In March, Caris released top-line results of its ACHIEVE study, testing its MCED test in real cases. Press release here. Active Linked In author Alex Dickinson wrote a set of 5 articles about the results. One, two, three, four, five.
Out of curiousity, I asked what Chat GPT could make out of the six documents.
AI CORNER
###
Overview
Caris reports striking interim performance for its Detect MCED assay using deep whole-genome sequencing, with unexpectedly strong early-stage sensitivity in common cancers. However, enriched cohorts, limited follow-up, and incomplete blinded validation constrain interpretation. Dickinson’s analyses highlight a differentiated WGS multi-signal strategy with potential advantages over methylation-first approaches.
Consolidated Article (Caris + Dickinson)
Focusing first on the press release, the key point is that Caris reported an interim analysis, not a completed prospective screening validation. The Achieve 1 dataset includes 2,122 subjects (1,505 undiagnosed; 617 cancers), but the undiagnosed group is enriched, not general-population screening.
Only 22.5% had ~1-year follow-up, with ~7% later diagnosed with cancer—again indicating high-risk enrichment. About 865 samples remain in blinded validation, so current results are signal-generating, not definitive.
The reported performance is notable. Stage-specific sensitivity was 56.8% (I), 70.1% (II), 77.1% (III), 99.1% (IV), with 61.3% for stage I–II. Early-stage sensitivity in key cancers included 53% breast, 78.9% prostate, 86.7% lung, and 62.2% colorectal. Specificity was 99.1% in a small asymptomatic subset (n=121) and 95.3% in the broader undiagnosed cohort. These are the central empirical results.
An Expert Discusses The Data
Dickinson’s posts provide useful context. He frames Caris as entering MCED from a position of scale and infrastructure—large tumor databases, clinical profiling, and sequencing capacity—suggesting Detect is an extension of an existing oncology data platform rather than a stand-alone assay.
Scientifically, Dickinson highlights the assay design: ~250x whole-genome sequencing of plasma with paired buffy coat sequencing to remove CHIP, extracting mutational, fragmentomic, and nucleosome-positioning signals for ML classification. This multi-signal WGS framework plausibly explains the observed sensitivity pattern.
His most provocative point concerns tumor-type performance, especially breast and prostate. He argues methylation-first MCED approaches may underperform in these high-incidence cancers, making Caris’s relatively strong early-stage sensitivity clinically meaningful. This reframes MCED evaluation toward incidence-weighted performance, not aggregate sensitivity alone.
Key Limitations
However, key limitations remain. The 99.1% specificity estimate rests on only 121 asymptomatic subjects. The broader 95.3% specificity is more relevant operationally but less reassuring for screening. The cohort is enriched, follow-up incomplete, and blinded validation pending—raising the risk of optimistic interim estimates.
Caris also claims superiority over methylation approaches, but provides no head-to-head comparison, making this an interpretive rather than evidentiary claim. Cross-study comparisons, including Dickinson’s, remain hypothesis-generating.
Strategically, Dickinson emphasizes Caris’s ability to pursue a sequencing-intensive design due to its scale. He notes current costs may be high (e.g., overlapping 2x150 reads), but could fall with alternative platforms. This underscores that economic feasibility will be as important as analytical performance.
Finally, Caris signals future expansion to whole transcriptome integration, reinforcing its view of MCED as a multimodal inference problem, though at the cost of added complexity.
Bottom line: Detect represents a biologically coherent and potentially important MCED approach with strong interim early-stage signals. But given cohort design, small screening subsets, and pending validation, the appropriate stance is serious interest with disciplined caution, not definitive conclusion.
