Multi-omics technologies and artificial intelligence are redefining how cancer begins, evolves, and can be detected earlier. In this plenary talk at the Precision Oncology Summit, Dr. Michael Snyder of Stanford Medicine explores how integrating genomics, transcriptomics, proteomics, metabolomics, spatial biology, and AI is transforming cancer research, prevention, and precision health—using familial adenomatous polyposis as a powerful model for early cancer formation.

Dr. Michael Snyder is a Professor of Genetics at Stanford Medicine, where he served as Chair of Genetics for many years. A pioneering force in precision medicine development, Dr. Snyder has been instrumental in advancing multiomics techniques—studying not just genomics, but all molecules, including genes, RNA, proteins, metabolites, and their three-dimensional interactions in tissues across health and disease states.

This video features a plenary speech at the Precision Oncology Summit, where Dr. Snyder presented groundbreaking research on familial adenomatous polyposis using comprehensive multi-omics approaches, alongside emerging applications of artificial intelligence in genetic medicine and personalized health monitoring.

The transcript below has not been reviewed by the speaker and may contain errors.

Introduction: Multiomics and AI in Genetic Medicine

While advances in genetic medicine could fill days of discussion, this presentation focuses on select topics with strong research foundations and clear clinical ramifications. Two primary areas frame this discussion: familial adenomatous polyposis (FAP) as a model for multi-omics cancer research, and artificial intelligence applications across medicine. These topics represent the cutting edge of how comprehensive molecular profiling and computational approaches are transforming our understanding of disease mechanisms and treatment possibilities.

Understanding Familial Adenomatous Polyposis: A Model for Early Cancer Formation

Familial adenomatous polyposis represents a genetic disease where individuals carry an APC mutation, thought to be the initiating event in most colon cancers. Consequently, these patients face a 100% lifetime cancer risk. They develop hundreds to thousands of polyps, typically leading to complete colon removal. Current treatment includes baby aspirin to mitigate progression, a therapeutic approach whose molecular basis becomes clearer through multi-omics analysis.

This condition provides an exceptional model for understanding the earliest steps of cancer formation. By the time colons are removed, numerous polyps of varying sizes exist—some with dysplasia, some without, larger ones with more dysplasia, and occasionally adenocarcinomas. This spectrum offers a unique window into cancer evolution.

Multi-Omics Research Approach in FAP Studies

A collaborative research effort involving multiple investigators established comprehensive patient consent protocols to obtain tissue samples from FAP patients. Each patient contributed multiple polyps for analysis using an extensive multi-omics approach. Whole genome sequencing or exome sequencing (sometimes both) was performed on all polyps to determine genetic composition. Bulk profiling included RNA sequencing, proteomics, metabolomics, and lipidomics. Additionally, extensive single-cell experiments were conducted, including single-cell RNA sequencing and single-cell ATAC sequencing, which maps regulatory regions throughout the genome. Modern techniques now enable simultaneous single-cell RNA and ATAC analysis in the same experiment, called multiome profiling.

This comprehensive approach across numerous polyps aimed to better understand early cancer formation steps, yielding several significant observations.

Discovery of Polyclonal Origins in Colon Polyps

One of the first major discoveries emerged from deep DNA and exome sequencing of individual polyps, even from the same patient. Initial analysis focused on oncogene mutations present in these polyps, which could originate from different intestinal regions. Results revealed that polyps contained shared mutations, but in a surprising hodgepodge or mix-and-match pattern. For example, certain mutations would pair together in one polyp, but the same mutation would pair with different oncogenic mutations (such as RAS) in another polyp from different colon regions.

The explanation challenges initial assumptions. Rather than polyps being monoclonal—arising from individual cells growing out—the evidence strongly suggests they are polyclonal. This means that early in colon formation, groups of cells (not individual cells) divide, and subsets migrate to different colon regions to form polyps. Different cell groups carry different mutation sets, with some mutations shared and others unique to specific groups.

Mathematical modeling provided a convincing demonstration that not just oncogenic mutations, but the entire mutation spectrum supports this polyclonal origin theory. This finding carries important therapeutic implications. If polyps were monoclonal, a single therapy might potentially eliminate many simultaneously. However, polyclonal origins complicate treatment because polyps can evolve differently with distinct oncogenic mutation sets, requiring more nuanced therapeutic approaches.

Mapping Early Cancer Formation: Stepwise Versus Continuous Progression

Understanding whether early cancer formation occurs stepwise or in a more continuous fashion required multiple experimental approaches. Single-cell methods called multi-M enabled simultaneous mapping of RNA and regulatory regions in the same cells across numerous patient samples.

