NEW YORK – A multiomic, machine learning-driven method for biomarker and therapeutic target discovery could lower the barrier to precision medicine access in resource-constrained settings by identifying minimal marker sets for a variety of contexts.
The method, published recently in Nature Cancer, describes a precision oncology platform called Molecular Twin that combines multi-analyte molecular analyses of a person and their tumor with machine learning models.
Dan Theodorescu, director of the Cedars-Sinai Cancer Center and senior author of the study, used the platform in this proof-of-principle study to develop a multiomic assay to predict disease survival in pancreatic ductal adenocarcinoma (PDAC), a highly lethal cancer where relatively few advances have been made, resulting in a high unmet need for better diagnostics, prognostics, and therapeutics.
The idea, Theodorescu said, was to determine whether the platform could be used to create minimalistic marker panels that are economically viable to societies that don't have a lot of funds.
"The dream is basically to democratize precision medicine with this system," he said.
Molecular Twin builds disease models via molecular profiling from any combination of omic analyses performed on paired tumor and plasma specimens.
The Molecular Twin platform is powered by proprietary machine learning algorithms developed by Bay Area-based digital medicine startup Betteromics.
Betteromics CEO Angela Lai explained that many current computational approaches for multiomic analysis analyze each type of analyte such as RNA and proteins separately and only combine the results at the end, which is suboptimal.
"Scientists … would have one team that works on protein data, [and] one team that works on the transcriptomics data," she said, which results in reduced dimension single-omics datasets that may not combine well because they lose information on characteristics from one dataset that dynamically interact with those from another.
"RNA and the protein are related," she said by way of example. "So if you don't analyze them together, you would end up potentially cutting off signals that are significant when [they] have interplay with other omics datasets."
Instead of looking at one "omic" at a time, Lai said that Betteromics built a clustering algorithm to search for characteristics that span all the omic modalities that represent different outcomes for patients in the study dataset.
"We then developed another algorithm to find the smallest number of dimensions that retains the cluster, thereby shining the light on the most significant subset of omics characteristics that affect patient outcomes," she said.
This parsimonious methodology is what Theodorescu relied on to develop the most stripped-down marker panel possible in his study.
In that pilot study, retrospective tumor and plasma samples were taken from 74 patients with surgically resected PDAC at stage I or stage II. Molecular profiling was performed via targeted next-generation DNA sequencing (NGS), full-transcriptome RNA sequencing, paired (tumor and normal from the same patient) tissue proteomics, unpaired (tumor from patients and normal unrelated controls) plasma proteomics, lipidomics, and computational pathology.
The Molecular Twin machine learning algorithms identified 6,363 clinical and multiomic features predictive of disease survival, which were later pared down to 589 multiomic features.
The findings were validated in patient data collected from a variety of other institutions, including Johns Hopkins University, Massachusetts General Hospital, and The Cancer Genome Atlas consortium.
A key finding of the analysis was that proteins found in plasma outperform serum carbohydrate antigen 19-9 (CA 19-9) in predicting survival. CA 19-9 is currently the only US Food and Drug Administration-approved biomarker widely used for diagnostic management and preoperative prognostication of PDAC, although it is known for having a high false-positive rate due to other pathologic conditions and can result in false negatives in roughly 10 percent of the population.
"This demonstrates that potentially, you could predict who has pretty much zero chance to be benefited by surgery," Theodorescu said, adding that such patients may be ideal candidates for clinical trials.
Theodorescu hopes to eventually develop a set of preoperative survival biomarkers for PDAC.
"We want to basically have only markers that are available before a surgical decision is made," he said. "The reason is because removal of the pancreas is really morbid."
A study from 2021, for example, found that 4.4 percent of patients died within 90 days of pancreatic resection. Other studies have found similar mortality rates.
"This study represents a meaningful breakthrough in pancreatic cancer research," said Nelson Dusetti, director of research at INSERM, the French National Institute of Health and Medical Research. "By outperforming the standard test and demonstrating the viability of the Molecular Twin platform, the study opens new horizons for improving treatment outcomes and expanding precision medicine availability. The potential impact extends beyond pancreatic cancer, offering hope for advancing cancer treatment on a global scale."
Dusetti said that the study's strength lies in its proof of concept but emphasized that while the identification of predictive signatures based on plasma proteins is noteworthy, the true challenge in PDAC lies in predicting treatment response rather than solely focusing on patient outcomes.
"In my opinion, prognostic markers have limited utility in pancreatic cancer due to its typically poor prognosis," he said. "Moreover, it is imperative to shift our focus towards discovering new targets and proposing innovative treatments. "
Michael Goggins, a professor of pathology, medicine, and oncology at Johns Hopkins Medical School, agreed that the Molecular Twin study is an important step forward while potentially limited by its focus on prognosis.
"The study is novel in examining many hundreds of potential markers across [multiple biomarker] types," he said, while adding as a caveat that there isn’t a clinical test ready for use to predict prognosis based on the study.
"And tests that predict prognosis," he said, "are not that useful unless they help predict response to therapy."
Theodorescu pointed out that while his study's primary focus was identifying prognostic markers, the Molecular Twin platform can be applied to therapeutic target discovery.
A copy number variation analysis performed during the molecular profiling, for example, identified the genes FOXQ1 and KDM5D as top predictors associated with survival in PDAC and as potential therapeutic targets.
While the Molecular Twin platform appears to mark an important advance in its respective field, multiple other research groups are pursuing computationally driven, multiomic approaches to profiling PDAC, as well as other cancers.
Researchers from the Cancer Research Center of Marseille in France, for example, developed a regulatory network inference-based PDAC model that integrated transcriptional networks, epigenomic states, and metabolomics pathways to identify metabolic phenotypes and epigenetic profiles underlying the disease's heterogeneity.
Similarly, researchers affiliated with the Beijing Advanced Innovation Center for Genomics published research in which they analyzed the methylome, chromatin accessibility, and transcriptome of individual cells from PDAC patient tumors to identify a set of biomarker genes whose expression levels correlated with a better prognosis.
Cedars-Sinai filed the intellectual property related to the Molecular Twin platform as the principal agent, with Betteromics listed as an author and inventor. Cedars-Sinai plans to commercialize it in the future, although it is too early at present for firm plans on how that will happen.
"Cedars-Sinai and Betteromics are collaborators in developing this platform and IP, and we do not yet have concrete plans on how to commercialize it together," Betteromics' Lai said. "However, the algorithms developed for this study are already available as reference algorithms on the Betteromics platform for our customers."
Cedars-Sinai also collaborates with Tempus on the Molecular Twin project. The Chicago-based informatics company performed genomic and transcriptomic analyses as well as H&E slide digitization on patient samples.
Theodorescu commented that the Molecular Twin platform must still undergo a number of further studies. He is currently writing up a second paper based on the cohort from the current study that will include computational radiology along with computational pathology as one of the analytes added to the model.
Additionally, Theodorescu's lab has begun to add microbiome data to its Molecular Twin PDAC models.
"We're doing it on the on the patients that were part of the first study," he said, "and reanalyzing [them] with [the] microbiome as an extra variable."
Theodorescu envisions prospective studies of the Molecular Twin platform in the near future.
"We've now collected 2,000 patients [who have] consented for the Molecular Twin across all cancer types here at Cedars," he said.