Machine learning image analysis predicts Parkinson disease genotype and mitochondrial lysosomal abnormalities in iPS neurons

Abstract

Mitochondrial and lysosomal dysfunction are central features of Parkinson’s disease (PD) across major genetic forms including PRKN, SNCA, and LRRK2. We applied cell morphomics, a machine-learning-based framework combining high-content imaging with quantitative feature extraction, to analyse mitochondrial and lysosomal morphology at single-cell resolution in iPS cell-derived cortical neurons from PD patients and healthy controls (13 lines total). Supervised machine-learning models distinguished PD neurons from controls with high accuracy (AUC = 0.87) and reliably separated individual genotypes. Feature importance and attribution analysis revealed genotype-specific organelle biases in the relative contribution of mitochondrial and lysosomal features, with mitochondrial features dominating classification in PRKN neurons, balanced mitochondrial and lysosomal contributions in SNCA neurons, and a greater lysosomal contribution in LRRK2 neurons. Multi-class models retained strong performance, and findings were reproduced across two independent laboratories using different dyes and imaging conditions. These results demonstrate that morphomics provides a robust and scalable framework to quantify genotype-specific organelle abnormalities in PD neurons and supports its application for cellular stratification and biomarker discovery.

Introduction

Parkinson’s disease (PD) is clinically heterogeneous, with patients carrying different genotypes presenting variability in age at onset, rate of progression, and severity of motor and non-motor symptoms¹. This heterogeneity is attributed to the downstream effects of genetic variation and their impact on cellular function. Mitochondrial and lysosomal dysfunction are consistently observed across the most common genetic forms of PD PRKN, SNCA, and LRRK2, with each mutation disrupting these cellular functions through distinct mechanisms². PRKN mutations are classically associated with defective mitophagy and impaired clearance of damaged mitochondria, which secondarily compromises lysosomal function. Dopaminergic neurons carrying PRKN mutations display reduced differentiation efficiency and neuritic complexity³, impaired mitochondrial respiration and membrane potential^4–6, and defective mitophagy under oxidative stress conditions⁷, together with enlarged lysosomes and disrupted mitochondria–lysosome contacts^8,9. SNCA mutations such as duplications and triplications are primarily associated with α-synuclein accumulation, that can initiate mitochondrial stress and imposes a secondary burden on lysosomal clearance. Dopaminergic neurons derived from SNCA lines show mitochondrial depolarisation, oxidative stress, and calcium dysregulation^10–12, followed by lysosomal enlargement, impaired trafficking, and reduced clearance of α-syn aggregates^11,13,14. LRRK2 mutations are linked to kinase-dependent defects in vesicular trafficking and lysosomal maturation, with secondary consequences on mitochondrial quality control and mitophagy. Neurons carrying LRRK2 mutations consistently exhibit mitochondrial dysfunction and reduced membrane potential^15–17, together with impaired autophagosome transport and defective lysosomal processing driven by aberrant kinase activity^18,19.

Given the consistent involvement of mitochondria and lysosomes across PD genotypes, we asked whether alterations in their morphology could be leveraged to classify and predict disease genotype and to determine if differences in the relative contribution of mitochondrial and lysosomal features could be detected across genotypes. To achieve this, we used high-content confocal imaging coupled with computational analysis, enabling the extraction of hundreds of quantitative features from thousands of single cells, generating multiparametric “morphomic” profiles that can be interrogated using machine learning to detect subtle multi-organelle phenotypes. Initial studies have provided proof-of-concept that cellular morphology encodes disease-relevant signatures. D’Sa et al profiled induced pluripotent stem (iPS) cell-derived cortical neurons and trained a machine-learning classifier on mitochondrial and lysosomal morphological features that distinguished PD subtypes including chemically induced complex I dysfunction (rotenone-induced), induced mitophagy (oligomycin/antimycin-induced), proteinopathy (α-synuclein oligomer-induced), and patient-derived SNCA triplication and PINK1 mutations²⁰. Although this study demonstrated that morphomics can resolve both genetic and mechanistic subtypes, its conclusions were limited by using only a single line per group, restricting generalisability.

