Friday, July 4, 2025

Large-scale Validation of LBX DNA-RNA Tumor Test from CARIS (Abraham et al.)

 A few weeks ago, we saw a large-scale validation of transcriptome and genome tumor testing by Yudina et al., Boston Gene.  Blog here.

Now, we have a large-scale validation of RNA-DNA tumor testing from Caris, in Nature Communications as Abraham et al.    Find it here, open access.  The Caris Assure test is plasma-based.   

Validation of an AI-enabled exome/transcriptome liquid biopsy platform for early detection, MRD, disease monitoring, and therapy selection for solid tumors.  Abraham et al.

###

I attach the abstract below.

###

Effective clinical management of patients with cancer requires highly accurate diagnosis, precise therapy selection, and highly sensitive monitoring of disease burden. Caris Assure is a multifunctional blood-based assay that couples whole exome and whole transcriptome sequencing on plasma and leukocytes with advanced machine learning techniques to satisfy all three clinical testing needs on one platform. 

Caris Assure for therapy selection was CLIA validated using 1,910 samples. 376,197 tissue profiles along with 7,061 paired blood and tissue profiles were used to engineer features for three machine learning models. The MCED model was trained on 1,013 patients and validated on an independent set of 2,675 patients. 

The tissue of origin for MCED model was trained on 1,166 samples and validated using 5-fold cross validation. 

The MRD & Monitoring model was trained on 3,439 patients and validated on two independent sets of 86 patients for MRD and 101 patients for monitoring. 

For early detection, sensitivities for stages I-IV cancers (n = 284, 129, 90, 23 respectively) were 83.1%, 86.0%, 84.4%, and 95.7%, all at 99.6% specificity (n = 2149). The diagnostic first-line procedure for tissue of origin was determined for 8 categories with a top-3 accuracy of 85% for stage I and II cancers. Detection of driver mutations for therapy selection from blood collected within 30 days of matched tumor tissue, demonstrated high concordance (PPA of 93.8%, PPV of 96.8%) using CHIP subtraction. For MRD and recurrence monitoring, the disease-free survival of patients whose cancers were predicted to have an event was significantly shorter than those predicted not to have an event using a tumor naïve approach (HR = 33.4, p < 0.005, HR = 4.39, p = 0.008, respectively). 

The data presented here demonstrate a unified liquid biopsy platform that uses blood-based whole-exome and transcriptome sequencing coupled with artificial intelligence to address the important clinical needs in multi-cancer early detection, monitoring of MRD and recurrent cancers, and precision selection of molecularly targeted therapies.

###

###
###

AI CORNER
###

Summary.

Here is a detailed review of the 2025 Nature Communications paper titled:

"Validation of an AI-enabled exome/transcriptome liquid biopsy platform for early detection, MRD, disease monitoring, and therapy selection for solid tumors"
by Abraham et al. (Caris Life Sciences) — for MD/PhD readers who are experts in genomics and laboratory diagnostics.

Executive Summary

This study presents Caris Assure, an ambitious blood-based platform that integrates whole-exome sequencing (WES) and whole-transcriptome sequencing (WTS) of both cell-free DNA/RNA and leukocyte DNA. Coupled with a proprietary multi-omic machine learning algorithm (ABCDai), the assay supports applications in:

  • Multi-cancer early detection (MCED)

  • Tissue-of-origin prediction

  • Minimal residual disease (MRD) detection

  • Recurrence monitoring

  • Therapy selection through blood-based CGP

The authors report impressive analytical and clinical validation across a large and diverse sample set (>370,000 tissue and 7,000 plasma-tissue pairs), making this one of the most expansive liquid biopsy validation studies to date.

Scientific and Technical Strengths

1. Broad Molecular Scope

  • Caris Assure captures WES and WTS data from both cfTNA (cfDNA + cfRNA) and buffy coat DNA, enabling:

    • SNV, indel, CNV, MSI, and fusion detection.

    • CHIP subtraction from buffy coat to reduce false positives.

    • Integrated transcriptomic features.

  • Custom hybrid capture targets 720 cancer-relevant genes deeply, plus >20,000 genes at lower coverage.

  • Uses NovaSeq 6000 for high-throughput sequencing, with sophisticated extraction chemistry and bioinformatics to handle both DNA and RNA.

2. Multi-pillar Machine Learning Architecture (ABCDai)

  • Caris engineers nine molecular "pillars":

    • Canonical genomic features: Mutationome, Copyome, Fusionome, Transcriptome.

