The clinical stakes could hardly be higher. Locally advanced rectal cancer has traditionally been treated with neoadjuvant chemoradiotherapy followed by surgery, an approach that saves lives but often exacts a heavy toll, including permanent colostomies, radiation-related bowel dysfunction, urinary problems, and sexual health consequences. Total neoadjuvant therapy (TNT), which delivers chemotherapy and radiation before any operation, has improved outcomes and opened the door to organ preservation for patients whose tumors vanish completely. The TORCH trial took this concept further by adding immunotherapy to the neoadjuvant regimen, creating immunotherapy-based total neoadjuvant therapy, or iTNT, for MSS patients who would typically be excluded from immunotherapy-based strategies because their tumors lack the microsatellite instability that makes them immunotherapy-sensitive.

In the new study, the research team analyzed clinical and genetic data from 63 patients enrolled in the prospectively registered TORCH trial between May 1, 2021, and September 15, 2022. The investigators collected baseline tissue mutations, magnetic resonance tumor regression grade (mrTRG) scores from treatment-time MRI scans, and circulating tumor DNA (ctDNA) positivity measured at multiple points across the treatment course. Their goal was twofold: first, to determine which features measured before, during, or after iTNT best distinguished patients who achieved a complete response from those who did not, and second, to identify early markers of recurrence risk that could guide follow-up intensity and decisions about whether surgery could be safely avoided.

The genetic findings were striking. A baseline mutation in the SMAD4 gene, a well-known tumor suppressor involved in the TGF-beta signaling pathway that helps regulate cell growth and tissue architecture, was significantly associated with poor response to iTNT, with the association reaching statistical significance at P equals 0.007. Patients carrying SMAD4 mutations were markedly less likely to achieve a complete response, suggesting that this single genetic alteration encodes a fundamental resistance to the combined chemo-radiation-immunotherapy approach. The finding aligns with a growing body of evidence that SMAD4 loss reshapes the tumor microenvironment in ways that blunt immune attack and may drive more aggressive, metastasis-prone biology.

Beyond SMAD4, three additional features correlated with complete response rates: post-treatment mrTRG assessed on MRI (P less than 0.0001), tumor mutational burden, or TMB, which quantifies the number of mutations carried by the tumor (P equals 0.03), and the latest ctDNA status, a measure of whether fragments of tumor DNA were still detectable in the bloodstream after treatment (P equals 0.03). TMB is a classic immunotherapy biomarker, since tumors with more mutations tend to produce more neoantigens that the immune system can recognize, but its relevance in MSS rectal cancer treated with combination iTNT had been uncertain. The ctDNA measurement, meanwhile, represents the concept of a liquid biopsy: rather than sampling the tumor itself, clinicians can track fragments of tumor-derived DNA shed into circulation, gaining a real-time view of residual disease that imaging may miss.

The true advance, however, came when the researchers integrated these features into multivariable models. A model combining post-treatment mrTRG, SMAD4 mutation status, TMB, and latest ctDNA status achieved an area under the curve, or AUC, of 0.92 for discriminating complete responders from non-responders, compared with just 0.75 for a model based on mrTRG alone, a statistically significant improvement (P equals 0.0098). An AUC of 0.92 reflects excellent discriminative accuracy, while 0.75 represents only modest performance, so the jump is clinically meaningful. Importantly, the researchers also tested whether prediction could happen earlier in the treatment course. An early integrated model that used mid-iTNT mrTRG along with the genetic features achieved an AUC of 0.89 versus 0.76 for the mrTRG-only model (P equals 0.047), meaning clinicians might one day gauge a patient’s trajectory partway through therapy rather than waiting until the end.

The prognostic analysis added another layer of clinical utility. Using Cox proportional hazards regression and Kaplan-Meier survival analyses, the team found that baseline SMAD4 mutations and ctDNA positivity at later time points were both correlated with an increased risk of disease progression, including local recurrence and distant metastasis. In practical terms, a patient whose blood still shows molecular traces of tumor after treatment completion, or whose original tumor carried a SMAD4 mutation, faces a higher probability that the cancer will return, even if standard assessments look favorable. This kind of molecular residual disease detection is becoming one of the most actively pursued frontiers in oncology, because it can flag recurrence months or even years before radiographic evidence appears, when intervention may be most effective.

The implications for treatment strategy are substantial. For patients with locally advanced rectal cancer, the decision of whether to proceed to surgery or pursue nonoperative management, sometimes called watch and wait, hinges on the confidence with which a complete response can be established. Current assessments rely on a patchwork of digital rectal examination, endoscopy, MRI, and serum markers such as carcinoembryonic antigen, none of which is perfectly reliable. A molecular framework that adds SMAD4 status and serial ctDNA to the mix could give clinicians the confidence to safely spare surgery in true complete responders, while identifying poor responders early enough to intensify their therapy, enroll them in alternative trials, or avoid prolonging an ineffective course of immunotherapy. The study’s authors emphasize that these models are best understood as complementary risk stratification tools rather than standalone decision instruments, augmenting rather than replacing clinical judgment and imaging.

