September 5, 2024 | It now appears possible to develop a series of blood-based tests looking for the “sparse” protein signatures of 67 pathologically diverse diseases, which would enable the early detection of high-risk individuals in primary care settings who could then be closely monitored for development of those conditions. These newly identified protein panels are ready to be developed into clinical-grade assays that might additionally be used as “dynamic risk indicators” in clinical trials or to predict patient response to specific therapies, according to Julia Carrasco-Zanini-Sanchez, Ph.D., postdoctoral research assistant in computational genomics and multiomics at Queen Mary University of London.
The predictive performance of the protein panels, which each include between five and 20 proteins, proved superior to that of standardized tests (e.g., blood cell counts, cholesterol, kidney function, and glycated hemoglobin, or HbA1c, tests) and basic clinical information, as reported recently in Nature Medicine (DOI: 10.1038/s41591-024-03142-z). The improvement in prediction for multiple myeloma—even a decade before overt disease development—was “simply astonishing, beyond anything we’ve seen reported in the literature for any disease before,” says Carrasco-Zanini-Sanchez, who conducted the research while working at the University of Cambridge as a research student with GlaxoSmithKline.
Using machine learning, researchers produced sparse (i.e., 20 and under proteins) prediction models for the 10-year incidence of 218 common and rare diseases using integrated measurements of more than 3,000 plasma proteins with linked clinical information on 41,931 randomly selected people from the UK Biobank. When pitted against 67 disease-specific models developed using basic clinical information alone and another 52 developed using a combination of clinical information and data from 37 clinical assays, the sparse models came out on top.
Among the 52 diseases are multiple myeloma, non-Hodgkin lymphoma, motor neuron disease, pulmonary fibrosis, and dilated cardiomyopathy. Notably, single-cell sequencing from the bone marrow of newly diagnosed patients indicated that four of the five proteins on the predictive panel were specifically expressed in plasma cells.
Moreover, the sparse protein models were replicated externally in the European Prospective Investigation into Cancer (EPIC)-Norfolk study that enrolled 25,639 middle-aged individuals from the general population of a county in Eastern England. For prediction of the six diseases tested—primary malignancy-prostate, heart failure, chronic kidney disease, type 2 diabetes, COPD, and cataract—the models were shown to generalize “very well,” says Carrasco-Zanini-Sanchez.
“The main challenge we face with some of these tests is that they work in terms of relative quantification,” she adds. “We don’t know exactly what [a protein level] means in terms of absolute concentration,” which is true of all the tests currently implemented in clinical practice. To do that will require further work to determine precisely the reference ranges of the proteins in a healthy population.
But the stage has been set to “hopefully in the not-far-away future be able to improve early detection of [67] different diseases” with standardized assays for predicting them, says Carrasco-Zanini-Sanchez. It is well appreciated that the early detection of cancer significantly improves the prognosis, but this is equally true for many other conditions.
Timelier diagnosis of multiple conditions has huge clinical implications. “Currently, healthcare systems in general are a little bit reactive in terms of intervening when a patient already presents with symptoms,” Carrasco-Zanini-Sanchez says.
The central issue is the lack of objective biomarker tests, or at least ones that aren’t complicated and time-consuming, she continues. Proteomic signatures of disease can be revealed simply by taking a blood sample.
The study team found that higher levels of the proteins TNFRSF17 and TNFRSF13B were strongly predictive, respectively, of multiple myeloma and MGUS (monoclonal gammopathy of undetermined significance), a non-cancerous condition that can lead to the blood cancer, reports Carrasco-Zanini-Sanchez. Decreased plasma TNFSF13B was also shown to be predictive of a higher risk for multiple myeloma, in support of earlier treatment of a subset of patients with anti-TNFRSF17 agents already approved for refractory disease—a list that includes antibody–drug conjugates, T cell engagers bispecific antibodies, and cellular therapy with chimeric antigen receptor T cells.
