A team of computational scientists at Lund University (Sweden) has built a deep co-learning proteomic model to improve diagnostic accuracy for a small number of dementia-related conditions that remain challenging in primary care settings due to a lack of useful biomarkers. Predicting these different but related diseases with a single blood test would facilitate rapid and confident differential diagnosis, according to postdoctoral researcher and model creator Lijun An, Ph.D.
Editor’s Note: This article was originally published on Diagnostics World. Visit the original article here.
For six related and often coexisting conditions – Alzheimer’s disease, Parkinson’s disease, amyotrophic lateral sclerosis, frontotemporal dementia, previous stroke, and healthy aging controls – individuals have a 50-50 chance of being misdiagnosed or underdiagnosed by their general practitioner today, he points out. This is largely due to shared biological features and symptoms, but also to the great diversity in patient groups and the way functional domains are measured from one clinic to another.
The newly innovative paradigm, called Proteomics-Based Artificial Intelligence for Dementia Diagnosis (ProtAIDe-Dx), is the first step toward a more accurate diagnostic approach using the largest neuroproteomics samples to date that will enable early intervention and tailored treatments for neurodegenerative diseases that are rapidly growing around the world. Its potential was the subject of a study recently published in Natural medicine (Digital ID: 10.1038/s41591-026-04303-y).
The modeling exercise exploited the world’s largest neurodegenerative disease plasma proteomics dataset, compiled by the Global Neurodegenerative Proteomics Consortium (GNPC). A subset of more than 17,000 memory clinic patients was selected for the study, based on the availability of SomaLogic 7K proteins, sampled across 19 contributing sites.
The SomaLogic platform identifies and measures thousands of proteins, “functional effectors” in the bodies of living people, notes Jakob Vogel, Ph.D., Anne’s supervisor and an assistant professor who leads a research group as part of the government-backed MultiPark translational program at Lund University. Applying the most advanced AI models to a large proteome set has shown that ProtAIDe-Dx can significantly improve biomarker-based differential diagnosis by identifying proteins that drive decisions at the patient level.
The field was not devoid of progress. Over the past few years, researchers have discovered blood-based biomarkers for Alzheimer’s disease (for example, p-tau217 and amyloid ratios) that are expected to soon improve diagnostic accuracy of this most common cause of dementia, An says.
This is not the case with other neurodegenerative diseases where a final diagnosis is often only made at autopsy. Patients are usually diagnosed while alive based on clinical history, brief cognitive screening tools, and various laboratory tests to rule out reversible causes.
A simple blood test that can sort out molecular clues is the hope, Ahn says. A starting point to address the problems of misdiagnosis, confounded by age-related comorbidities, is the development of blood-based biomarkers to identify underlying neurodegenerative diseases with high specificity.
Blood samples from diverse groups from the United States and Europe were analyzed using SomaLogic’s proteomics assay performed on Illumina high-throughput DNA sequencing platforms. From an initial set of 7,595 proteins, several hundred most relevant proteins were selected for a multi-task co-learning approach to allow the ProtAIDe-Dx model to indicate common pathology.
ProtAIDe-Dx has outperformed several machine learning models and state-of-the-art deep learning models in terms of predictive accuracy. When generalized to multiple independent data sets, it also produced a better differential diagnosis compared to currently accessible clinical biomarkers, reports An.
The deep learning architecture essentially learned one task (say, is it Alzheimer’s?) to help learn another task (say, is it Parkinson’s?), all at the same time, Vogel explains. “Learning them together gives you a net advantage on each individual task, because deep within the data there is information relevant to all (six) tasks.”
After a few thousand iterations, the model internally determines the “weight” assigned to the informational proteins, An explains. Feature importance techniques were also used to classify, filter and identify the most predictive variables while reducing noise.
The multiple diagnosis prediction model has been tested across dozens of different data sets, providing an accurate estimate of how it would perform in a real-world clinic, Vogel says. Ultimately, understanding how a model behaves “in reality” is more important than finding the best models based on computational techniques.
Vogel says it was somewhat surprising that protein profiles predicted cognitive decline better than a clinical diagnosis. The GNPC data set “reflects real life… (where) everyone diagnoses themselves a little differently.” Not only are doctors likely to reach different conclusions, but levels of confidence in those diagnoses are related to the availability of sophisticated technology such as positron emission tomography, magnetic resonance imaging, and cerebrospinal fluid analysis.
“Overall across all these different data sets, we don’t know how accurate the diagnostic classifications are,” he says. “It’s very difficult to diagnose neurodegenerative disease without biomarkers, and we don’t have accurate biomarkers for most of them.”
Another interesting finding of the study is that individuals with the same clinical diagnosis appear to have different biological subtypes. The way the model groups patients “won’t always match exactly what doctors have identified as their diagnosis, but it tells us something… (namely), that there are people who may have a similar biological trait but who may have symptoms that present differently.”