Samples ranged from normal, unaffected tissue (though importantly, from APC individuals, so not truly normal as subsequent analysis revealed) to polyps of various sizes, larger polyps, and adenomas, with carcinomas added for comparison. Single-cell analysis revealed how cancer cells are similar or different at individual cell levels and how they change across the polyp spectrum.

Cell Type Differentiation Across Cancer Progression

Examining the three major polyp cell types—epithelial, stromal, and immune cells—revealed that even at the earliest stages, certainly by the polyp stage, cells begin differentiating from normal cells. Although representing the same cell types, they show substantial differences in gene expression. Additionally, new cell types emerge, such as cancer-associated fibroblasts, which are unique to cancerous cells. These observations enable the identification of new cell types forming during oncogenesis.

Addressing whether cancer occurs stepwise or continuously, analysis of overall cell expression and regulatory patterns proved revealing. When examining RNA patterns and regulatory patterns across normal tissue (green), polyps (purple), and adenocarcinomas/colorectal cancers (red), polyps appear to progress in a more continuous fashion toward oncogenesis rather than giant discrete leaps. This suggests progression occurs through numerous small steps rather than dramatic transformations.

Systematic Computational Analysis of Regulatory Regions

Detailed mapping through systematic computational analyses of regulatory regions revealed complex activation patterns. Each cell type contains hundreds of thousands of regulatory regions throughout the genome. By mapping when these regions become active (shown in red) versus inactive (blue), patterns emerge that can be clustered. Some regulatory elements activate early and remain on, while others activate quite late. Specific transcription factors associated with these patterns can be identified from sequence motifs.

Gene expression analysis follows very similar patterns, allowing informatics extraction of detailed information. Ultimately, this data combines to create comprehensive regulatory maps tracking progression from the very first steps through the final differentiation of different cell types. For cancer-associated fibroblast cells, for example, all information can be integrated to identify key regulatory factors for each cell type. These cancer-associated fibroblasts appear to have specific key factors sending cells down particular differentiation pathways, with implications for controlling or eliminating these cells.

Spatial Organization of Cells: CODEX and Advanced Imaging Techniques

Understanding cell organization represents a critical challenge because while single-cell analysis reveals what cells are present, it doesn't necessarily show how they're organized. Using a painting analogy: knowing all pixels in a painting doesn't reveal what the painting depicts. Understanding pixel arrangement is essential for comprehending the complete image, much like the Mona Lisa.

CODEX Methodology and Protein Localization

The CODEX method uses antibodies to stain for approximately 50 different proteins, enabling precise visualization of protein localization throughout polyps. The laboratory also employs transcriptomic methods like Xenium, which track several hundred to now thousands of markers. This approach reveals which cell types are present and how they differ or remain similar across different polyps or oncogenesis stages.

Analysis included normal healthy mucosa from individuals without FAP, FAP patient "normal" mucosal layers, polyps of varying sizes, and adenocarcinomas or colorectal cancers. Staining different cell types requires substantial antibody validation work, with this study validating over 40 antibodies to track immune cells, epithelial cells, and all previously mentioned cell types.

Early Cellular Changes in FAP Patients

A striking finding revealed that even at the very earliest polyps showing no dysplasia—and even before that, in mucosal layers—tissue differs substantially from normal healthy intestinal or colon regions. Comparing normal cancer tissue (green) across different cell types to FAP patients' "normal" tissue, polyps at different stages, and adenocarcinomas/colorectal cells demonstrated that numerous cell populations shift. Epithelial cell numbers change, and cancer-associated fibroblasts, rare in normal cells, appear even at the very earliest stages in normal FAP patients. The critical point: "normal" tissue in these patients isn't truly normal.

Tracking Specific Cancer Markers Across Disease Progression

Individual markers can be tracked with precision. For example, a CA antigen—a recognized cancer hallmark—can be detected even at the very earliest stages before cancer initiation, before polyp formation begins. Staining intensity increases with cancer stage progression.

Summary analysis of cell type changes shows stem cells increasing from normal tissue through FAP mucosa to polyps and adenocarcinoma. Goblet cells (secretory cells) decrease. Cancer-associated fibroblasts dramatically increase with cancer. Interestingly, immune cell populations show complex changes, particularly macrophages. T regulatory cells, which regulate immune response, appear very late. By the colorectal cancer stage, substantial evidence indicates T cell exhaustion.