It is worth underscoring what makes the MSS population such an important target. Roughly the overwhelming majority of colorectal cancers are microsatellite stable, and for years, immune checkpoint inhibitors were considered largely futile in this group. The TORCH trial’s demonstration that adding immunotherapy to total neoadjuvant therapy can produce promising complete response rates in MSS patients was itself a paradigm-shifting result, and the current analysis addresses the critical next question: how do we know who is benefiting? Without reliable biomarkers, the field risks either overtreating patients who will never respond or undertreating those who would have responded with more time. The integrated models described here offer a data-driven way to navigate that uncertainty, converting a one-size-fits-all regimen into something closer to adaptive, biology-informed care.

Limitations remain, as they do in any single-trial biomarker analysis of 63 patients. The findings will need validation in larger, independent cohorts, and the logistics of serial ctDNA monitoring, including assay standardization, cost, and the optimal sampling schedule, remain active areas of investigation. The trial was registered as NCT04518280, and the analysis was funded by the National Natural Science Foundation of China and the Beijing Xisike Clinical Oncology Research Foundation. Still, the convergence of tissue genomics, liquid biopsy, and quantitative imaging in a prospectively registered trial represents a template for how biomarker-driven personalization should be developed in rectal cancer. If validated, the combination of SMAD4 mutation status, tumor mutational burden, serial ctDNA tracking, and mrTRG assessment could become a standard part of the decision-making toolkit for MSS locally advanced rectal cancer, helping more patients keep their organs, avoid futile toxicity, and face their disease with a clearer, molecularly grounded picture of what lies ahead.

Subject of Research: Predicting response and prognosis of immunotherapy-based total neoadjuvant therapy in microsatellite stable locally advanced rectal cancer using genetic features and serial circulating tumor DNA

Article Title: Application of genetic features and serial ctDNA in estimating response and prognosis of immunotherapy-based total neoadjuvant therapy (iTNT) for microsatellite stable locally advanced rectal cancer: data from TORCH trial

Article References: Wang, Y., Xu, Y., Lin, Y., Liu, Y., Jing, Q., Shen, L., Wan, J., Wang, Y., Zhang, H., Wu, R., Zhou, S., Chen, Y., Zhang, Z., Wang, J., Sun, Y., Bao, H., Ou, Q., Bao, H., Shao, Y., … Xia, F. (2026). Application of genetic features and serial ctDNA in estimating response and prognosis of immunotherapy-based total neoadjuvant therapy (iTNT) for microsatellite stable locally advanced rectal cancer: data from TORCH trial. Genome Medicine. https://doi.org/10.1186/s13073-026-01755-2

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01755-2

Keywords: rectal cancer, immunotherapy, ctDNA, SMAD4 mutation, tumor mutational burden, total neoadjuvant therapy, microsatellite stable, TORCH trial, liquid biopsy, organ preservation, prognosis, complete response

At the core of this pioneering work lies the concept of "contextualized modeling," a sophisticated family of algorithms designed to tailor gene network analyses to the unique biological makeup of individual tumors. Led by graduate student Caleb Ellington and Professor Eric P. Xing of CMU’s School of Computer Science, the researchers applied these methods to nearly 8,000 tumors spanning 25 distinct cancer types, constructing models that capture the individual complexity and heterogeneity present across patients. The findings, published in the prestigious Proceedings of the National Academy of Sciences, reveal how these personalized gene networks can unearth previously hidden cancer subtypes and enhance the precision of survival predictions, with particular benefit to rare cancers that have historically been understudied.

Traditional biomedical modeling typically depends on segmenting patient populations into broad categories, forming aggregate models that may inadvertently obscure critical biological differences. This aggregation stems from limitations inherent in existing methods, which require large homogeneous patient cohorts to ensure statistical power and model accuracy. Consequently, researchers and clinicians face dilemmas: either include a limited number of stratifying factors, thus glossing over subtleties, or create increasingly granular groups at the risk of generating less reliable models. Ellington articulates this conundrum by highlighting how such practices result in models insensitive to individual variation, impairing their clinical applicability, especially in multifactorial diseases like cancer, Alzheimer’s disease, and diabetes.

Contextualized modeling fundamentally shifts this paradigm by producing individualized gene network models conditioned on each patient’s distinct clinical, genetic, and lifestyle features. This approach not only recognizes but actively leverages the complexity of thousands of potential contextual factors. By learning which elements are most informative for differentiating patient profiles and disease behaviors, the model automatically weeds out irrelevant variables, thus circumventing the contentious debates surrounding patient grouping criteria. This elegant solution yields models with enhanced specificity and predictive power, empowering physicians to tailor treatments based on a holistic view of the patient’s biological context.

Beyond individualization, contextualized models exhibit a unique generative capability that addresses a critical blindspot in traditional approaches: the prediction and understanding of novel or rare disease forms. Whereas conventional models rely on pre-established patient clusters, contextualized models can be synthesized anew to correspond with previously uncharacterized medical contexts. This adaptive versatility was demonstrated when the researchers applied their framework to gene expression data for tumor types not previously encountered in model training. The ability to extrapolate in this manner suggests a new frontier for modeling biological complexity across scales — from molecular interactions to system-level dynamics.

