🧬 biology

Genome based subtyping identifies Parkinson’s disease subtypes with reproducible motor progression differences and clinical trial enrichment potential

This study demonstrates that genome-based subtyping of early Parkinson's disease identifies three reproducible clusters with distinct motor progression rates and clinical outcomes, offering a strategy to enrich clinical trial cohorts and significantly reduce required sample sizes.

Original authors: Rafaela Ravazio, Daniel Teixeira-dos-Santos, Christian Mattjie, Alvaro de Oliveira Franco, Marco Antônio De Bastiani, Andrei Bieger, Eduardo R. Zimmer, Joana Meneguzzo Pasquali, Lucas S. Kupssinskü, R

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rafaela Ravazio, Daniel Teixeira-dos-Santos, Christian Mattjie, Alvaro de Oliveira Franco, Marco Antônio De Bastiani, Andrei Bieger, Eduardo R. Zimmer, Joana Meneguzzo Pasquali, Lucas S. Kupssinskü, Rodrigo C. Barros, Marco A. M. Schlindwein, Lara Angi Souza, Thomas Hugentobler Schlickmann, Artur F. S. Schuh

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine Parkinson's disease isn't just one giant, messy blob of a problem, but more like a huge box of different kinds of LEGO sets. For a long time, doctors have tried to sort these sets by looking at the finished models (the symptoms) or asking the builders how they feel. But sometimes, the instructions are hidden, the pieces shift around, or the builders are having a bad day, making it hard to tell which set is which.

This paper suggests a totally new way to sort the boxes: look at the instruction manual before you even start building.

The DNA "Instruction Manual"

The researchers took a massive look at the genetic "instruction manuals" (DNA) of 1,419 people with early Parkinson's. They didn't look at how the patients were moving or feeling at that moment. Instead, they used a super-smart computer program (an "autoencoder") to scan 98 specific genetic switches known to be linked to Parkinson's risk and progression.

Think of these 98 switches as a unique barcode for each person's disease. The computer grouped these barcodes into three distinct clusters, which the authors named C1, C2, and C3.

The Big Reveal: Three Different Speeds

Here is the cool part: even though the computer only looked at the DNA and ignored the symptoms, the three groups turned out to be moving at very different speeds.

Group C1 is the "Fast Lane." These people's motor symptoms got worse the quickest.
Group C3 is the "Slow Lane." Their symptoms progressed much more slowly.
Group C2 was somewhere in the middle, but with a twist: they tended to get stiffer (rigidity) faster than the others.

The team tested this in two different groups of people (a "discovery" group and a "validation" group) to make sure it wasn't just a fluke. The results were consistent:

In the first group, C1 got worse by 3.21 points per year on a standard movement test (MDS-UPDRS III), while C3 only got worse by 2.14 points per year.
In the second group, the gap was even wider: C1 worsened by 3.21 points per year, while C3 barely moved at 0.49 points per year.

The authors explicitly argue against the idea that these differences were just because the groups started out different (like being older or sicker at the beginning). They checked the starting line carefully, and everyone was pretty similar. The speed difference seems to come from the genetic blueprint itself.

What This Means for the Future (and the Lab)

The paper suggests that knowing a patient's "barcode" early on could help doctors predict their future.

The DBS Connection: In the group followed for up to 10 years, people in the "Fast Lane" (C1) were more likely to need a specific surgery called Deep Brain Stimulation (DBS) sooner. The math suggests C1 patients were about twice as likely to need this surgery compared to C2 or C3, even after adjusting for age.
The Clinical Trial Game-Changer: This is where it gets really exciting for science. Imagine you are trying to test a new drug that is supposed to slow down Parkinson's. If you mix "Fast Lane" and "Slow Lane" people together in your test, the drug might look like it's not working because the "Slow Lane" people weren't going to get worse anyway, or the "Fast Lane" people got too sick too fast.

The authors ran a simulation (a computer guess based on their data) to see what would happen if they only picked "Fast Lane" (C1) people for a trial. They found that if they enriched the group with more C1 participants, they could shrink the size of the trial needed by up to 45.4%. That means fewer people needed, less money, and faster answers.

What the Paper Doesn't Say

It's important to keep our feet on the ground. The paper does not say this is a magic cure or that we can change the speed of the disease yet.

They ruled out that the groups were different because of how much medication they were taking or their race (mostly).
They suggest but do not prove that these groups might have different levels of a specific protein (alpha-synuclein) in their spinal fluid, but the data there was a bit fuzzy and needs more checking.
The trial size reduction is a simulation, not a real-world result yet. It shows the potential to save time, not a guarantee.
The study mostly looked at people of European ancestry, so we don't know yet if this "barcode" works exactly the same way for everyone else.

The Takeaway

This study is like finding a secret code in the DNA that predicts how fast a car will drive down a bumpy road, even before the car starts moving. It doesn't fix the bumps, but it tells us which drivers are heading for the rough patch and which are cruising on smooth pavement. By sorting patients this way, scientists hope to run better experiments and maybe, one day, give the right treatment to the right person at the right time.

Technical Summary: Genome-Based Subtyping of Parkinson's Disease

Problem Statement
Parkinson's disease (PD) is characterized by significant clinical and biological heterogeneity, which complicates the prediction of progression, treatment response, and the identification of underlying biological mechanisms. Current subtyping strategies often rely on clinical variables (e.g., tremor-dominant vs. postural instability gait disorder), which can be unstable over time and vulnerable to confounding factors such as medication state, rater variability, and disease stage. While genetic variation offers a stable, objective input for stratification, it remains underutilized in PD subtyping. The authors address the challenge of identifying reproducible PD subtypes using only baseline genomic data, independent of clinical information, to determine if such subtypes predict distinct longitudinal motor trajectories and treatment milestones.