In this study, we sought to overcome these limitations by applying a morphomics framework to a multi-donor iPS cell-derived neuronal cohort spanning three major PD genotypes PRKN, SNCA, and LRRK2, with three independent patient lines per genotype and four healthy controls (13 lines in total). Using a Cell Painting–style approach^21,22, we systematically profiled mitochondrial and lysosomal morphology at single-cell resolution. In total, 1320 quantitative features per cell, capturing intensity, texture, and spatial organisation, were extracted and used to train machine-learning classifiers to distinguish disease from control and between genotypes. To ensure robustness and generalisability, the workflow was validated independently across two laboratories using distinct mitochondrial stains and imaging conditions. This design enables not only accurate genotype classification but also quantitative attribution of the relative contribution of mitochondrial and lysosomal features across PD genotypes.

Results

Mitochondrial and lysosomal morphomic profiles distinguish PD from control cells

We first established a cell morphomics workflow to examine mitochondrial and lysosomal morphology at single-cell resolution (Fig. 1). Patient- and control- iPS cell-derived neurons were stained to visualise mitochondria and lysosomes and imaged using high-content confocal microscopy. Quantitative features were extracted from these images (Supplementary fig. 1) and used to train a machine-learning model to distinguish genetic PD samples from controls. A total of 1320 quantitative features were extracted per cell (66,061 single cells analysed across the study). To understand which organelle variables drove the predictions, we applied SHAP (SHapley Additive exPlanations), which ranks features according to their influence on the model.

Fig. 1. Cell morphomics workflow to identify Parkinson’s disease specific cellular signatures.

iPS cell-derived neurons were seeded onto 96-well plates and labelled with live-cell dyes to visualise mitochondria and lysosomes. High-content imaging was performed to capture cellular morphology. Quantitative morphomic features were extracted and used to train an AI-based classifier that predicts cell genotype and reveals relative mitochondrial and lysosomal feature contribution underlying disease pathology.

The developed classifier achieved an AUC of 0.87 (95% CI: 0.87–0.88) (Fig. 2A) with a balanced accuracy of ~79% (95% CI: 78–79%) (82% of control cells and 75% of PD cells correctly identified) (Fig. 2B). Performance was consistent across donors, indicating that the signal captures shared cellular changes. Misclassified cells may reflect overlap in morphology, with some control cells appearing stressed and some PD cells relatively preserved. SHAP-based feature attribution highlighted contributions from both mitochondria and lysosomes (Fig. 2C). While the top 10 features were evenly divided, weighted SHAP contributions showed mitochondrial features account for a larger share of the model’s explanatory signal ( ~ 59% vs 41% for lysosomal features) (Fig. 2D). These proportions reflect SHAP-based feature attribution within the trained model and do not represent standalone classification performance of individual organelles. These results confirm that morphomic analysis can robustly distinguish PD from control cells and implicates both organelles in the disease phenotype.

Fig. 2. Classification of Parkinson’s disease neurons and relative organelle feature contribution.

A ROC curve showing model performance in distinguishing Parkinson’s disease (PD) neurons from controls (AUC = 0.87, 95% CI 0.87–0.88) in the test dataset. B Confusion matrix illustrating model accuracy for PD and control classifications. C SHAP (SHapley Additive exPlanations) summary plot showing the top features contributing to model prediction, with mitochondrial and lysosomal intensity and texture parameters among the highest-ranked contributors. D Relative contribution of mitochondrial (59%) and lysosomal (41%) features to overall model performance, indicating that both organelle feature sets contribute substantially to disease classification.

To further interpret the organelle-level features driving classification, we examined the SHAP summary plot (Fig. 2C). In this plot, features positioned toward the right (positive SHAP values) increase the model’s predicted probability of PD, whereas features positioned toward the left (negative SHAP values) increase the predicted probability of control; point colour represents the relative feature value, ranging from low (blue) to high (red). The morphological features using the “cell-painting” style approach capture patterns of staining intensity, spatial distribution, and texture (i.e., differences in pixel intensity distribution), enabling quantitative description of cellular phenotypes without explicit organelle segmentation.

Analysis of the top SHAP-ranked features indicates that mitochondrial and lysosomal contributions to classification arise from distinct morphomic signatures, with features from the outer cellular region (peripheral cell region, furthest away from nucleus) contributing prominently for both organelles. Highly ranked mitochondrial features were predominantly intensity-based (e.g. Outer Region Minimum and Median), indicating alterations in how mitochondrial signal is distributed across the cell. In contrast, highly ranked lysosomal features were largely texture-based (e.g. SER and Haralick descriptors), again primarily derived from the outer cellular region, reflecting differences in the spatial organisation and pixel intensity distribution of lysosomal staining.

