Predicting Parkinson's from genetics: when cohort enrichment turns a risk gene into a "protective" one

A researcher at Claremont Graduate University trains, on 1,987 participants from the Fox Insight cohort, a battery of models — from a penalized logistic regression to an ensemble of algorithms — to predict a Parkinson's diagnosis from a handful of targeted genetic variants. The best model reaches an AUC of 0.929 on a held-out test set, a seemingly excellent figure. But the cohort holds 1,861 patients for only 126 controls, and the field's best-known risk variant comes out "protective": the performance mostly measures a recruitment artifact, not a screening ability.

The context

Parkinson's disease is the world's second most common neurodegenerative disease, and one of the most genetically studied. Genome-wide association studies (GWAS: comparing the DNA of tens of thousands of patients and controls to spot the genomic regions linked to the disease) have identified, since 2017, more than ninety risk regions. A few genes recur systematically: GBA, LRRK2, PRKN. The GBA N370S variant in particular is one of the best-characterized hereditary risk factors for the disease.

Hence a temptation: what if the diagnosis could be predicted from genotype alone? Earlier work on the same cohort plateaued around an AUC of 0.74 — the AUC (area under the ROC curve) measuring a model's ability to correctly rank a randomly drawn patient above a randomly drawn control: 0.5 = coin flip, 1.0 = perfect discrimination. This preprint, single-authored and not yet peer-reviewed, aims to do better while staying interpretable and "economical", by combining classical statistics and machine learning on a small set of targeted variants.

The method

The data come from Fox Insight, a large online cohort of the Michael J. Fox Foundation, where volunteers self-report their diagnosis and share genetic data. The analysis sample holds 1,987 participants (1,132 men, 855 women), mean age 64, of whom 1,861 self-reported as having Parkinson's and 126 as controls. The predictors are a small panel of targeted variants — GBA N370S (rs76763715), LRRK2 G2019S (rs34637584), GBA E326K, GBA T369M, PRKN R275W, plus a few extra SNPs — alongside age, sex and handedness.

Two model families are compared. First a Firth penalized logistic regression: a logistic regression whose estimates are corrected to avoid aberrant values when a variant is rare and almost always tied to the same status (the so-called "separation" problem). It produces odds ratios (OR: above 1, the factor raises the odds of being classed as a patient; below 1, it lowers them), with confidence intervals and a correction for multiple comparisons via the Benjamini-Hochberg procedure (which caps the false discovery rate, or FDR, when many variants are tested at once).

Then an ensemble of algorithms assembled automatically (GBM and XGBoost — boosted tree forests —, random forest, neural network, generalized linear model), topped by a meta-model that learns to combine their predictions (a technique called stacking). Training, run in six minutes on a GPU, ultimately keeps a plain GBM as the best model. A low-rank model (GLRM), which compresses and "denoises" the variables, is tested as an add-on to enrich the inputs. The data are split into 70% training, 15% validation, 15% strictly held-out test. Interpretability rests on permutation variable importance, individual conditional expectation curves, and local Shapley values (SHAP: a method that attributes to each variable its contribution to an individual prediction).

The results

On the analysis sample, the Firth regression shows an AUC of 0.905 and a Brier score of 0.028. Male sex comes out associated with the diagnosis at an OR of 2.02 (95% confidence interval: 1.21 to 3.41; p = 0.007), consistent with the disease's known epidemiology. The most striking result lies elsewhere: carrying the GBA N370S variant comes out at an OR of 0.01 (p < 0.001, FDR < 0.001), a massive "protective" effect — the exact opposite of the established literature, where this variant raises risk. The author states it plainly: this signal reflects a recruitment enrichment, not biological protection. The other variants (rs12456492, rs823118, rs2280104) show nominally significant associations, but none survives the correction for multiple comparisons.

