Multi-Omic Graph Diagnosis (MOGDx)

MOGDx, shown in S1 Fig, is a flexible tool to integrate multiple omic measures and perform classification tasks. It uses a patient similarity measure to identify patients who have similar molecular, epigenetic, and demographic disease characteristics and performs node classification using a GCN. The performance of MOGDx was benchmarked on cancer data and achieved state-of-the-art performance compared to similar research [11].

There are two main components in the MOGDx framework; PSN generation and Graph Convolutional Network with Multi-Modal Encoder training and classification. A single PSN is built per modality. Feature selection is performed, and similarity is measured between these features using Pearson correlation, where suitable, otherwise Euclidean distance. Each PSN is constructed from the similarity matrix using the k nearest neighbours algorithm. Similarity Network Fusion (SNF) is then used to combine individual PSN’s into a single network. As is common practice, all patient information is used to construct the network, with train, validation and test labels created during the training phase of the GCN-MME [12, 13]. In this approach, non-informative features are discarded during the similarity calculation to discourage uniform similarity scores for modalities such as DNAm which will have a large number of redundant or similar features.

The fused PSN and the modalities are input into the GCN-MME for training and classification. Each modality is compressed using a two layer encoder. The first layer of the encoder is of fixed length, with the second layer being tuned to each modality by performing a hyperparameter search. The compressed encoded layer of each modality is then decompressed to a shared latent space using mean pooling. This methodology follows similar encoder architectures established in other works [14, 15]. Median imputation is also performed at the second layer of each encoder to retain patients if they are missing from that modality. The shared latent space corresponds to the patient node features, which are combined with the PSN and input into the GCN for classification. The GCN-MME is trained under the semi-supervised setting for graph neural networks outlined by Hamilton (2020) [16]. For a detailed description of the MOGDx architecture, please refer to Ryan et al. 2023 [11].

MOGDx is a suitable tool to perform analysis on the PPMI dataset due to its flexibility. It can integrate any number of modalities, whilst simultaneously retaining the maximum number of patients possible, in contrast to other existing methodologies. As discussed by Chan et al. (2022) and as per Table 1, there are fewer HC participants [9]. Moreover, not every participant will be present in each modality at each time point. In order to avail of the full PPMI dataset, a method which can incorporate the maximum number of participants is required. MOGDx achieves this by utilising SNF and imputation to retain patient nodes without a large degradation in performance [11]. MOGDx also provides a high level of interpretability. Due to the flexibility of integration, ablation experiments can be performed to identify the most predictive modalities. As feature selection is performed in the MOGDx pipeline, these features can be further analysed to identify important pathways, traits or interactions of the target application.

PPMI dataset

Data was obtained from the publicly available PPMI dataset [17]. The participant count per year, modalities analysed and number of raw features per modality are summarised in Table 1. In total, 6307 samples from 2188 participants were included in the analysis. Participants in the analysis were identified as PD, PL or HC and by disease subtype of genetic, idiopathic, Rapid eye movement Behaviour Disorder or hyposmia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Breakdown of PPMI dataset by number of participant and modality features.

https://doi.org/10.1371/journal.pone.0313772.t001

The four genomic measures analysed, namely Messenger RNA, Micro RNA, DNA Methylation (DNAm) and Single Nucleotide Polymorphisms, were generated from whole-blood samples by PPMI. Each genomic dataset was further processed in a complimentary bioinformatics pipeline, if available. For example, DNAm was normalised using the wateRmelon package in R [18]. See Ryan et al. (2023) for more detail on how specific genomic modalities were handled [11]. In general, processing included the removal of zero variance or missing features, normalisation, imputation of missing values and conversion from categorical to numerical features. The top 300k most variable CpG sites were retained to allow computation on this dataset to fit into memory. For the same reason, a principal component analysis was performed on the SNPs dataset to reduce the dimensionality of the dataset, and the first 20 PC’s were retained.

Genomic datasets were supplemented with additional measures of 1472 Cerebral Spinal Fluid (CSF) protein markers extracted from participants and clinical descriptors. Clinical descriptors included individual phenotypes of age, sex and years of education; These were supplemented with measures for smoking, alcohol and BMI generated from DNAm profiles [19]. These DNAm profiles were derived from models trained on up to 5087 individuals in a national study in Scotland and tested on two separate cohorts also based in Scotland [19]. The Movement Disorder Society Unified Parkinson’s Disease Rating Scale by Goetz et al. (2008) is a measure of disease severity in those with PD and PL [20]. This scale combines measures relating to both motor and non-motor symptoms of PD. It consists of both self-assessment and clinical assessments and is a proxy of disease stratification [20]. It was used as a baseline comparative model to identify if the biological signal for PD found in the blood is stronger than clinical assessment using MOGDx. These modalities were similarly processed for feature selection, conversion and normalisation.