Genotype-specific separation and organelle attribution

The morphomics workflow was also able to distinguish each major genetic form of PD from controls. Classifier performance was strong across all comparisons, with AUC = 0.88 (95% CI: 0.87–0.89) for PRKN vs Control, 0.94 (95% CI: 0.93–0.95) for LRRK2 vs Control, and 0.90 (95% CI: 0.9–0.91) for SNCA vs Control (Fig. 3A). Balanced accuracy was also high 79% (95% CI: 78–80%) for PRKN vs Control, 87% (95% CI: 85–88%) for LRRK2 vs Control and 82% (95% CI: 81–83%) for SNCA vs Control, with 76-88% of PD cells correctly classified, showing that the cellular features captured by morphomics are sufficiently distinctive to resolve genotype-specific signatures (Fig. 3B).

Fig. 3. Genotype-specific classification and organelle feature attribution in Parkinson’s disease.

A ROC curves showing machine-learning model performance in distinguishing PRKN, LRRK2, and SNCA patient-derived neurons from controls in the test dataset. Classification accuracy was highest for LRRK2 (AUC = 0.94, 95% CI 0.93–0.95), followed by SNCA (AUC = 0.90, 95% CI 0.90–0.91) and PRKN (AUC = 0.88, 95% CI 0.87–0.89). B Confusion matrices illustrating prediction accuracy for each genotype relative to controls. C–E SHAP feature attribution plots and organelle contribution charts for PRKN (C), LRRK2 (D), and SNCA (E) neurons. Mitochondrial features were the dominant contributors to model prediction, accounting for 82%, 59%, and 76% of explanatory power in PRKN, LRRK2, and SNCA neurons, respectively, with lysosomal features contributing the remainder.

Feature attribution analysis revealed organelle-specific differences across genotypes (Fig. 3C-E). Weighted SHAP contributions showed that mitochondrial features accounted for the largest share in PRKN (82%), were moderate in SNCA (76%), and lowest in LRRK2 (59%). Lysosomal features followed the opposite pattern, contributing 18% in PRKN, 24% in SNCA, and 41% in LRRK2. Across all genotypes, the most informative features for both organelles were predominantly derived from the outer cellular region, defined as the cell area distal to the nucleus (Fig. 3C-E). PRKN cells showed a predominantly mitochondrial-associated signature, with highly ranked features reflecting changes in mitochondrial staining intensity in the outer cellular region, indicating altered mitochondrial distribution in the cell periphery. LRRK2 cells, in contrast, showed a relatively stronger lysosomal-associated signature, with top features enriched for lysosomal texture and intensity measures from the outer cellular region. SNCA cells displayed an intermediate profile, with both mitochondrial intensity-based features and lysosomal texture-based features from the outer cellular region contributing to classification, indicating concurrent alterations affecting both organelles.

Multi-modal classification across PD genotypes

We next evaluated a multi-class model to discriminate between PRKN, LRRK2, and SNCA genotypes. The classifier achieved AUCs of 0.87 (95% CI: 0.86–0.88) for PRKN vs LRRK2, 0.89 (95% CI: 0.88–0.9) for LRRK2 vs SNCA, and 0.82 (95% CI: 0.81–0.82) for PRKN vs SNCA (Fig. 4A). Per-class correct assignment was 65% (95% CI: 63–70%) for PRKN, 82% (95% CI: 79–84%) for LRRK2, and 60% for SNCA (95% CI: 59–62%) (Fig. 4B). Misclassifications occurred most frequently between PRKN and SNCA, whereas LRRK2 showed the highest correct classification rate. Weighted SHAP values in the multi-class setting showed contributions of 83% mitochondrial and 17% lysosomal for PRKN, 50% mitochondrial and 50% lysosomal for LRRK2, and 82% mitochondrial and 18% lysosomal for SNCA (Fig. 4C-E).

Fig. 4. Multi-modal classification and organelle feature attribution across Parkinson’s disease genotypes.