Cellular Neighborhoods, Communities, and Tissue Organization

Cell biologists are deeply interested in understanding cellular organization patterns. Computational approaches can determine individual cell types' propensity for proximity to one another, defining "neighborhoods." Some neighborhood patterns make intuitive sense. Analysis can then examine how neighborhoods cluster together into "communities," which map onto recognizable morphological structures like follicles and other cell types. Further analysis reveals how communities organize into "tissue units."

Cancer progression shows total organizational reorganization, as expected. The most interesting changes involve immune cells, with major shifts in CD8 T cells, microvasculature alterations, and changes in macrophages and other cell types. This reveals how entire organizational structures shift during cancer development.

Multi-Omics Pathway Analysis: The Eicosanoid-Arachidonic Acid Connection

Complete multi-omics analysis—incorporating metabolomics, lipidomics, RNA-seq, and other modalities—provides comprehensive views of pathway changes during cancer progression. While numerous immune system changes occur, one of the most interesting pathways identified was the eicosanoid-arachidonic acid pathway.

This finding provides a molecular explanation for baby aspirin treatment in FAP patients. This pathway activates very early, even in the earliest polyps, suggesting inflammation occurs early in disease progression. This may explain one mechanism through which baby aspirin provides control. While aspirin remains somewhat controversial in the broader field, with studies supporting and questioning its use, evidence suggests it is beneficial in this context.

The bottom line: deep multi-omics analyses provide a detailed understanding of colorectal cancer origins and events occurring across various disease stages, with significant therapeutic implications.

Artificial Intelligence Transforming Healthcare and Research

Working with big data has long been central to this research program, and artificial intelligence is fundamentally transforming data interpretation approaches. AI has become deeply embedded in healthcare across virtually every application area. Beyond patient management benefits, AI proves extremely beneficial in research, diagnostics, treatment selection, and numerous other domains.

Addressing the Genetic Basis of Complex Diseases

Much discussion at precision medicine conferences focuses on Mendelian forms of disease—BRCA1 and BRCA2, ACMG genes, and similar single-gene conditions. However, these genes account for only approximately 15% of cases, even in breast cancer. Expanded gene lists still account for only 30% or fewer cases, leaving substantial missing heritability.

The explanation: most diseases aren't single-gene mutations but complex diseases caused by multiple mutations. This complexity has been largely ignored due to the difficulty in analysis, though polygenic risk scores exist. However, traditional polygenic risk scores work poorly, showing predictive value only for the top 2-4% of individuals. They fail because genetics isn't linear—mutations don't simply add up with one plus one equaling two. Classical genetics demonstrates this non-linearity clearly.

AI-Powered Nonlinear Genetic Analysis

New AI-powered approaches, developed even before AI's current popularity, address this limitation. AI's power lies in its nonlinearity—it can identify gene groups in a nonlinear fashion and detect patterns enriched in disease cases. While details become technically complex, results prove compelling.

Abdominal aortic aneurysm provides an illustrative case. This condition shows 70% heritability, yet is typically discovered when aortas burst, and patients die—a terrible diagnostic method. Better predictive methods are desperately needed. Traditional approaches identified three questionable genes. AI analysis of a relatively small study identified 60 genes and built reasonably effective polygenic risk scores using genetics. Electronic health records can eventually provide decent scores, but genetic approaches enable prediction from birth. Combining multiple data sources improves predictions further.

Other applications demonstrate even greater impact. Amyotrophic lateral sclerosis (ALS), known to be highly heritable, has seven known genes from large GWAS studies. New AI methods identified 690 genes, increasing explained heritability from 6% to 36%. Typically, these approaches find 10 to 100 times more genes and at least quadruple (usually more) the explained heritability. For severe COVID cases, a large study analysis enabled explaining approximately 77% of heritability.

These new methods should enable a better understanding of the genetic disease basis. The hope is that beyond discussing Mendelian forms, the field can incorporate these complex genetic contributions. However, the predictive aspects haven't proven as straightforward as initially hoped.

Wearables and Continuous Health Monitoring

Tracking cancer occurrence and general health through wearables and similar technologies represents another powerful application area. Systems have been established to detect illness from COVID, respiratory viruses, and similar conditions using simple smartwatches. Heart rate increases provide signals that learning systems can interpret.

This represents just the beginning. AI can extract extensive information from smartwatches, which make numerous measurements by shining light into blood, measuring red blood cell counts, and oxygenation from oxygenated hemoglobin. Machine learning and AI enable predictive models for red blood cell count, hemoglobin, and related measures. While not clinical-grade tests, they're sufficient for detecting shifts from individual baselines, enabling health tracking via a simple smartwatch.

