Parkinson’s disease is a clinically heterogenous neurodegenerative disorder, meaning that it is likely that different cellular mechanisms drive its pathology in different individuals. As it is currently impossible to know which Parkinson’s disease subtype a patient has while they are still alive, it is therefore hard to know which treatment course is likely to be the most effective. A new study published in Nature Machine Intelligence uses AI to not only accurately distinguish between Parkinson’s disease and healthy tissue, but also to predict the disease subtype, based on images of stem cells.
The age of onset, rate of disease progression and severity of motor and non-motor symptoms display considerable individual variation in Parkinson’s disease, making it difficult to classify the disorder into definitive subtypes. These differing symptoms are thought to be due to the underlying mechanisms in each disease subtype. If these underlying mechanisms could be characterised, we may be able to choose treatment courses in line with what is likely to be most effective for that particular patient.
However, there are no widely used approaches that can define the molecular heterogeneity within a patient’s brain, and therefore the mechanisms that drive these different phenotypic subtypes are currently unknown. If this challenge could be met, this could enable early and accurate molecular-level diagnosis allowing personalised therapeutic interventions, at the earliest possible timepoint.
A beneficial collaboration
Researchers at the Francis Crick Institute and the UCL Queen Square Institute of Neurology collaborated with the technology company Faculty AI to apply a deep learning approach to human cellular models of Parkinson’s disease, generating a predictive model of different disease mechanisms.
This model focussed on the two key pathways that are known to drive Parkinson’s disease pathology:
2. The accumulation of abnormal mitochondria with impaired bioenergetic function and reduced mitochondrial clearance.
Using patient-derived, induced pluripotent stem cell (iPSC)-derived cortical neurons to model Parkinson’s disease, the researchers were then able to classify four cellular subtypes which relate to these two key pathways (Figure 1):
Subtype 1: Patients with overexpression mutations in the SNCA gene encoding α-Syn, who develop an autosomal dominant aggressive form of Parkinson’s with predominant protein aggregation.
Subtype 2: Patients with proteotoxic stress caused by protein aggregates of α-Syn spreading from cell to cell in the brain in a prion-like manner.
Subtype 3: Patients with mitochondrial complex 1 impairment who can develop Parkinson’s disease following exposure to toxins such as pesticides.
Subtype 4: Patients with mutations in the PINK1 and PARKIN genes who develop autosomal recessive early-onset Parkinson’s disease, in association with impaired clearance of damaged mitochondria and accumulation of abnormal mitochondria in neurons.
Following fluorescent labelling of the nucleus, mitochondria and lysosomes of these iPSC-derived neuron subtypes, high-content live single-cell imaging of 1,560,315 cells was used to generate disease models (Figure 2). These were then used to create a highly accurate deep learning classifier that can distinguish between the presence or absence of Parkinson’s disease and, if present, the subtype of the disease. The classifiers achieved 80-100% accuracy at distinguishing between healthy and diseased tissue across the different disease states, with an average 82% accuracy for distinguishing between the four disease subtypes.
This study showed that AI analysis can distinguish between healthy and diseased tissue and between these four disease subtypes with high accuracy. This accuracy, alongside the unbiased nature of machine learning, is one of the main benefits of using this technology for this kind of analysis compared to traditional methods.
The researchers hope this study can generate useful predictions about the mechanistic subtypes of Parkinson’s disease. As this disease is highly heterogeneous, this may enable patient-derived cells to be classified into a subtype, which could significantly improve both diagnosis speed and treatment efficacy.
Identification of cellular mechanisms may indicate a patient’s likely response to proteinopathy treatments, such as those targeting α-Syn, compared to mitochondrial treatments, such as antioxidant therapies. In the future, AI and machine learning techniques such as this could be used to assess which pathway predominates in a patient, indicating whether specific medications could reverse the cellular phenotypes present and improve disease outcomes.