The research staff member at the IBM Research-Australia lab discussed how he and colleagues utilized machine learning to identify a set of proteins in blood that can predict the concentration of amyloid-beta in spinal fluid.
Benjamin Goudey, PHD
In a recent study, Benjamin Goudey, PhD, a research staff member at the IBM Research-Australia lab, and colleagues showed that by analyzing as few as 4 proteins, a blood-based signature can predict the status of cerebrospinal fluid (CSF) amyloidβ1-42.
Since direct measurement of CSF biomarkers is too invasive for a screening test, Goudey and colleagues developed a blood-based signature through a machine learning approach that can provide a cheap and minimally invasive estimation of an individual’s CSF amyloid status. This is the first study to show that a machine learning approach using plasma protein levels, age, and APOE4 carrier status can predict CSF Aβ1-42 levels with high accuracy.
To further explore the study and its results, NeurologyLive spoke with Goudey in an interview.
Benjamin Goudey, PhD: Our study explored how machine learning could be used to predict levels of amyloid beta, one of the key proteins associated with Alzheimer disease, in the cerebrospinal fluid from hundreds of proteins and metabolites measured in blood, demographics, and genetics. People had conducted similar studies predicting whether individuals had abnormal amyloid levels measured from PET imaging, rather than spinal fluid. However, recent evidence indicates that changes in amyloid levels in the spinal fluid may be detectable over a decade before a change of amyloid from PET imaging can be detected.
We showed that our model could correctly predict patients as having abnormal amyloid-beta levels in their spinal fluid with an accuracy of 77%. Moreover, we found that similar levels of accuracy could be achieved when using only a handful of proteins. Finally, we examined a set of individuals with mild cognitive impairment who were not used to build the predictive model and found that individuals from this group who were predicted to have abnormal levels of amyloid in their spinal fluid were 2.5 times more likely to receive a diagnosis of Alzheimer compared to individuals predicted to have normal amyloid levels.
These type of models have been explored previously to predict levels of amyloid measured via PET scans, as opposed to in the spinal fluid as we have done here. However, levels of amyloid in the brain (measured by PET) and spinal fluid have different properties, with the 2 measurements representing different aspects of Alzheimer disease pathology and with changes in the spinal fluid detectable up to a decade before those by PET imaging. Hence, it wasn't clear whether applying a similar machine learning approach to spinal fluid would work. The strong predictive performance of our models indicates this is an interesting avenue for further research.
We also found that the inclusion of metabolite levels measured in plasma wasn't helpful for predicting whether an individual had abnormal levels of amyloid in their spinal fluid. This contradicts some of the previous literature and it is unclear why this would be the case. It may be that the measurements in our study were too heterogeneous across the individuals in our study, or that we need more individuals to extract a clear signal from the metabolite data.
This work adds to a growing body of evidence that it may be possible to develop a test that combines demographics, genetics and protein levels to provide predict whether an individual has abnormal levels of amyloid affecting their brain. The immediate utility of such tests is likely to be in clinical trial enrichment, with the test used to select which individuals who have are cognitively unimpaired but are at risk of developing Alzheimer disease and hence should be enrolled. However, this research is in its early stages and more work needs to be done to demonstrate that our modeling replicates across multiple cohorts, especially those primarily made-up of individuals who are cognitively unimpaired.
Blood-based tests for Alzheimer disease are rapidly developing. Their use in clinical trial enrichment is already becoming a reality with some cohorts being screening using plasma amyloid-beta, whereby mass spectrometry is used as a very sensitive measure of amyloid in plasma and is highly correlated with PET amyloid levels. We believe that the use of multiple plasma analytes, identified using machine learning, is likely to be a complementary approach to other Alzheimer blood tests which focus on very sensitive measures of single analytes. In both cases, replication remains a key challenge.
One of the advantages of the machine-learning approach utilized in this work is that the same framework can be adapted to examine other proteins which may be relevant to Alzheimer disease or other neurodegenerative diseases. We are presenting work on adapting the same framework to predict levels of tau in spinal fluid and expanding the use of genetic markers at AD/PD at the end of March.
Transcript edited for clarity.