A ROC curves showing machine-learning based multi-class model performance in distinguishing PRKN, LRRK2, and SNCA neurons in the test dataset. Model separability was highest for LRRK2 vs SNCA (AUC = 0.89, 95% CI 0.88–0.90), followed by PRKN vs LRRK2 (AUC = 0.87, 95% CI 0.86–0.88) and PRKN vs SNCA (AUC = 0.82, 95% CI 0.81–0.82). B Confusion matrix summarising prediction accuracy across all genotypes. C–E SHAP feature attribution plots and organelle contribution charts for PRKN (C), LRRK2 (D), and SNCA (E) neurons compared to other genotypes. Mitochondrial features dominated classification performance for PRKN and SNCA contributing 83% and 82% of explanatory power for PRKN and SNCA, respectively, while lysosomal features were equal with mitochondrial features for LRRK2’s weighted SHAP explanations.

When compared with the binomial genotype-vs-control models (Fig. 3), only small to moderate shifts (1–9%) were observed in organelle contributions between the multi-class and binomial models. PRKN showed a slight increase in mitochondrial contribution (82–83%) with a corresponding decrease in lysosomal contribution (18–17%). LRRK2 showed a moderate decrease in mitochondrial contribution (59–50%) with a corresponding increase in lysosomal contribution (41–50%). SNCA showed a moderate increase in mitochondrial contribution (76–82%) and a corresponding decrease in lysosomal contribution (24–18%). Across both binomial and multi-class models, mitochondrial features accounted for the largest relative share of SHAP-based feature attribution for PRKN and SNCA, while LRRK2 showed balanced contributions for multiclass classification.

Robustness of morphomics across laboratories

To test the robustness of the assay, we compared morphomic predictions generated from two independent laboratories using different dyes and imaging conditions. Lab 1 used Mitotracker Red with 40× magnification, while Lab 2 used TMRM with 20× magnification. Despite these methodological differences, classifier performance was similar across the two sites.

In Lab 1, the classifier achieved an overall test accuracy of 66% (95% CI: 64–67%) with an average pairwise AUC of 0.83 (Fig. 5A). Pairwise AUCs were strong for PRKN vs LRRK2 (0.84, 95% CI: 0.82–0.85) and LRRK2 vs SNCA (0.86, 95% CI: 0.85–0.87), while PRKN vs SNCA was more challenging (0.79, 95% CI: 0.78–0.79). Correct classification rates were highest for LRRK2 (77%, 95% CI: 74–80%), followed by PRKN (63%, 95% CI: 61–64%), with SNCA showing greater misclassification (57% correctly classified, 95% CI: 55–59%).

Fig. 5. Cross-laboratory validation of multimodal genotype classification in Parkinson’s disease neurons.

A Lab 1: Mitotracker imaging at 40× magnification. ROC curves and confusion matrix showing multi-class classification performance for PRKN, LRRK2, and SNCA neurons in the test dataset. Model separability was highest for LRRK2 vs SNCA (AUC = 0.86, 95% CI 0.85–0.87), followed by PRKN vs LRRK2 (AUC = 0.84, 95% CI 0.82–0.85) and PRKN vs SNCA (AUC = 0.79, 95% CI 0.78–0.79). B Lab 2: TMRM imaging at 20× magnification. ROC curves and confusion matrix showing consistent classification trends in the test dataset across imaging modalities, with comparable performance for PRKN vs LRRK2 (AUC = 0.85, 95% CI 0.85–0.86), LRRK2 vs SNCA (AUC = 0.81, 95% CI 0.81–0.82), and PRKN vs SNCA (AUC = 0.81, 95% CI 0.80–0.81).

In Lab 2, similar results were obtained (Fig. 5B). The classifier reached an overall accuracy of 64% (95% CI: 63–65%), with pairwise AUCs of 0.85 (95% CI: 0.85–0.86) for PRKN vs LRRK2, 0.81 (95% CI: 0.81–0.82) for SNCA vs LRRK2, and 0.81 (95% CI: 0.81–0.82) for PRKN vs SNCA. Correct classification was again strongest for LRRK2 (68%, 95% CI: 66–69%) and PRKN (66%, 95% CI: 65–68%), while SNCA remained the least distinct (59%, 95% CI: 58–60%).

Together, these findings demonstrate that the morphomics workflow is robust across laboratories, remaining stable despite differences in staining dyes and imaging magnification. The reproducibility of predictions supports the generalisability of the assay and strengthens confidence in its biological readouts.

Discussion

This study establishes mitochondrial and lysosomal morphomics as a robust framework for distinguishing genetic PD patient cells from controls and for resolving genotype-specific differences. By combining high-content imaging, machine learning, and SHAP-based attribution, we demonstrate that disease- and genotype-specific profiles can be consistently detected at single-cell resolution, and that these signals can be attributed directly to mitochondria and lysosomes.

