Revisiting The Cancer Genome Atlas From a Whole-Genome View

Artificial intelligence has the potential to transform cancer research, but its impact depends heavily on the quality, scale, and completeness of the data used to train it. Many oncology datasets were generated during the exome sequencing era, when analysis focused largely on protein-coding genes rather than the full genome, leaving much of the genomic landscape unexplored. Now, as sequencing costs fall and computational tools become more sophisticated, the field is shifting toward whole-genome analysis to achieve a more comprehensive view of tumor biology.

A collaboration between Inocras and the Broad Institute aims to explore the unmapped portions of the genome by reanalyzing more than 8,000 public cancer genomes from The Cancer Genome Atlas (TCGA) using harmonized whole-genome sequencing pipelines. The initiative intends to create one of the largest standardized references of somatic mutations across multiple cancer types, providing a high-quality whole-genome variant dataset that could support both biological discovery and next-generation computational models.

Technology Networks spoke with Jehee Suh, chief executive officer at Inocras, whose background spans biotech leadership, strategy, and commercialization, and Dr. YoungSeok Ju, co-founder of Inocras, director of the Genome Insight Institute, and associate professor at KAIST, whose research focuses on cancer genomics and somatic mutation biology.

They discussed why the shift to whole-genome analysis matters now, how harmonized datasets could strengthen AI-driven discovery, and what the next phase of precision oncology may need.

The perks of pivoting from exome to whole-genome analysis

Your collaboration with the Broad Institute builds on data from The Cancer Genome Atlas, but shifts the focus from exome to whole-genome analysis—what are the key benefits to adopting a genome-wide view?

Genome-wide analysis allows researchers to move beyond the small fraction of DNA captured by exome sequencing and study cancer as a disease shaped by alterations across the genome. Rather than focusing only on protein-coding regions, whole-genome sequencing enables analysis of the broader genomic landscape that can shape how tumors emerge, evolve, and respond to treatment.

“Whole-genome analysis gives researchers access not only to coding mutations, but also to non-coding regulatory alterations, structural variants, copy-number changes, and mutational signatures that can be central to tumor biology yet are often missed by panel- or exome-based approaches,” explained Suh.

For a program such as TCGA, this creates an opportunity to revisit one of cancer research’s most influential datasets through a more comprehensive lens. Researchers can ask new questions about disease mechanisms, biomarker development, and therapeutic vulnerabilities that were difficult to address using exome data alone.

From Inocras’s perspective, the shift is also about scale and usability. “Our public materials consistently emphasize that whole-genome data can provide a more comprehensive picture for research and clinical interpretation, and this collaboration extends that philosophy to one of the field's most important public cancer datasets,” said Suh.

Key benefits of whole-genome analysis:

• Whole-genome sequencing captures coding and non-coding alterations missed by exome-based approaches.
• Reanalysis of TCGA could unlock new insights into biomarkers, tumor biology, and treatment opportunities.
• Richer genomic datasets can improve computational approaches and help to direct future precision oncology strategies.

AI needs consistent, curated data to advance cancer biology

You describe this as an “AI-ready” reference dataset with harmonized variant calls—what specific challenges did you solve in standardizing and curating this data at scale, and how do you see it enabling the next generation of AI models in cancer research and drug discovery?

One of the biggest challenges in large-scale whole-genome analysis is not simply generating variant calls but ensuring they are sufficiently consistent across thousands of tumor-normal pairs to serve as a reliable reference for downstream research. With only a small number of genomes, minor inefficiencies or inaccuracies can often be corrected manually. At the scale of more than 8,000 samples, however, pipelines must be accurate, robust, and fully automated.

In this collaboration, the Broad Institute and Inocras refined their bioinformatics workflows for population-scale analysis, ran the pipelines in parallel, and then harmonized the resulting calls into a single frozen, curated dataset.

Every sample was processed using the same analytical framework. That level of standardization is particularly important because, according to Suh, “AI models are highly sensitive to noise, batch effects, inconsistent calling conventions, and differences in curation standards.” Without those controls, models risk learning technical artifacts rather than meaningful biology.

“What makes this especially powerful for AI is the breadth of signals now available within a single framework: single nucleotide variants, indels, structural variants, copy-number alterations, and non-coding events across a large pan-cancer cohort with clinical annotation,” said Suh.

“A harmonized dataset like this can help train models that are better at identifying driver events, stratifying tumors, linking genomic patterns to therapeutic hypotheses, and benchmarking new computational methods.”

He added that being “AI-ready” means more than simply having large data. The data must also be coherent, reproducible, and suitable for both model development and validation. In practice, this kind of infrastructure gives researchers a stronger foundation for building tools that generate more reliable biological insights—rather than correlations that fail when tested using datasets or in the “real-world”.

Why standardized data matters for AI in cancer research:

• AI models require consistent, high-quality inputs to avoid learning technical noise instead of biological signals.
• More than 8,000 samples were processed through a harmonized analytical framework to improve reliability.
• Genome-wide data types can support models for tumor classification, driver detection, and therapeutic hypothesis generation.
• Reproducible reference datasets are essential for validating AI tools and translating them into research or clinical use.

Whole-genome analysis provides a richer view of tumor evolution

At the AACR Annual Meeting 2026, you presented new insights from TCGA whole-genome sequencing across both the Educational Session and the “TCGA and Beyond” spotlight—what would you highlight as the most important takeaways from these data, and what should the field be paying closest attention to?