It is known that family members can have the same genetic mutation, with one person developing amyotrophic lateral sclerosis and another person developing frontotemporal dementia, for example. “The question becomes what information is relevant – the symptoms or the biological profile.”
To choose a treatment, doctors may want to know the patient’s basic biological profile. But if it’s behavioral therapy, they may also want to know more about the symptoms. These are “disparate entities, and both are relevant to our pursuit of treatment and management of patients,” Vogel says.
In an ideal world, Vogel imagines, multiple diseases would be diagnosed via a blood-based panel. “We’re not there yet, at least with this iteration of plasma proteins.”
While the ProtAIDe-Dx model is highly correlated with multiple aspects of many different diseases, it is not at the point where it can be used alone to make a definitive diagnosis. But adding the model’s results to information from Alzheimer’s disease biomarkers, imaging scans, and cognitive tests would greatly improve the situation.
“If you use very stringent thresholds with this proteomic data, you can also come up with fairly reliable biological diagnostics,” he adds. The problem is that the diagnosis may be uncertain for all but 10% to 20% of individuals.
What’s missing is data, Vogel says, comparing the decades-long process of developing a blood test for Alzheimer’s disease, which has been progressing in fits and starts. SomaLogic’s affinity technology provides only one view of the protein. Other perspectives are available from proximity extension assays (eg, Olink Proteomics), which uses next-generation sequencing to identify and quantify proteins, and mass spectrometry, which offers a much broader and more detailed array of peptides, fragments, and post-translational modifications. Brain-derived extracellular vesicles present in the blood are also promising non-invasive biomarkers for early detection of neurodegenerative diseases.
“The AI methods we use are great because they can take the needle out of the haystack and combine information in ways that someone with a spreadsheet couldn’t do,” he says. “It’s just a matter of continuing to generate data in large quantities and characterize it…until we find what we want.”
At that point, the work will begin to narrow down the candidate proteins to the smallest panel possible, without losing diagnostic value, Vogel says. This would make the test scalable by making it less expensive and easier to reproduce and maintain quality control.
“If you have something that is incredibly accurate, people will use it regardless,” he says. If there is high demand for the test, the price will eventually fall as the supply side of the equation adjusts. The goal will be to obtain maximum clinical benefit with minimum overhead.
A single blood test to diagnose multiple neurodegenerative diseases is the long-term goal here. “It cannot be understated how important biomarkers are in our pursuit of treatment and management of patients,” Vogel says. “When you have a disease, there is a very powerful relief that comes from having someone say, ‘I know exactly what you have…’ It opens up patients to a world of possibilities that is not available to them if they don’t know.”
For many years, one of the major obstacles in finding treatments for Alzheimer’s disease was that many people enrolled in clinical trials didn’t actually have the condition because there were no biomarkers to diagnose them, he says. “Everything about the pursuit of dementia research and treatment has improved with biomarkers,” including tracking disease progression and response to medications.
On a single-disease basis, biomarkers are increasingly being used in clinical trials to improve confidence in diagnosis. This includes in particular Parkinson’s disease, where measurements of alpha-synuclein in the cerebrospinal fluid are common, Vogel says.
Recently, biomarkers have emerged for 4R tauopathies—a group of neurodegenerative diseases defined by the pathological accumulation of the four-repeat isoform of tau protein in the brain—which often manifests as atypical parkinsonism. Through proteomic analyses, studies have also found proteins strongly associated with progression to cognitive impairment based on damage to synapses.
All of this suggests that biomarker panels could be assembled at some point. “Progress has been excellent in this area over the past couple of years,” says Vogel, who hopes the multidiagnostic test will be ready for market much sooner than the decade or more it once was preparing for.
A major focus of the Dementia Multi-Omics and Neuroimaging Laboratory, where Vogel and his team do their work, is to partner with their colleagues to better understand the heterogeneity of neurodegenerative diseases and how that affects response to treatment, Vogel and his team say. There may also be a role for the ProtAIDe-Dx model in clinical trials being conducted by World-Wide FINGERS, a network of studies based on the landmark Finnish Study of Geriatric Intervention to Prevent Cognitive Impairment and Disability, which shows that concurrent lifestyle changes can improve cognition in at-risk older adults.
It’s easy to underestimate the complexity of the data being analyzed, says An, who learned about clinical proteomics precision medicine projects in Vogel’s lab. He emphasizes that crowded, high-dimensional data require specialized and advanced computational methods to produce interesting results that can be validated to support precision medicine.
“There’s a lot of diversity in what we’re looking at, and when you put it all together, it becomes a very difficult task to make anything out of it,” Vogel says. “(Researchers) all know that if you add more data, you’re more likely to be able to find the truth, but it would be easier… if we had better coordination and standardization. Maybe we just need to work together more and… pool our data (in advance).”
