

All FDA approval details—FoundationOne CDx (PMA P170019, November 2017), Guardant360 CDx (PMA P200010, August 2020), FoundationOne Liquid CDx (PMA P190032, August 2020), Tempus xT CDx (PMA P210011, April 2023), TruSight Oncology Comprehensive (PMA P230011, August 2024), and MI Cancer Seek (PMA P240010, late 2024)—were verified against the FDA’s Premarket Approval database and associated Summary of Safety and Effectiveness Data documents.

The computational tools and frameworks described—NeuSomatic, DeepSomatic, Similarity Network Fusion, MOFA/MOFA+, JIVE, iCluster, SHAP, Integrated Gradients, LIME, TRIPOD+AI, PROBAST+AI, the Multimodal Co-Attention Transformer, SurvPath, DeePathNet, GraphSynergy, and DriverOmicsNet—were each verified as real, published tools with documented results in peer-reviewed literature. All remain research-stage technologies unless explicitly noted otherwise.

The description of the Warburg effect was checked against the foundational 2009 analysis in Science by Vander Heiden, Cantley, and Thompson. The article reflects the modern understanding that aerobic glycolysis in cancer cells is a metabolic adaptation supporting biosynthetic demands, not a consequence of mitochondrial dysfunction as Warburg originally proposed.

Data governance claims were verified against primary policy documents: the NIH Genomic Data Sharing Policy (effective January 25, 2015), the Global Alliance for Genomics and Health framework, Health Canada’s Medical Devices Active Licence Listing, the FDA’s January 2026 Clinical Decision Support Software guidance, and the EU’s AI Act. AACR Project GENIE’s documentation regarding tumour-only sequencing at participating centres and demographic underrepresentation in genomic datasets was verified against the consortium’s published data guide and peer-reviewed literature.

The goal of this article was not to produce advocacy for a technology, but to produce an honest, detailed, and accessible account of one of the most important transformations in modern medicine—grounded entirely in verifiable facts.

— END —

The Media Glen Publishing • Synexmedia.com

Original games, tools, books, music, and journalism. Free. No ads. No tracking.