Large-scale identification of clinical and genetic predictors of Parkinson’s disease motor progression in newly-diagnosed patients: a longitudinal cohort study and validation

Introduction

Parkinson’s disease (PD) is a chronic, debilitating neurodegenerative disorder characterized clinically by progressive motor dysfunction and various non-motor features ¹. There is substantial heterogeneity in the presentation of these symptoms and the rates of their progression among PD patients, making it difficult for care-providers to give accurate prognoses to patients and challenging for researchers to develop drugs to modify the course of disease ². Increasing evidence supports a complex interplay between genetic, biological, and molecular abnormalities of the disease explaining this heterogeneity between patients. Understanding the etiologic and physiological factors that contribute to this variability in the evolution of PD symptoms is therefore a high priority area of PD research ². The Parkinson’s Progression Markers Initiative (PPMI) study was initiated with the support of the Michael J. Fox Foundation for Parkinson’s Research to address this research gap by providing a uniquely comprehensive set of longitudinal clinical, imaging, and bio-sample data from de novo PD patients and controls.

Previous analyses of PD progression data have generally focused on investigating the individual associations of predictive features such as age, sex, baseline scores, clinical subtypes and varied potential biomarkers, as opposed to developing comprehensive multivariable prognostic models ^3–5. There have been exceptions, however, including logistic regression and Bayesian classification models to predict cognitive impairment in PD^6,7, backwards selection models to predict negative outcomes (e.g. postural instability, dementia or death)⁸, and machine-learning random survival forests to predict time to initiation of symptomatic treatment⁹. However, models that predict the rate of change of motor scores assessed using the Movement Disorder Society-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), instead of specific clinical events, remain elusive, despite that changes in this rating scale score and its subcomponents serve as the primary endpoint in many clinical trials of PD medications ¹⁰. Here, we used a novel Bayesian machine-learning platform, Reverse Engineering and Forward Simulation (REFS^TM), with the PPMI compendia to identify an ensemble of predictive models, instead of a single best model, of the rate of motor progression.

This analysis is differentiated from previous studies not only in the outcome studied, but also in the scope and complexity of the predictive relationships that are explored by examining the entire complement of available genetic, molecular and clinical variables^11–14. In particular, REFS allows for the detection of higher order interactions across these different data types and, consequently, identification of subpopulation-specific effects. This allows the crucial distinction between modifiers that are specific to PD cases from those that are more generally related to natural aging or other risk factors.

Taken together, these features permit three complementary objectives for this study. (1) The construction of clinically useful predictive models to identify PD patients at risk of rapid disease motor progression. (2) A comparison of the predictive utility of different types of potential biomarkers of motor progression. (3) The identification of novel progression markers for subsequent validation in an independent test sample (LABS-PD, N=317), which, notably, uncovered a significant epistatic genetic association to disease progression upon replication. We also demonstrate the utility of these models in designing PD clinical trials that aim to test potential disease-modifying therapies.

Materials and Methods

Results

A total of 117 controls and 312 PD cases with complete genetic, molecular and 2+ years of longitudinal data were available for analysis of motor progression (Supplemental Table S1). As expected, the mean (±standard deviation) rate of motor progression was significantly higher (i.e. more severe) among PD cases compared to controls (5.05 ± 3.3 vs. −0.14 ± 0.64 points/year, p = < 2e-16), with modest but significant reductions in progression rates observed during treatment periods with symptomatic therapy among PD cases (Table S1).

A total of 17,499 features were included as potential predictors in the models. These included 53 a priori selected PD-related single nucleotide polymorphism (SNPs); 17,403 LD-pruned SNPs from across the genome; 7 cerebrospinal fluid (CSF) protein biomarkers; 8 DaTscan imaging variables; 10 genetic principal components; and 18 clinical and demographic variables. The included variables are described in more detail in Supplemental Tables S2.a and S2.b, and the Supplemental methods.