As per Table 1, there were two methods for feature selection. Where suitable pairwise linear regression between the three strata was performed using the DESeq2 package in R [21]. In other modalities, penalised elastic net regression was performed using the glmnet package in R [22]. Differentially expressed genes with a statistically significant FDR () and logistic regression coefficients with an absolute weight greater than zero were used as selected features in the similarity calculation. The number of features selected depended on the subgroup being analysed. If no informative features were found, all features were retained.

Table 1 shows the participant availability per year. The time points cover the first three years of the disease in the PD cohort. The first time point (labelled year 0) corresponds to participants with PD who have had a diagnosis for less than 2 years, have not begun taking any PD medication and are not expected to require PD medication for at least 6 months [17]. Those in the genetic subgroup of PD have a mutation in one of three genes: LRRK2, SNCA or GBA. Idiopathic individuals do not have mutations in any of these three genes. PL participants have been identified as being of high risk for the disease, but have not yet met a clinical threshold for diagnosis. The first time point, year 0, in this cohort corresponds to their enrolment in the study. The genetic PL subgroup also have a mutation in one of the three aforementioned genes as aligned with the genetic PD subgroup. As per Table 1 the PL participants in the genetic subgroup far outnumber the participants in the RBD and hyposmia subgroups. Participants in these groups have one of two non-motor symptoms associated with PD. RBD is a sleep disorder and hyposmia is a smell disorder, both of which have been identified as early indicators for PD [23, 24]. HC participants were screened by PPMI to ensure they did not meet the criteria for either PD or PL. As with PL, their first time point, year 0, aligns with their enrolment in the PPMI study. PD idiopathic and PL genetic are the two most prevalent subgroups in the dataset. As identified in Chan et al. (2022) there are fewer HCs compared to PD and PL however the numbers presented in Table 1 show higher counts compared to their analysis which was subset to participants who had image data available [9]. A distinguishing factor of this analysis is the utilisation of the longitudinal data in the PPMI dataset. As per Table 1, over time, the number of participants decreases across all strata and subgroups. This is in part due to participant dropout (n = 401), missing samples for a participant at a time point or the transition from PL to a clinical diagnosis for PD (n = 33). Summary of the criteria for participant stratification and disease subgroups are summarised in S2 Fig.

Cross-sectional and longitudinal experiment design

In this analysis, we perform both cross-sectional and longitudinal experiments at 4 time points over three years. In all experiments we classify whether participants have PD, are PL or are a HC. We perform cross-sectional experiments on all participants, regardless of their subgroup and on two subsets based on participants’ subgroup. The first subset, referred to as genetic, includes PD and PL participants in the genetic subgroup. The other subset, referred to as idiopathic, includes participants in the idiopathic PD, RBD and hyposmia subgroups. HC participants are included in both subsets as a control. We use a brute-force approach, testing all combinations of modalities in each experiment to identify the modalities at each time point with the highest accuracies and F1 scores. For fairness of comparison, we train each model independently and perform cross-validation on each model to obtain standard errors for each models’ performance. We also account for differences in the number of participants by using F1-score and comparing each model to a baseline, which only predicts the most common class. In this manner, we can fairly attribute improved performance due to the inclusion or exclusion of a modality. Where two models achieve comparable performance, we report the model which includes the greatest number of participants. The top performing model at each time point, per subgroup is reported in Table 2 and shown in Fig 1 with all model’s performance reported in S1 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Cross-sectional performance of MOGDx in different subgroup experiments.

https://doi.org/10.1371/journal.pone.0313772.t002

In the longitudinal analysis, we repeat the best performing cross-sectional experiment on the genetic PD (n = 70) and genetic PL (n = 84) subgroups. Once again, HC (n = 150) are included as a control. This analysis is restricted to participants who are present at each time point in at least one of the included modalities. Only the optimal combination of modalities which maximised both accuracy and patient retention was analysed. This was identified by averaging the F1 scores of each cross-sectional model across all time points. A model integrating these modalities is trained and tested at each time point. These models are then tested on unseen networks and modalities of the same patients but from the alternative time points. Participants were split into training and test sets, thus performance for all models at all time points are reported on the same test set. Networks were reconstructed during this phase using only the features selected at the test time point. For example, when testing the year 3 network at year 1, the network being tested is reconstructed only using the features selected at year 1. This is undertaken to assess if the biological signal learnt at one time point is present at other time points.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Performance of Best vs. Baseline (MDS-UPDRS) models at each time point.