On the machine-learning side, the best GBM reaches, on the 15% of never-seen data, an AUC of 0.929, a Gini coefficient of 0.859, a log-loss of 0.108 and an MCC of 0.743. The MCC (Matthews correlation coefficient) is a balanced measure that accounts for all four cells of the confusion matrix, useful precisely when classes are imbalanced. The GLRM-augmented variant does slightly worse (AUC 0.902, MCC 0.727): compressing the variables adds nothing here.

That leaves the clinical translation, and this is where it breaks down. The test set holds about three hundred participants, an overwhelming majority of them patients and a handful of controls. The preprint's text in fact reports contradictory confusion-matrix figures (it mentions "280 controls" and "17 cases" in the test, which inverts the cohort's actual composition, 94% patients). On a set where almost everyone is already diagnosed, a model that answered "patient" to everyone would already score flatteringly: the metrics must therefore be read with the utmost caution.

What is good

Rare honesty about its own artifact. The author does not hide that the inverted GBA N370S signal comes from the way the cohort was recruited, and explicitly distinguishes in-distribution predictive value from any causal inference in the population. Many papers would have sold this "protective gene" as a discovery; this one defuses it itself.

Fitting statistical choices. Firth regression is the right tool for rare, separation-prone variants, and applying a Benjamini-Hochberg correction — which knocks out all candidate associations — is an honest discipline: it prevents over-reading fragile signals rather than celebrating them.

Systematic interpretability. Variable importance, conditional expectation curves, local SHAP values and per-model individual predictions: together they allow auditing why a prediction is made, rather than delivering a black box. It is the right direction for collaborative human-AI use.

What is less good

A population bias that hollows out the AUC. With 1,861 patients for 126 controls, prevalence in the sample nears 94%, against under 0.3% in the general population. An AUC of 0.929 on such an imbalanced set says almost nothing about real screening: it is the textbook of population bias coupled with the misleading metric. Worse, the model may learn markers of who signs up for Fox Insight rather than the disease's biology — a case of shortcut learning, where the algorithm exploits a spurious recruitment correlation.

No out-of-distribution validation. The test set comes from the same cohort as training; the author himself acknowledges not knowing whether the models hold on other populations. Yet that is precisely where this kind of predictor usually collapses. Add an unexplained inconsistency in the confusion-matrix figures (280 controls claimed in a test drawn from a cohort that holds only 126 in total), which undermines any clinical reading of false positives and false negatives.

Limited power and robustness. Only 126 controls, a non-peer-reviewed preprint, a single author, and long philosophical digressions that dilute the methods section. The "best model" is picked after the fact among several AutoML passes, exposing it to selection over-optimism; and the only comparator is internal (Firth versus ensemble), with no confrontation to a standard polygenic risk score that would have given a credible reference.

What it changes

For the research community, the value is mainly pedagogical: the paper illustrates, almost in pure form, how a self-recruited patient cohort traps genetic prediction and manufactures inverted associations. It is a reminder that a model is worth only as much as its cohort's representativeness, not its AUC figure. The pipeline (AutoML plus interpretability) remains technically replicable for anyone with better-balanced data and an external validation cohort.

For clinicians, nothing changes in practice. There is no deployable tool here: a model trained on a population 94% of whom are patients cannot screen for a rare disease in the general population. The practical message is defensive — beware of high AUCs from imbalanced cohorts, and demand external validation before any conclusion.

For patients and the general public, this must not be read as a "Parkinson's genetic test", nor should the advertised "protective gene" be believed, being only a statistical mirage due to the recruitment method. It is, however, an excellent case for understanding why an impressive figure can mean nothing the moment one leaves the sample on which it was computed.

To go further

The preprint is available on medRxiv (10.64898/2026.07.11.26357847), under a CC-BY-NC-ND 4.0 license; it is single-authored from Claremont Graduate University and has not been peer-reviewed. The data come from the Michael J. Fox Foundation's Fox Insight cohort (access on request). On genetic risk prediction and the importance of calibration and interpretability, see our decryptages on a stroke-risk score combining XGBoost and polygenic risk, on using SHAP to interpret a thromboembolism model, and on the calibration of an ICU mortality model.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.