Skip to main content
Premium Trial:

Request an Annual Quote

Caris Shares Data for AI-Driven Tumor Origin Prediction, Hinting at Diagnostic, Patient Benefit


NEW YORK – Caris Life Sciences has detailed the development and validation of its AI-driven tumor origin predictor, MI GPSai, which the firm has been offering clinically since December 2019 to help direct treatment for patients with carcinoma of unknown primary (metastatic cancer that has no clear origin) and, in certain cases, to inform physicians of possible cases of non-CUP misdiagnosis.

Published last week in Translational Oncology, the study described the training of MI GPSai on genomic data from nearly 60,000 cancer cases and its validation in another 19,555 tumors. Across this validation set, the classifier predicted the correct cancer type with an accuracy of over 94 percent in 93 percent of cases. When the second highest predicted origin was also considered this accuracy increased to 97 percent.

The results — which also included data on the predictor's performance and clinical impact in a cohort of commercially tested patients — suggest MI GPSai has the potential to overcome limitations seen with previous tumor type classifiers, the study authors wrote. Moreover, because of its integration of origin prediction with comprehensive tumor genomic profiling in a single test, Caris' approach could potentially lead to even better improvements than origin-agnostic precision medicine efforts now being explored for CUP, they added.

Genomic technologies have provided significant new hope in the challenge of treating CUP, which occurs in 3 to 5 percent of patients and is marked by dismal outcomes when treated empirically, with median overall survival less than one year.

Molecular predictors of tumor origin have proposed to address this by providing information on tumor origin that can better inform treatment decisions. At the same time the emergence of precision oncology has enabled new strategies to treat CUP based purely on its molecular features, regardless of origin.

But recent trial results have suggested that CUP patients failed to benefit significantly when older tumor-origin classifiers were used on their own to direct treatment. Meanwhile, results from origin-agnostic targeted therapy trials, like the Roche-sponsored CUPISCO study, are still pending.

Caris hopes that its AI approach, by combining both context-specific lineage information and genomic profiling, can not only do better for CUP patients but also provide a quality-control backstop for definitively diagnosed cancer patients, who in rare instances may have their tumor mischaracterized by standard laboratory analyses.

David Spetzler, Caris' president and chief scientific officer, said that the company's broad goal with tools like MI GPSai is to support an emerging paradigm shift in the field of oncology, whereby even accurate diagnoses based on location and morphology are only one factor among many that inform therapeutic decision-making.

Classification and subclassifications of tumors based on their genomic characteristics and their molecular lineage is playing a larger and larger role, with new insights now emerging from epigenetic characteristics, the interaction between a tumor and its microenvironment, and the functioning of the human immune system. For Caris, the goal is to develop algorithmic approaches that can integrate as many of these factors as possible.

"I think what we're going to see is this type of work begin to stratify and characterize patient subpopulations based upon their underlying biological similarity. And that's so informative in understanding how to treat patients differently," Spetzler said. "It's going to start to evolve the way that we diagnose patients and how we characterize them, and I think that's going to be really, really important to advance our understanding of cancer."

The recently published study included both a retrospective validation and data from the use of MI GPSai in both CUP and non-CUP clinical cases profiled by Caris.

Launched a little more than a year ago, Caris' MI GPSai isn't a standalone test, but is essentially an interpretation that can be applied to the company's established genomic and transcriptomic cancer sequencing service. As such, any oncologist who orders that sequencing can get an MI GPSai classification result if they request it.

According to Spetzler, these requests are now being made for virtually all CUP patients. "It's been almost immediate and instantaneous global adoption for those types of cases," he said. Almost no non-CUP cases opt in, which he said makes sense because doctors who have a diagnosis in hand already would have little reason to request a tumor origin or lineage prediction for their patient.

Despite this, the company does calculate origin predictions in the background for every case it processes and if an MI GPSai classification conflicts with a patient's histological diagnosis and meets a high-enough confidence threshold, the company contacts the ordering physician to tell them they might have gotten something wrong.

In the new study the authors reported on 46 cases in which Caris alerted pathologists to discrepancies between submitted diagnosis and MI GPSai prediction, based on the predictions exceeding a score threshold of 0.999. These alerts resulted in a change in diagnosis in 19 cases.

"Considering that the rate of inaccurate diagnosis ranges between 3 percent and 9 percent, inclusion of MI GPSai in clinical routine could improve diagnostic fidelity overall," the authors wrote.

"We're finding these every day," Spetzler added. "Right now, it's happening about 3 percent of the time … but it comes down to what the threshold is that [we set]."

"Right now, we have that set at 97 percent and as we're collecting data to find out how often we are right … we may drop that."

Spetzler said that Caris is finding that if a change in diagnosis would not result in a change in treatment, physicians don't bother to change the diagnosis. The authors reported that another common reason for not changing a diagnosis was a lack of assays to test and confirm a GPSai prediction.

As a hint of the potential clinical impact of this process, the group highlighted one case where a diagnosis change resulted in the patient's treatment course being altered. The company received a sample from a 61-year-old man for molecular profiling for whom the referring pathologist had assigned a diagnosis of poorly differentiated squamous cell carcinoma, but who had not responded well to squamous cell carcinoma-directed therapy.

The MI GPSai predicted urothelial carcinoma, which was confirmed by additional immunohistochemical workup, prompting new PD-L1 testing with protocols specific to the label indications for the immunotherapy atezolizumab. "This PD-L1 score was positive, the patient therapy changed, and the patient response will be tracked in our prospective study," the authors wrote.

For the 1,292 CUP cases in the clinically tested cohort, MI GPSai rendered a high-accuracy prediction — a score above the company's reportable cutoff — for 71.7 percent.

The investigators detailed one case, an 82-year-old man whose standard workup had not identified a primary tumor. MI GPSai predicted a prostate adenocarcinoma and review of the gene expression data showed high expression of androgen receptor, with follow-up IHC confirming. The backbone sequencing data identified pathogenic variants in BRCA2 and PTEN.

According to the authors, the patient had a follow-up biopsy of the prostate which confirmed prostatic adenocarcinoma and, after discussion with the ordering physician, the diagnosis was changed from CUP to metastatic prostatic adenocarcinoma.

In earlier discussions of its GPSai approach, Caris has reported that it can identify an origin for 100 percent of CUP patients and Spetzler said that's still the case, but the overall prediction accuracy drops when lower-confidence calls are included.

"What we landed on is that we really want to have right around 93, 94 percent accuracy [for clinical reporting]," he said.

In the study, the authors cautioned that a current limitation of MI GPSai is that it hasn't yet been trained on all tumor types, with current outliers including non-uterine sarcoma and hematologic cancers. But as data accumulates from the company's routine molecular profiling additional categories for these cancer types can be incorporated into the classifier, which Spetzler said is intended to be plastic and ever-evolving.

Other signals that provide information about tumor origin and lineage could also improve future versions of the algorithm, the authors wrote. These include things like digital pathology image analysis and also epigenetic information, which, in emerging data from companies developing non-invasive cancer screening technologies, has shown the ability to predict the location of a tumor with high accuracy.

Caris is currently collecting prospective data on the overall accuracy of MI GPSai and on resulting clinical outcomes for both CUP patients and those with diagnosis-revisions.

"I think in probably another year or so, we should have enough cases with enough longitudinal data to really start to see the impact," Spetzler said. "There are kind of two ways to think about a study like that. One is just, how often does it agree … but that treatment decision component is arguably the best source of confirmation that we got it right, and [for that] we've got to wait for progression or lack of progression events to really start to understand how well the patient did or didn't do."