(A) Idiopathic & Genetic, (B) Genetic, (C) Idiopathic. The best model in each subgroup at each time point corresponds to the combination of modalities reported in Table 2. The performance of the MDS-UPDRS baseline models are reported in S1 File.

https://doi.org/10.1371/journal.pcbi.1012857.g001

Performance and evaluation

The performance metrics used to compare the classification performance of MOGDx were accuracy, F1 score and improvement in accuracy. The F1 score was calculated by the mean F1 score of each class, weighted by the size of that class. Improvement in accuracy is a metric used to compare how much the accuracy improved compared to a baseline model which only predicts the most common class. Stratified k-fold cross validation was performed with 5 randomly generated splits to obtain the mean and standard deviation metrics reported. Within each split, the training set was further randomly split into training and validation sets to produce an overall train/validation/test split of 68%/12%/20% respectively.

A multi-modal approach is optimal when stratifying individuals with PD over time

The optimal combination of modalities at each time point in each subgroup are reported in Table 2 and shown in Fig 1 as the best model. Confusion matrices are also provided in S3–S5 Figs for multi-modal performance assessment. The results from the cross-sectional experiments favour a multi-modal approach. The power of flexible integration when stratifying participants in the PPMI dataset with PD is apparent as all 6 modalities are included in at least one experiment. In Table 2, only two experiments, years 1 and 3 with all participants (genetic + idiopathic), do not integrate modalities. In Fig 1, performance in the combined genetic + idiopathic and in the idiopathic subgroup alone are significantly poorer than the MDS-UPDRS baseline model. Only the genetic subgroup achieves a consistent improvement in accuracy greater than the MDS-UPDRS assessment. This could motivate the use of these modalities for early disease diagnosis, as motivated below. The combination of CSF and DNAm in the idiopathic subgroup shows promising performance, particularly later in the disease course. Comparing to a model trained on the MDS-UPDRS assessment is a difficult task, given it consists of clinical assessment scores of both motor and non-motor symptoms [20]. Thus, the results show encouraging performance when integrating combinations of modalities in subgroups of PD.

Disease signatures for PD can be learnt from whole-blood samples and protein markers in PPMI study participants

Table 2 shows there is an improvement in accuracy compared to a model which predicts the most common class in all experiments. This improvement increases with time, indicating an increased biological signal for PD as the disease progresses. In both genetic and idiopathic subgroups, years 1 and 2 are the most predictive time points. This could indicate that these time points are capturing both early and late signatures of PD. This is particularly evident in the idiopathic subgroup, where there is a change in predictive modalities over time, with common modalities early and late in the disease. The SNPs modality is predictive early, whereas protein CSF markers along with DNAm are more prominent in later stages. This supports the work of Wüllner et al. (2023) who found that there may be different disease mechanisms at different stages of PD and highlights the necessity of flexible modality integration [1]. The caveat is that the prediction accuracy at year 0 in this subgroup is low.

Conversely, in the genetic subgroup, the modalities which are most predictive do not change with time. SNPs are a fixed description of participant’s genetic status [25] and the sub-population of participants that carry pathogenic mutations with varying penetrance in the PD and PL groups are likely to share other features across their genomes that influence the extent and nature of the onset of PD symptoms. The ability of SNPs derived signatures alone to distinguish between PD/PL and HC participants in our study but not between PD and PL participants supports this idea. We demonstrate that this further separation can be achieved by inclusion of an additional modality, in this case DNA methylation. Whether the biological signal learnt changes over time requires further work to understand the drivers of variability in DNAm at each time point. The integration of these two modalities outperforms the MDS-UPDRS baseline, highlighting the predictive power of using whole-blood samples to extract genomic information relating to a neurological disease. Overall, we show there are disease signatures early, but particularly late, in the blood of individuals with PD, despite it being a neurological disorder.