The final REFS ensemble is summarized as a reduced set of weighted constituent models including most of the important predictors identified from the full ensemble (N=12 weighted models; Supplemental Table S3) that optimally approximates the full ensemble of 128 total unweighted constituent models ²⁶.

Discussion

Using a hypothesis-free machine learning ensemble approach suitable for large-scale multivariate modeling, we have developed predictive models of motor symptom progression in early-stage Parkinson’s disease cases and age- and gender-matched controls. We have further identified the relative contribution of individual patient factors and sets of factors, and replicated several specific associations in an independent PD cohort, including a novel epistatic interaction. While previous progression modeling efforts have often focused on specific binary clinical outcomes (such as initiation of dopaminergic treatment or development of MCI), we used linear mixed effects models to estimate the continuous treated and untreated rates of clinical progression for each individual. Looking at strata-specific model performance, the predictive accuracy of the models was unsurprisingly greater among the PD cases, as there was very little variability in motor symptoms over time in the controls. Drug-naïve progression was similarly modeled more accurately than treated progression both in and out of sample, potentially due to heterogeneity in treatment regimens among those receiving PD medications. Additionally, it is well-recognized that there is large inter-individual variability in response to anti-Parkinsonian treatments, with this variability becoming more prominent over time as disease progresses ^27,28.

While a substantial drop-off in the amount of variability explained was observed in the LABS-PD test set, this was expected, as several predictors were unavailable in LABS-PD (including CSF biomarkers and the baseline SCOPA exam). The required conversion of UPDRS to MDS-UPDRS scores for the outcome data also likely introduced significant variability. Furthermore, among cases, predictive accuracy for motor progression was greater in the earlier stages of disease among the PPMI cohort. In contrast, prediction in later-stage patients was only significant in the LABS-PD validation cohort. These discrepancies in the accuracy of predictions were likely due to differences in the composition of cohorts, as the large majority of the sample in LABS-PD (95%) was followed for longer than 5 years as opposed to a minority in PPMI (19%). As the models assume a linear trend in progression, another explanation may be that progression in later disease stages is associated with the same factors as early-stage, but the relationships are non-linear over time ^4,28. Although these limitations in comparability between the test and training cohorts resulted in reduced (but still significant) R², they also provide an important demonstration of the robustness of the ensemble modeling to diverse cohorts and data missingness. We can see that despite the differences from the training sample, the models demonstrate an unequivocal ability to prospectively differentiate between patients in the test set who would manifest slow or moderate progression and those whose condition would deteriorate more rapidly (Figure 3).

Variable importance measures determined using leave-one-out cross validation provided insight into the comparative predictive value of three different general types of potential predictive markers or tests available to clinicians (i.e. genetic variation, CSF biomarkers, and DaTscan imaging). The relatively limited prognostic utility of the DaTscan imaging data was in contrast to a previous study showing association between baseline imaging measures and various long-term PD outcomes ²⁹. As our study is focused on early-stage PD progression, as opposed to long-term outcomes, this is intriguingly suggestive of distinctions between the factors determining disease progression at different disease stages. Of the CSF biomarkers, alpha synuclein levels, which have previously been inversely associated with motor symptom severity³⁰, were the most common predictor of motor progression, with higher levels similarly predictive of slower progression.

Examination of the implicated features available for study in the LABS-PD cohort replicated a majority of clinical features, including sex, which showed women having a slower rate of decline in both cohorts. Sex differences are consistently observed in PD, most notably in prevalence, with men being more commonly affected, but also in age of onset and disease presentation ³¹. An interaction between sex and treatment was also observed, but only reached nominal significance in the replication cohort. Treatment and SWEDD status, both as main effects and in interaction, failed to replicate, likely as a consequence of the significant differences in prevalence of each factor in the two cohorts.