    • Liquid biopsy-specific features: Fragmentome, Motifome, Positionome (NU & TF), Entropyome.

  • Uses XGBoost-based gradient-boosted decision trees and a pan-omic feature selection strategy.

  • These features capture not only somatic mutation burden but also nucleosomal positioning, transcriptional start-site fragment patterns, and motif entropy—some of the most cutting-edge concepts in cfDNA biology.

Key Validation Results

A. Therapy Selection (CGP)

  • PPA of 93.8%, PPV of 96.8% vs. matched tumor tissue in metastatic cancer (blood drawn within 30 days).

  • CHIP subtraction shown to be critical: >50% of patients had CHIP-associated mutations in genes like KRAS, ATM, CHEK2.

B. Multi-Cancer Early Detection (ABCDai-MCED)

  • Stage I–IV sensitivity: 83.1% (I), 86.0% (II), 84.4% (III), 95.7% (IV) at 99.6% specificity.

  • Demonstrated superior sensitivity vs. GRAIL's MCED test in early-stage gastric, uterus, kidney, breast, and prostate cancers.

C. Diagnostic Pathway Prediction (ABCDai-GPS)

  • Trained on 660 early-stage cancer patients and 506 normals.

  • Top-3 accuracy of 85% for selecting the optimal diagnostic imaging/test modality (e.g., mammogram, colonoscopy, PSA).

D. MRD Detection and Monitoring (ABCDai-M&M)

  • Tumor-naïve approach—no need for baseline tumor sequencing.

  • MRD cohort: HR = 33.4 (p < 0.005); 89% sensitivity, 98% NPV, 76% specificity.

  • Monitoring cohort: HR = 4.39 (p = 0.008); 45% sensitivity, 92.6% NPV, 83% specificity.

  • Average lead time vs. imaging = 261 days.

Innovative Concepts and Contributions

  1. CHIP-aware Plasma Genotyping
    This is arguably one of the most carefully designed CHIP filtering pipelines in a commercial setting—essential for improving cfDNA test specificity and avoiding false treatment signals.

  2. Use of Plasma Transcriptome Data (cfRNA)
    While rare in the liquid biopsy space due to RNA degradation, Caris shows that plasma WTS can yield meaningful transcriptomic features for detection and classification.

  3. Panomic Feature Integration
    The authors unify expression (WTS), nucleosome positioning, motif analysis, and fragmentomic signatures into one robust AI platform—an integrative approach unmatched in commercial assays today.

  4. Tumor-Naïve MRD
    This circumvents logistical burdens (tumor tissue access, turnaround time), potentially enabling broader adoption.

Limitations and Open Questions

  • Sample sizes for certain cancer subtypes and stages are small, which limits granularity for rare cancers.

  • The model's reliance on proprietary features and non-public algorithmic code raises reproducibility questions, even though code is available upon legal request.

  • RNA-based liquid biopsy features remain biologically compelling but technically vulnerable to variation—though Caris reports RNA extraction optimization, real-world performance remains to be seen.

  • For MRD, comparative head-to-head studies vs. tumor-informed methods (e.g., Signatera) are pending, though the paper states these are planned.

Clinical Implications for Lab Directors and Genomics Experts

  • Caris Assure represents a paradigm shift toward a single-platform solution across early detection, CGP, and MRD—challenging the market silo of GRAIL, Guardant, Natera, and Foundation.

  • The platform may allow laboratories to consolidate test ordering, logistics, and reporting, streamlining decision-making and reducing redundancies.

  • The tumor-naïve MRD approach could reduce turnaround time and tissue acquisition burden, making it attractive to community oncology settings.

  • CLIA validation and >1,900 samples tested across assays indicate high readiness for clinical adoption—though CMS and MolDx coverage pathways remain a key unknown.

Conclusion

Abraham et al. present a landmark study validating Caris Assure, a unified exome + transcriptome liquid biopsy platform empowered by sophisticated AI. The assay’s breadth, sensitivity, specificity, and clinical versatility are notable. While prospective studies will be essential for regulatory acceptance and clinical guidelines integration, this work clearly positions Caris Assure among the most advanced multi-purpose liquid biopsy systems currently in development or clinical use.

Let me know if you'd like a slide deck summary, side-by-side product comparison (e.g., with GRAIL or Natera), or Medicare coverage pathway assessment.


###

###

Here is an AI_generated reimbursement assessment.   

THIS IS PROVIDED FOR CURIOUSITY ONLY AND SHOULD NOT BE TAKEN LITERALLY.