Professor Eric P. Xing emphasizes the deep biological insight this method affords. He explains that biology consists of intricately interconnected systems extending over multiple organizational scales, from molecules to ecosystems. Until now, congruence between these scales had been mostly intuitive and piecemeal. Contextualized modeling offers a rigorous computational framework to probe these nested layers of complexity and evaluate individual variability systematically. This mechanistic clarity feeds into the ongoing development of GenBio AI, an ambitious project aiming to integrate multi-scale simulators that ultimately form an AI-driven digital organism (AIDO). The vision is a simulator capable of mirroring the distinctive biological makeup of each person, capturing not only commonalities but also the idiosyncrasies of individual biology.

A particularly compelling application of this method was the study of thyroid carcinoma, a cancer traditionally associated with relatively favorable outcomes. Due to its high survival rates, thyroid cancer may receive less research attention relative to other malignancies, potentially masking subgroups with more aggressive phenotypes. By deploying contextualized gene network models, the researchers identified a novel thyroid cancer subtype with significantly worse prognosis, a discovery that may pave the way for targeted therapeutic development. This result showcases the power of individualized modeling not merely to stratify risk but to reveal actionable biological insights that could influence clinical management.

However, the significance of this work extends far beyond thyroid carcinoma. The study encompasses 25 cancer types, including notoriously complex malignancies affecting the lung, brain, and stomach, among others. Through this expansive coverage, the analytic framework simultaneously extracts both tumor-specific and pan-cancer biological information, deepening our understanding of oncogenic processes at multiple scales. By examining shared and unique features of individual tumors across diverse cancer types, the approach enables a more nuanced understanding of cancer biology, potentially guiding cross-cutting therapeutic strategies.

To facilitate broader exploration and foster collaborative research, the team has also provided a publicly accessible web tool that allows users to visualize and interrogate the extensive pan-cancer dataset. This resource democratizes access to complex multi-omics data and promotes integrative analyses that can uncover novel patterns and hypotheses. Such open science initiatives accelerate the translation of computational innovations into clinical applications by bridging the gap between data generation and actionable knowledge.

One of the paramount challenges addressed by contextualized modeling is the frequent lack of sufficiently large, uniform datasets in biomedical research. In most experimental settings, increasing sample sizes enhances statistical accuracy. Yet in medicine, expanding sample numbers often entails incorporating heterogeneous patient populations, complicating modeling efforts due to varying disease stages, environmental exposures, and genetic backgrounds. Conventional models struggle to reconcile these confounding factors, resulting in oversimplified or misleading conclusions. Alternatively, contextualized models embrace and leverage this intricacy by explicitly modeling the influence of multiple varied conditions simultaneously, leading to more robust and generalizable predictions.

This modeling approach fundamentally changes the scientific workflow by enabling improvements in prediction accuracy through diversification rather than mere repetition. Instead of conducting repeated measurements under identical conditions, researchers can introduce greater variation in conditions and rely on the model’s capacity to discern relevant signals from noise. This permutes the traditional tradeoff between complexity and accuracy, allowing scientists to exploit the richness of real-world clinical data where patient heterogeneity and incomplete information have historically hindered precise modeling.

The research group demonstrated that contextualized models consistently outperform standard methods across a spectrum of challenging datasets, particularly those characterized by limited or noisy data. This superiority stems from the model’s ability to identify and prioritize contextual factors most critical for outcome prediction, effectively adapting to variance in the data instead of being confounded by it. Such adaptability suggests wide-ranging applicability across biomedical domains beyond oncology, wherever data complexity and nuance pose fundamental barriers to understanding and intervention.

As highlighted by Caleb Ellington, this work heralds a new era in biological modeling—one that transcends the constraints of reductive grouping and embraces the full diversity of biological and environmental inputs. By acknowledging and integrating this complexity, researchers and clinicians can achieve greater fidelity in their models, produce deeper insights, and ultimately improve patient care. The framework not only advances computational methodology but also champions a shift toward truly individualized medicine.

Looking ahead, the Carnegie Mellon team aims to refine these models with the ultimate goal of personalizing therapeutic regimens in real clinical settings. Recognizing the translational potential, they have released a comprehensive toolkit available at contextualized.ml, fostering adoption and further innovation. This transparent and accessible platform positions the scientific community to capitalize on contextualized modeling’s capabilities, accelerating the journey from computational discovery to bedside impact.

Subject of Research: Development of individualized gene network models for cancer using contextualized machine learning methods

Article Title: Learning to estimate sample-specific transcriptional networks for 7,000 tumors

News Publication Date: 23-May-2025

Web References:

• Carnegie Mellon School of Computer Science
• Eric P. Xing’s Homepage
• Proceedings of the National Academy of Sciences Article
• GenBio AI
• Contextualized Modeling Toolkit
• Pan-Cancer Web Tool

References:
10.1073/pnas.2411930122

Image Credits: Carnegie Mellon University

Keywords: Cancer genomics, personalized medicine, contextualized modeling, gene networks, computational biology, oncology, tumor heterogeneity