Several clinical trials looking at therapeutic timing of anti-TNFRSF17 agents have already started providing evidence of their safety and effectiveness in early lines of treatment for multiple myeloma, she adds. One of the latest of these was published last year in Blood Cancer Discovery (DOI: 10.1158/2643-3230.BCD-22-0074).
Protein predictors could theoretically play a role in future clinical trials, to address the lengthy wait for primary or even secondary outcomes to happen, says Carrasco-Zanini-Sanchez. One proposal is that the protein signatures serve as a dynamic estimator of how an individual’s risk changes in response to a therapeutic intervention, which would also enable the efficacy of the predictive model to be evaluated.
“There is certainly also an interest in using this kind of framework to predict... who is at high risk of a poor prognosis such as mortality or adverse effects,” she adds. Similarly, the sparse prediction models might help identify groups of patients who will respond better to specific drugs.
Carrasco-Zanini-Sanchez grew up in Mexico, where her undergraduate work was dedicated primarily to gestational diabetes. It’s what sparked her interest in prediction, she says. “We would receive women in the clinic in their first trimester and try to predict who of those women were going to develop gestational diabetes between their 24th and 28th week of pregnancy.”
While working on her Ph.D., she landed in the group led by Claudia Langenberg, Ph.D., director of the Precision Healthcare University Research Institute at Queen Mary University of London and a visiting scientist at the epidemiology unit of the University of Cambridge. An initial focus on epidemiology quickly pivoted to working on the proteomic signatures, starting with a proof-of-concept study on impaired glucose tolerance, otherwise known as prediabetes (Nature Medicine, DOI: 10.1038/s41591-022-02055-z), which afflicts around 7% of the adult population worldwide.
The condition can only be detected by an oral glucose tolerance test, which takes over two hours to perform and is more costly than a simple blood test, she continues. For these reasons, the test is not routinely performed in clinical practice and many people remain undiagnosed for years until they present with some of the later consequences of diabetes.
“We were extremely successful in developing the protein predictor for this condition specifically,” says Carrasco-Zanini-Sanchez, and “it set us off to be able to propose this at a large scale across many different diseases systematically and to then be able to pinpoint the most promising clinical specialties in which this sort of strategy might be helpful.”
‘Disease-agnostic Approach’
Next steps include validating study results in a more diverse patient cohort, says Carrasco-Zanini-Sanchez, noting that the UK Biobank represents a mostly European population. “It is very important to test the flexibility of these protein tests across different ethnicities to make sure we are not amplifying health disparities.”
This might be accomplished with a couple of cohorts—the All of Us Research Program of the National Institutes of Health in the U.S. and Our Future Health, its ambitious U.K. equivalent—both of which have proteomic projects underway. Even standard clinical tests are known to work differently based on a person’s ethnicity.
Some of the calculators used to evaluate candidates for kidney transplants rely on metrics that may not be as generalizable across ethnicities, she offers as an example. “People of African ancestry tend to have a disadvantage with the standard risk model that is used in clinical practice.” Similarly, genetic variants in certain ethnic groups affect the lifespan of red blood cells that in turn affect levels of HbA1c used to diagnose diabetes and prediabetes.
The validation exercise across ethnicities needs to be done in a hospital, says Carrasco-Zanini-Sanchez, adding that Queen Mary University of London has close ties to clinicians with Barts Health NHS Trust that runs five hospitals throughout the City of London and East London. But further technical validation of the assays, and the quantification work, will need to happen before larger scale testing can begin.
Limiting the predictive signature to no more than 20 proteins reflects the desire to bring a new generation of practical and affordable blood tests to busy primary care settings, she says. The full platforms, targeting thousands of proteins, are all quite expensive.
The research team has adopted a “disease-agnostic approach” to the risk prediction models, says Carrasco-Zanini-Sanchez. Their hope is that disease specialists will now step up to help develop tests clinically useful for their area of focus.