##

Here is a Medicare coverage pathway assessment for the Caris Assure liquid biopsy platform, tailored for genomics experts and lab directors evaluating strategy under CMS and MolDx frameworks:

🔍 Medicare Coverage Pathway Assessment for Caris Assure

I. Summary of Caris Assure’s Clinical Functions

Caris Assure is a multifunctional WES/WTS blood-based assay that supports:

  1. Therapy selection in metastatic disease (via CGP from cfDNA/RNA)

  2. Multi-cancer early detection (MCED)

  3. Tumor-naïve MRD detection and recurrence monitoring

  4. Tissue-of-origin prediction (diagnostic aid)

Each of these clinical use cases faces separate coverage hurdles under Medicare, with different lead MACs, LCDs, and regulatory precedent.

II. Coverage Pathways by Clinical Use Case

1. Therapy Selection via Plasma CGP

Goal: Coverage under the existing MolDx LCDs for blood-based comprehensive genomic profiling (e.g., LCD L37810 or successor policies).

  • ✅ Pathway Exists: Caris can pursue coverage through Palmetto/MolDx under the existing cfDNA CGP policies, which require:

    • Validated detection of guideline-relevant biomarkers (e.g., EGFR, ALK, BRAF).

    • Concordance data vs. tissue (which Caris reports: PPA 93.8%, PPV 96.8%).

    • CHIP correction is a favorable feature, mitigating false positives.

  • 🔍 Considerations:

    • Must publish or submit clinical utility data showing treatment decisions changed and improved outcomes.

    • MolDx prefers inclusion in NCCN guidelines or demonstration that the test identifies FDA-approved therapy indications.

2. MRD Detection (Tumor-Naïve)

Goal: Coverage for MRD surveillance following curative-intent treatment (e.g., in stage II/III colorectal, breast, or lung cancers).

  • ⚠️ Pathway Exists but Challenging: MolDx’s existing LCD for MRD (L38779) supports coverage of tumor-informed platforms like Signatera or RaDaR, especially in colorectal cancer.

  • ❗ Caris Assure uses a tumor-naïve approach, which:

    • Bypasses logistical delays of tumor sequencing (a strength).

    • But lacks direct precedent in current LCD language.

  • ✅ Supportive Evidence in Paper:

    • HR = 33.4 for MRD recurrence, with 98% NPV.

    • Concordance with events and lead time advantage (+261 days).

  • 🧩 Coverage Path:

    • Submit dossier for LCD expansion or reconsideration via MolDx.

    • Support inclusion in CIRCULATE-US–type trials or joint publications.

    • Engage clinical KOLs in oncology to advocate for tumor-naïve MRD as an efficient alternative.

3. Recurrence Monitoring (Longitudinal Surveillance)

  • ⚠️ Limited Pathway: Medicare rarely covers indefinite recurrence monitoring unless tied to:

    • Specific disease states with expected recurrence (e.g., high-risk colorectal cancer).

    • Actionable downstream changes in management.

  • 📈 Caris data supports clinical utility (HR = 4.39, p = 0.008), but:

    • Needs longer-term outcomes and real-world treatment change documentation.

    • Could be packaged as a "surveillance bundle" with MRD under the same policy.

4. Multi-Cancer Early Detection (MCED)

Goal: Screening use in asymptomatic patients or high-risk cohorts.

  • ⛔ No Current Pathway for MCED Screening Under Medicare

    • USPSTF has not endorsed any MCED test.

    • Statutory CMS rules prohibit coverage of screening tests without USPSTF A/B rating under SSA §1861(ddd).

    • No existing LCD or NCD provides a pathway unless under Coverage with Evidence Development (CED).

  • 🧠 Options for Caris:

    • Collaborate with NCI or CMS on a prospective MCED study under CED.

    • Pursue inclusion in programs like PREVENTABLE or PATHFINDER-style cohorts.

    • Pilot diagnostic aid framing, e.g., "rule-in" for symptomatic patients with nonspecific findings.

III. Regulatory and Reimbursement Milestones Required

Use Case Coverage Pathway Caris Status Next Steps
CGP – Therapy MolDx LCD (L37810) Strong Submit concordance + clinical utility
MRD – Tumor-Naïve LCD (L38779 – CRC) Moderate Reconsideration with tumor-naïve data
Recurrence Monitoring Not clearly defined Weak/Exploratory Package with MRD evidence
MCED – Screening Not eligible for FFS Medicare None Consider CED or commercial plans

IV. Coverage Strategy Recommendations

  1. Near-Term Focus:

    • Submit for MolDx coverage of CGP and MRD (target colorectal and breast first).