There are possible shared DNAm signatures in the idiopathic and genetic disease subgroups of PD

As per Table 2, there is a clear genetic driver in participants who have a mutation in a causal gene of PD. This highlights the homogeneity between participants in this subgroup. Conversely, the idiopathic cohort is heterogenous, as this group combines participants with unknown and likely different causes of PD. Despite these significant differences between participant subgroups, most experiments include DNAm as a predictive modality, with the idiopathic subgroup at year 0 being the only experiment where it is not included. Given this prominence, it indicates the presence of epigenetic modifications between PD, PL and HC participants. When considering all participants (genetic + idiopathic), there is a mix of a homogeneous and a heterogeneous group, which makes learning very difficult. Despite this, there is a robust improvement in accuracy between 10% and 20% across all time points, as per Fig 1. This suggests that there may be a shared signal between the two subgroups in the DNAm modality. A possible explanation for the decreased performance compared to the two subgroups is that the additional information added by other integrated modalities is not shared between the two subgroups. In order to confirm if the signal being learnt is similar, more research needs to be conducted to identify the common discriminating DNAm features, however our results suggests common signatures in the DNAm of genetic and idiopathic PD participants.

An integrative model trained at a late disease stage could form a viable early diagnostic tool for individuals with a genetic predisposition for the disease

In the cross-sectional experiments, we show the metrics for classifying participants with a mutation in a causal gene of PD to be very promising. As per Fig 1 and Table 2, the combination of DNAm and SNPs achieves a consistently high accuracy, F1 score and improvement in accuracy. These results motivate the integration of these modalities for early disease detection. We performed longitudinal experiments on a subset of participants from the genetic group who are present in either DNAm or SNPs at each time point. These experiments were designed to identify the optimal time point to train such a diagnostic tool and if the disease signal learnt early in the disease is present later and vice versa.

Table 3 highlights that an early PD detection model should be trained later in the disease course. Both the accuracy and F1 scores increase with models that are trained at later time points. Optimal performance was observed by the model trained at year 3. Poorest performance was observed by the model trained at year 0, with the performance of models trained at years 1 and 2 being comparable. S6–S9 Figs show the multi-modal performance of each model at each time point. These figures highlight the improved discrimination between PD and PL participants for models trained later in the disease course.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Longitudinal experiments performance metrics.

https://doi.org/10.1371/journal.pone.0313772.t003

In Fig 2 the within class accuracy is reported at each time point with interconnecting bands showing incorrect model predictions. We can see the model misclassifies more PD patients as PL at earlier time points, as denoted by the thickness of the red band (Fig 2A). We also assessed the longitudinal behaviour of the model to see whether it consistently identified the disease class of participants as PD, PL and HC throughout the time course (Fig 2B). There is a subset of ˜35% PD participants who are the most difficult for the model to classify. This subset only has either 1 or 2 correct classifications, whereas both the PL and HC classes maintain ˜90% accuracy at all time points. The HC class is the easiest to predict. As mentioned, both PD and PL participants have a genetic risk variant for the disease, thus, the SNPs modality can easily discriminate between them and the HC participants. The main differentiation between the models is their ability to distinguish PD from PL participants. As mentioned, the accuracy in predicting PD participants decays the further away in time you test the model from when it was trained. This can be observed both by the sharp gradients of the PD participants when assessing the number of consecutive correct predictions of the model and the decrease in accuracy over time in Fig 2A. Conversely, the PL class have much more stable and consistent predictions across all time points. This is evident in Fig 2A with the number of PL participants correctly classified being less variable over time and the flatter gradients in the consistency of predictions in Fig 2B.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Patient stratification for model trained at year three.

(A) Within class accuracy and predictions, with the interconnecting bands showing incorrect model predictions and the thickness of the bands reflecting the number of patients misclassified. (B) The percentage of times the model correctly classified an individual patient. The sharp decline in the PD class shows there is a subset of ˜65% of PD participants who are more easily identified at all time points.

https://doi.org/10.1371/journal.pcbi.1012857.g002

In Table 3, we show there is a much stronger signal discriminating PD from PL participants later in the disease course. This finding is expected as the PD participants, on average, will have a more severe disease at year 3 than they will at year 0. What these results therefore show is that by year 3 we have found a very accurate threshold for differentiating PD participants from PL. When we then back-propagate this threshold by testing the model over time, we find that the PL participants maintain a high predictive accuracy, but some PD participants cross this threshold and are misclassified as PL. As stated, differences between these groups can be largely explained by differences in their DNAm. Thus, we can attribute these findings to epigenetic modifications occurring in participants with PD as their disease progresses.