Parkinson’s disease (PD), a complex neurodegenerative disorder, has long perplexed scientists due to its heterogeneous nature, involving a range of genetic and environmental factors. Recent advancements in genetics and technology have allowed researchers to delve deeper into the molecular underpinnings of the disease.
Genome-wide association studies (GWAS) have illuminated the significance of genetic risk loci tied to protein homeostasis, protein trafficking, lysosomal function, and mitochondrial function in sporadic cases of the disease. These pathways play a pivotal role in disease development and progression.
In a groundbreaking study, scientists harnessed the potential of human induced pluripotent stem cell (hiPSC)-derived cortical neurons, which mirror the vulnerable neuronal type in PD. These neurons replicate key cellular traits associated with PD, enabling the researchers to model and categorize four mechanistic subtypes of the disease.
The subtypes are based on the presence of familial mutations, proteotoxic stress, mitochondrial stress, and induced mitochondrial clearance. Using a live high-content imaging system, the researchers meticulously tracked these disease mechanisms across crucial organelles – mitochondria, lysosomes, and nuclei.
This innovative approach holds immense promise as a preclinical platform with high predictive capabilities for PD. It offers a human-based model of brain disease that captures real-time information on the central organelles implicated in PD pathogenesis. This amalgamation of advantages culminated in the development of a remarkably accurate deep learning classifier. The classifier not only discerns the presence or absence of PD but also accurately identifies the specific disease subtype.
The classifier’s accuracy is grounded in its integration of diverse features derived from mitochondria, lysosomes, and nuclei. These features encompass intensity, morphological, and textural characteristics, extracted via the live high-content imaging system. Importantly, the researchers used SHAP (SHapley Additive exPlanations) analysis to rank the contribution of individual features to the model’s prediction.
This process unveiled the pivotal role of mitochondrial features, particularly in predicting mitochondrial pathways and inducing mitophagy. Lysosomal and nuclear features also featured prominently in predicting aggregation pathways.
Furthermore, the researchers explored the interactions between mitochondria and lysosomes, showing that tabular data based solely on these interactions effectively distinguished aggregation and mitochondrial toxicity pathways. This finding underscores the biological relevance of these interactions in PD.
While the tabular data-based model offers transparency and interpretability, it may face challenges related to experimental variations and software-related uncertainties. To enhance generalizability, the researchers employed convolutional neural networks (CNNs) to develop image classifiers using the same dataset. The deep CNN-based image classifiers demonstrated high accuracy in categorizing healthy and diseased states.
This groundbreaking approach offers several advantages over traditional image analysis methods. While conventional methods focus on specific structural properties and automation, they often overlook the wealth of information within imaging data. Machine learning, particularly CNNs, has the capacity to decipher cellular traits in an unbiased manner, outperforming traditional methods in accuracy.
The study also highlighted the versatility of their approach. By combining classifiers based on tabular features and image tiles, the researchers successfully translated the explainability gained from tabular data into image classifiers. Notably, the absence of nuclear signals in a CNN image-based model did not compromise accuracy.
The CNN trained solely on mitochondria outperformed the lysosome-based model, consistent with the importance of mitochondrial features highlighted by the tabular models.
This pioneering study extends its impact to chemically induced subtypes of PD, demonstrating proof of concept for their experimental paradigm’s efficacy. Recognizing the diverse genetic backgrounds of real-world patients, the researchers also developed a classifier centered on mutations tied to disease subtypes.
In conclusion, the study showcases the potency of employing deep learning techniques in predicting the intricate mechanisms underlying PD. With the disease’s heterogeneity posing a significant challenge, this platform offers the potential to classify disease mechanisms within patient cells.
This holds substantial clinical implications for diagnosis and treatment, tailoring therapeutic approaches based on cellular mechanisms. The platform’s potential extends to evaluating individual pathways and assessing the efficacy of medications in an unbiased manner. As researchers continue to explore these avenues, the future of PD diagnosis and treatment appears increasingly promising.
reference link : https://www.nature.com/articles/s42256-023-00702-9