    • Emphasize tumor-naïve efficiencies, especially in community settings lacking FFPE access.

  2. Mid-Term Focus:

    • Develop comparative trials vs. tumor-informed platforms.

    • Build real-world data registry aligned with CMS's "totality of evidence" approach.

  3. Long-Term Vision:

    • Seek commercial insurer partnerships for MCED pilots.

    • Join NIH or CMS-sponsored MCED umbrella projects for eventual CED petition.

V. Conclusion

Caris Assure has a robust technical foundation and meets multiple high-priority clinical needs. Its CGP and MRD use cases are ready for Medicare engagement, while MCED requires long-term strategic planning aligned with USPSTF and CMS evidence frameworks. A sequenced approach—starting with CGP and MRD under MolDx and building toward broader utility—is the optimal path to widespread reimbursement.

Let me know if you'd like a PowerPoint version, payer value proposition sheet, or LCD comparison grid next.


###

ALSO FOR YOUR CURIOUSITY ONLY, HERE IS AN AI GENERATED "OP ED"

###

Editorial Commentary: Toward a Unified Liquid Biopsy Platform for Cancer Care
By the Editors of Nature Communications

The study by Abraham et al., published in this issue of Nature Communications, introduces an ambitious and technically impressive liquid biopsy platform—Caris Assure—that brings together whole exome and transcriptome sequencing with a sophisticated machine learning framework (ABCDai) to address a broad range of clinical applications in oncology. From early cancer detection to therapy selection and recurrence monitoring, this single blood-based assay aspires to support the full continuum of precision cancer care.

That ambition alone warrants attention. But what sets this work apart—and makes it timely—is the integration of multi-omic liquid biopsy signals with machine learning at population scale. The authors trained and validated their models on over 376,000 tumor samples and 7,000 matched plasma-tissue pairs, a sample size unmatched by most published studies in the field.

Bringing Breadth and Depth to Liquid Biopsy

Caris Assure is more than another cfDNA panel. The assay combines:

  • High-depth WES and WTS on cfDNA and cfRNA,

  • Matched buffy coat sequencing for CHIP subtraction,

  • A machine learning platform built from nine separate molecular “pillars,” including novel fragmentomic and epigenomic surrogates.

This results in a panomic diagnostic signal—one that appears to outperform existing commercial benchmarks in early-stage cancer sensitivity, particularly in breast, gastric, and prostate cancer. The authors report >83% sensitivity for stage I cancers at 99.6% specificity, addressing the Achilles’ heel of many MCED platforms: early-stage detection with acceptable false positive rates.

The same architecture powers a tumor-naïve MRD tool that stratifies relapse risk with hazard ratios as high as HR = 33.4, enabling long lead times over conventional imaging and offering a compelling alternative to tumor-informed strategies.

From Concept to Clinic: Strengths and Challenges

From an editorial perspective, the paper’s strengths lie in its scope, rigor, and transparency:

  • The authors provide detailed methodological appendices, including feature engineering approaches and validation cohorts.

  • CHIP subtraction is treated not as an afterthought, but as a core analytic safeguard.

  • The test is already CLIA validated, enhancing translational relevance.

However, challenges remain. While the use of cfRNA is innovative, its reliability in real-world clinical workflows—where pre-analytical variables abound—will require further study. Likewise, while the multi-feature machine learning model shows superior performance, its interpretability and reproducibility in external settings are open questions. Readers should note that the algorithms are currently proprietary and available only upon request.

Moreover, multi-cancer early detection (MCED) remains a regulatory frontier. Without U.S. Preventive Services Task Force endorsement, MCED applications face barriers to clinical reimbursement. This means that Caris Assure’s path to widespread adoption may rest more securely—at least initially—on its use in therapy selection and MRD surveillance, particularly where tumor access is limited or tissue-based testing is infeasible.

A Platform for the Future

What this paper ultimately represents is a strategic vision for oncology diagnostics: a unified, scalable, blood-based testing system that can inform clinical decisions across the cancer care pathway. Whether the clinical community is ready for a single test that screens, stratifies, and guides treatment remains to be seen—but Caris Assure provides the technological groundwork for such a paradigm.

As liquid biopsy platforms evolve, studies like this one remind us that assay versatility, biological signal integration, and clinical scalability will be the benchmarks of next-generation precision medicine.

We commend the authors for their bold and comprehensive work and look forward to future data—particularly from prospective trials—that may define how such technologies change the standard of care.

—The Editors
Nature Communications