Fasting glucose and hemoglobin A1C show decent signals, sufficient for identifying concerning changes. Smartwatches also measure parameters unavailable in clinical settings, such as electrodermal activity or galvanic stress response—skin conductance. This proves powerful because skin becomes dry when diabetic or stressed, becoming wet under other conditions, and algorithms can detect these patterns.

Diabetes Subtyping Through Continuous Glucose Monitoring

Extensive diabetes research utilizes new continuous glucose monitors, which have proven remarkably powerful. This work focuses on dissecting type 2 diabetes, which most people don't realize encompasses not just type 1 and type 2, but effectively type 2A through 2Z and beyond. Numerous type 2 diabetes subforms exist but are largely ignored.

Research has been subtyping type 2 diabetes based on hemoglobin A1C levels across normal (blue), pre-diabetic (yellow), and diabetic (red) individuals. Subtyping considers whether patients show muscle insulin resistance, beta cell defects, incretin effects (the GLP pathways), hepatic resistance, and other factors. Glucose levels reveal distinct subphenotypes distributed across the spectrum. One person might show insulin resistance with normal incretin function, while another displays severe incretin defects—differences not apparent from glucose levels alone.

Personalized Nutrition and Lifestyle Based on Diabetes Subtype

The significance becomes clear through practical application. Continuous glucose monitors, when worn by individuals (including physicians who should monitor their hemoglobin A1C), reveal surprising patterns. Studies suggest 90% of pre-diabetic individuals don't know their status. Wearing these patches genuinely changes eating behavior because users observe what foods spike glucose and what doesn't. Glucose patterns vary dramatically between individuals, particularly in aging populations, where most become "spikers."

Machine learning and AI can determine diabetes subtype from glucose curve shapes with high predictive value, at least for some subtypes. This matters because diabetes type determines which foods cause glucose spikes. Muscle insulin-resistant individuals spike from potatoes and pasta. Beta cell defect individuals spike from potatoes. AI can match diabetes type with appropriate food choices.

Beyond food selection, comprehensive lifestyle programs can be built around these insights, including not just what to eat but when to eat it and how to match it with exercise. This enables truly personalized lifestyle management using AI. (Full disclosure: commercial involvement exists in this area, though academic research continues independently.) When incorporating 54 million foods from databases into personalized eating recommendations, AI becomes essential for breaking down this information. Simple interventions like 15-minute walks after eating spike-inducing foods reliably suppress glucose spikes, enabling personalized management programs.

AI in Medical Imaging and Diagnostics

Retinal imaging has become increasingly popular, with AI now detecting approximately 17 different conditions from retinal scans. Remarkably, about a dozen non-eye conditions—not necessarily eye-associated—can also be diagnosed from retinal scans, including Alzheimer's disease and cardiovascular disease.

The biggest AI impact area remains image analysis. Even though pathologists don't always agree, the best agreement study found two pathologists agreeing 84% of the time on canine glioma diagnoses. AI-powered image analysis and classification prove much more predictive for both detection and cancer phenotype subtyping. Genomic information can now be extracted from images. As of the current count, 1,247 FDA-approved medical tests exist, most incorporating image analysis algorithms.

Whole Body MRI and Longitudinal Monitoring

Strong belief exists in bringing comprehensive information into prevention, which drives most laboratory work through deep data profiling of healthy individuals. One spinoff company performs whole-body MRI, which remains controversial among physicians despite its potential value.

The critical insight: the question isn't whether nodules exist—everyone has nodules. (Personal example: nine nodules across 21 whole-body MRIs over nine years.) The question is whether any nodules are growing. Longitudinal tracking provides this answer. None of the example nodules has grown, but only longitudinal monitoring can establish this. Single MRIs might cause panic, but longitudinal series provide real power.

The company's first 100 patients demonstrated this value, catching early ovarian cancer, early prostate cancer, and even early pancreatic cancer, almost never detected early. Longitudinal tracking proves key to identifying these shifts. Integrating other data types enhances power further, as often multiple markers increase together rather than single indicators.

The Future of AI in Medicine: Beyond Empathy

Eric Topol's book "Deep Medicine" predicted AI would transform everything except empathy, devoting an entire chapter to this theme. One year later, AI demonstrated superior performance to physicians in empathy assessments. The reality: AI will be embedded in all medicine aspects, and appropriately so. It won't replace doctors, at least not for complex decision-making.