An interaction between two SNPs identified at high frequency in the motor progression ensemble was replicated in the LABS-PD cohort. As shown in Figure 2, cases who carry minor alleles for both SNPs have on average a substantially faster rate of motor decline in both the LABS-PD (1.2 points/per year faster) and PPMI (2.4 points/year). Neither SNP was among the a priori selected PD-related SNPS identified from GWAS, but intriguingly, rs9298897 is located in the fifth intron of the gene leucine rich repeat and Ig domain containing 2 (LINGO2). LINGO2 and its paralog LINGO1 encode for type I leucine-rich transmembrane proteins exclusively expressed in the central nervous system^32,33. Like LRRK2, which is also of the leucine-rich repeat protein family and is the greatest known genetic contributor to PD, LINGO1 and LINGO2 are thought to function in the regulation of cell signaling related to neuronal survival and growth as well as glial function^34–37. In experimental models of PD, increased expression of LINGO-1 were found in animal models of PD after parkinsonism-inducing neurotoxic lesions, while the absence or inhibition of LINGO-1 resulted in increased dopamine neuron protection³⁵. Polymorphisms at these genes have been previously implicated in risk of PD and essential tremor (ET)³⁸ but these findings have not been conclusive ³⁹. The interacting SNP rs17710829 is located in an intergenic region on 2q14.1. The nearest gene dipeptidyl peptidase like 10 (DPP10) has not been linked to neurodegenerative disease in GWAS or other studies of genetic variation, however abnormal expression of the encoded protein, DPP10, has been observed in Alzheimer’s disease and other neurodegenerative diseases ⁴⁰. Additional replication of this finding is warranted, as the functional link between these two genes is unclear, but further understanding of this interaction may lead to important mechanistic insights in the disease process.

Despite stringent sample inclusion criteria, robust data pre-processing, and penalization of model complexity via Bayesian ensemble inference, the generality of findings reported here are subject to several limitations associated with PPMI and LABS-PD study designs. The PPMI profiles used to develop the models are principally restricted to short-term follow-up after enrollment (median < 4 years), which may limit predictive accuracy when evaluated over longer periods of disease progression. Notably, in the LABS-PD cohort with average follow-up times exceeding 6 years, we see slower overall rates of progression (Supplemental Table S5) with a narrower range (Figure 2) suggestive of a potential plateau or other non-linear treads in progression. Additional characterization of the natural disease trajectories for each assessment domain, including more sophisticated incorporation of site and treatment effects may refine the clinically reliable window for linear rate model predictions beyond 5 years ^28,41–43. Independent validation and variable replication findings reported herein may be additionally biased by variation in follow-up time distribution, distinct outcome assessment tools, and missing covariates between discovery and validation cohorts.

In summary, this analysis highlights the ability of ensemble modeling to capture the complex interplay of clinical, genetic, and molecular predictors of the highly heterogeneous PD progression phenotype. Bayesian model inference using REFS identified a combination of established and novel patient factors predictive of PD motor progression, particularly in the earlier stages of disease. The ability of the Bayesian model selection process to integrate genetic data with CSF-based biomarkers is of notable clinical relevance, as these data classes showed the greatest predictive impact in our models. Further, the quantification of the comparative predictive importance of the different data types, particularly the limited predictive utility of the DaTscan imaging in early-stage progression, provides guidance to the effective deployment of clinical and research resources in longitudinal patient evaluations.

The ability of the ensembles to prospectively identify patients most likely to have rapid progression of symptoms at the earliest stage of the disease has immediate significance towards enabling more effective trial recruitment and clinical disease management. The identification and independent confirmation of predictive factors also provides potential mechanistic insight into the disease process. Identification of a PD-specific novel genetic interaction between an intronic LINGO2 (9p21.1) SNP rs9298897 and the 2q14.1 variant rs17710829 was made possible through the multi-dimensional, hypothesis-free methodologies implemented here. Taken together these results advance the goal of establishing validated biomarkers of PD progression rate and improving existing prognostic models to the benefit of trial design and evaluation, as well as clinical disease monitoring and treatment.

Supplementary Material