Methodology
The study employed an unsupervised machine learning framework applied to whole-genome sequencing (WGS) data from three independent cohorts: the Parkinson's Progression Markers Initiative (PPMI) as the discovery cohort ( $n=265$ ), and the Parkinson's Disease Biomarkers Program (PDBP) and STEADY-PD III trial (STEADY) combined as an external validation cohort ( $n=104$ ). The total analytic sample included 369 participants with early-stage PD (disease duration $\le$ 2–3 years).

Feature Selection: The model utilized 98 specific single-nucleotide polymorphisms (SNPs) previously associated with PD susceptibility (90 variants) and progression (8 variants) in European populations.
Model Architecture:
- Training: An autoencoder was trained on 1,050 non-analytic PD participants (from the same data sources but excluded from the analysis) to learn compact representations of the 98 SNPs. This step prevented information leakage.
- Clustering: The latent vectors generated by the autoencoder for the analytic participants were clustered using a Gaussian Mixture Model (GMM).
- Selection Criteria: Solutions were filtered for stability (using an Adjusted Rand Index threshold) and interpretability. The final model prioritized clusters that demonstrated consistent separation in MDS-UPDRS III burden or progression across both discovery and validation cohorts.
Validation and Analysis:
- Longitudinal Outcomes: Linear mixed-effects models were used to evaluate annual progression in MDS-UPDRS Parts II and III (total and subscores) from baseline to year 3.
- Extended Outcomes: In the PPMI cohort (up to 10 years follow-up), the study analyzed Deep Brain Stimulation (DBS) incidence using Cox proportional hazards models (adjusted for age), Levodopa Equivalent Daily Dose (LEDD) trajectories, and biomarker profiles (CSF/serum markers, including Seed Amplification Assay [SAA] kinetics).
- Enrichment Simulation: A theoretical simulation calculated the reduction in required sample size for a hypothetical two-arm clinical trial if cohorts were enriched with specific subtypes.

Key Results
The analysis identified three genetically derived clusters (C1, C2, and C3) that were comparable at baseline regarding demographics, disease duration, and polygenic risk scores, but diverged significantly in longitudinal outcomes.

Motor Progression:
- Discovery (PPMI): C1 exhibited the fastest MDS-UPDRS III progression (3.21 points/year) compared to C3 (2.14 points/year; $p=0.0072$ ).
- Validation (PDBP + STEADY): The pattern was replicated, with C1 showing 3.21 points/year progression versus 0.49 points/year for C3 ( $p=0.0034$ ).
- Subscores: Gait and posture progression was highest in C1. Rigidity progression was highest in C2. Tremor, bradykinesia, and speech scores did not show consistent differences across datasets.
Treatment Milestones (PPMI Extended Follow-up):
- DBS Surgery: C1 had the highest crude rate of DBS surgery (21.7%) compared to C2 (14.1%) and C3 (13.6%). Age-adjusted Cox regression showed C2 and C3 had significantly lower hazards for DBS compared to C1 (HR 0.40 and 0.35, respectively).
- Medication: C2 showed a lower yearly increase in LEDD compared to C1. No significant differences were found in levodopa response trajectories.
Biomarkers:
- Standard biomarkers (CSF A $\beta$ 1–42, tau, NfL, GFAP, $\alpha$ -synuclein) showed no adjusted differences across clusters.
- SAA positivity was numerically higher in C1 and C2 than C3 but did not reach statistical significance ( $p=0.052$ ). However, the mean time to maximum slope at 24 hours in SAA kinetics was significantly higher in C1 compared to C2 and C3 ( $p=0.026$ ).
Clinical Trial Enrichment:
- Simulations indicated that enriching a trial cohort with C1 participants could reduce the required sample size by up to 45.4% (reducing from 388 to 212 participants per arm) if the C1 fraction was increased from 32% to 100%.

Significance and Claims
The authors claim that this study demonstrates the feasibility of using baseline polygenic variation alone to identify early PD subtypes with reproducible differences in motor progression and treatment-related milestones. Key contributions include:

Genetic Stratification: The study provides evidence that common variant structure can stratify patients into groups with distinct motor trajectories before longitudinal clinical divergence becomes evident, offering a fixed, objective input for subtyping that avoids the confounding of medication state and assessment timing.
External Validation: Unlike many previous data-driven studies, this work successfully replicated the motor progression pattern in an independent external cohort (PDBP/STEADY), addressing a critical gap in the literature regarding the generalizability of PD subtypes.
Clinical Utility: The findings suggest that genomic subtyping can serve as a tool for "motor prognostic stratification." Specifically, the ability to identify a "fast-progressing" subgroup (C1) offers practical value for clinical trial enrichment, potentially increasing statistical power and reducing sample sizes for disease-modifying interventions.
Deployable Framework: The authors provide a publicly available algorithm (SNP2PDsubtype) for assigning individual patients to these genetically derived clusters, moving beyond theoretical clustering to a deployable framework for research and potential clinical workflows.

The paper concludes modestly, acknowledging limitations such as the restriction to European ancestry populations, the focus on early disease stages, and the exclusion of rare variants and gene-environment interactions. They emphasize that while the approach shows promise for prognostication and trial design, prospective validation in broader ancestries and disease stages is required.