“The most important takeaway is that whole-genome analysis materially expands what we can learn from TCGA. This is not just a technical upgrade from exome to genome; it changes the biological resolution of the dataset,” said Suh.

Improving Oral Bioavailability Prediction With Human Models

Join this webinar to explore how an integrated human gut/liver microphysiological system (MPS), combined with mechanistic modeling, provides deeper insight into oral drug disposition.
View Webinar

Advertisement

By moving beyond coding regions, whole-genome analysis brings non-coding driver candidates, structural variation, copy-number architecture, and genome-wide mutational processes into a single analytical framework. Together, these features can provide a more complete picture of how tumors develop and evolve, while helping researchers identify genomic traits that may be relevant for disease classification, prognosis, and therapeutic intervention.

Suh added that the opportunity extends beyond improved biological insight. Equally important is the value of harmonization at the cohort scale. “One of the reasons this effort matters is that it creates a common reference dataset that different groups can analyze, benchmark against, and build AI methods on top of.”

He believes that this shared foundation is essential if findings are to be reproducible, comparable, and extensible across institutions. In that sense, the project highlights two priorities for the field: generating new biological discoveries from whole-genome data, while also building the standardized infrastructure needed to accelerate future cancer intelligence initiatives.

Key insights from the TCGA whole-genome analysis:

• Whole-genome sequencing increases the biological resolution of TCGA beyond “exome-era” analyses.
• Integrated analysis of structural variants, copy-number changes, and non-coding events offers a fuller view of tumor evolution.
• Harmonized reference datasets enable benchmarking, collaboration, and AI model development.
• Future progress depends on both biological discovery and robust data infrastructure.

Detecting alterations in the “dark matter” of cancer genomes

If the non-coding genome is the “dark matter” of cancer, are we at risk of having built our current models—and even treatments—on an incomplete picture, and how does this new whole-genome dataset begin to correct that?

Only roughly 1% of the human genome’s ~3 billion base pairs encodes proteins. The remaining majority of the genome doesn’t code for proteins, but it is far from biologically inactive. Increasing evidence shows that large portions of this “non-coding” space contain regulatory elements that control when, where, and how genes are expressed.

Although this region has historically been described as “dark matter,” Suh notes that the term can be misleading. “Pathogenic mutations disrupting these regions can have tumorigenic effects and may represent novel therapeutic targets,” he explained. In other words, changes outside of protein-coding genes can still drive cancer biology by altering gene regulation and cellular behavior.

Beyond single regulatory elements, many structurally important alterations also occur in non-coding DNA. Large genomic rearrangements can reshape regulatory landscapes, while deep intronic variants can disrupt normal RNA splicing and lead to aberrant gene products. At a broader level, genome-wide mutational signatures—patterns left behind by different DNA damage and repair processes—provide critical insight into how cancers arise and evolve and can help inform therapeutic strategies such as homologous recombination deficiency (HRD)-guided treatment.

Taken together, these features are largely invisible to exome- or panel-based sequencing. Suh emphasizes that only whole-genome sequencing can systematically capture this layer of variation at scale, making it possible to move beyond a protein-centric view of cancer toward a more complete genomic understanding of disease.

How Agentic AI Is Transforming the Modern Lab

This whitepaper examines how agentic AI is reshaping laboratory operations and what this shift means for the future of scientific discovery.
View Whitepaper

Advertisement

Broadening data integration, expanding cohorts, and merging genomics data with medical records

As whole-genome sequencing becomes more scalable and integrated with computational methods, where do you see the field heading over the next 5–10 years—especially in terms of combining large, harmonized datasets like this with AI to uncover new therapeutic targets or biomarkers?

Ju suggests that progress in cancer genomics will depend on expanding datasets across three interconnected areas: cohort scale, clinical linkage, and multi-layered biological integration. Together, these represent not just incremental improvements in size or resolution, but a shift toward building datasets that are sufficiently rich and structured to support both AI-driven discovery and clinical translation.

First, cohort size will need to increase substantially to provide the statistical power required for identifying rare but high-impact drivers of disease. “Cohort sizes must scale by another order of magnitude to provide the statistical power for AI to identify rare, high-impact therapeutic targets,” Ju cautioned. While the current dataset of more than 8,000 genomes represents a major milestone, he noted that it still averages fewer than 300 cases per cancer type across 31 tumor types, limiting the ability to robustly detect less common genomic events and account for broader population diversity.

Second, he emphasized that genomic data must be meaningfully connected to the clinical context. Without integration of medical history, treatment response, and outcomes, the translational value of sequencing datasets remains limited. As Ju put it, “without linkage to clinical records, the utility of these datasets is substantially limited.” This linkage is essential if genomic insights are to move beyond discovery and into real-world clinical decision-making.

Finally, Ju highlighted the importance of integrating multiple data modalities, including multiomics, pathology, and imaging. Bringing these layers together would allow researchers to move beyond single-dimensional genomic interpretation toward a more comprehensive understanding of tumor biology, improving both mechanistic insight and clinical applicability.

Three pillars shaping the next era of cancer genomics:

• Scaling cohort size is essential to detect rare but clinically important genomic drivers.
• Linking genomic data with clinical records is critical for real-world translational impact.
• Multi-modal integration (multiomics, pathology, imaging) will enable a more complete view of tumor biology.

This content includes text that has been created with the assistance of generative AI and has undergone editorial review before publishing. Technology Networks' AI policy can be found here.