At the disease level, PD-derived neurons were reliably separated from controls, confirming that morphomic features capture convergent cellular alterations shared across patients. Both mitochondrial and lysosomal features contributed to classification; however, mitochondrial features consistently accounted for a greater proportion of the model’s explanatory signal based on SHAP attribution. This is consistent with known mitochondrial and lysosomal involvement in PD². At the genotype level, organelle-specific biases were apparent but broadly aligned with known biology: As shown by SHAP analysis in Fig. 3, PRKN cells were defined most strongly by mitochondrial descriptors, LRRK2 cells showed the greatest relative lysosomal involvement, and SNCA cells displayed a more intermediate profile. Extending beyond binary classification, multi-class modelling showed that genotypes can also be distinguished directly from one another, albeit with lower accuracy than in genotype-vs-control comparisons. Importantly, organelle-level contributions were largely stable across modelling frameworks, with only small (1-9%) shifts in mitochondrial versus lysosomal weighting. Notably, the highest misclassification rate in the multi-class model occurred between PRKN and SNCA neurons, consistent with overlap in their mitochondrial SHAP-weighted profiles (Fig. 4). One possible explanation is that PRKN and SNCA mutations, although acting through different upstream mechanisms, converge on mitochondrial dysfunction at the cellular level. Further functional studies will be required to define the mechanisms underlying this convergence and to determine whether these represent shared downstream pathways or distinct but overlapping mitochondrial phenotypes. To better resolve PD heterogeneity, particularly between PRKN and SNCA, incorporating additional cellular compartments, markers such as total and phosphorylated α-synuclein, or stress paradigms may enhance discriminatory power.

A recurring finding across all models was the prominence of features derived from the outer cellular region, defined as the cell area distal to the nucleus, for both mitochondrial and lysosomal classification. This observation is consistent with known principles of neuronal cell biology, where mitochondria and lysosomes are generated in the soma and trafficked anterogradely along microtubules to distal compartments. This transport process is ATP-dependent, linking efficient organelle distribution directly to mitochondrial function²³. As mitochondrial function declines, ATP availability for transport is reduced, which may impair both the delivery of newly formed organelles to distal regions and the clearance of damaged organelles. As a result, distal cellular compartments are likely to be affected earlier, consistent with the dominant contribution of peripheral-region features observed across all three genotypes. This interpretation is also consistent with the concept of “dying-back” neurodegeneration, in which distal axonal compartments are affected prior to overt cell body degeneration, as described in Parkinson’s disease and related neurodegenerative disorders^23–26. However, it is important to note that the mitochondrial and lysosomal features used in this study quantify intensity and texture within predefined cellular regions, rather than directly measuring organelle transport; hence, the biological interpretation of these spatial patterns remains indirect. Future studies incorporating direct approaches such as time-lapse imaging of organelle transport dynamics across neuronal compartments will be important to validate these findings. In line with the morphomic signatures, our earlier functional analyses of the same cortical neuron lines demonstrated that all three genotypes exhibited significantly reduced mitochondrial respiration, whereas lysosomal glucocerebrosidase activity was most severely impaired in LRRK2 neurons²⁷. The dominance of mitochondrial descriptors across prediction models, therefore, reflects a convergent mitochondrial deficit that was captured independently by oxygen consumption rate assays. Conversely, the enhanced contribution of lysosomal features to LRRK2 classification aligns with its pronounced lysosomal dysfunction, while the stronger mitochondrial contribution to PRKN classification suggests a genotype-associated morphomic pattern that was not evident in bulk respiration assays, which showed comparable reductions across genotypes. This suggests that the single-cell resolution of the morphomics approach is more sensitive to detecting genotype-specific mitochondrial alterations than population-level assays such as oxygen consumption rate measurements²⁸. This heightened sensitivity aligns with the established role of PRKN in maintaining mitochondrial quality control, making subtle mitochondrial defects in PRKN neurons more readily detectable through morphomic profiling.

Our results complement and extend prior iPS cell-based morphomic studies, where both mitochondria and lysosomes have been implicated as key contributors to PD phenotypes. Using profiled iPS cell-derived cortical neurons carrying SNCA triplication and PINK1 mutations, alongside chemically induced PD-like models including rotenone (complex I dysfunction), oligomycin/antimycin (induced mitophagy), and α-synuclein oligomers (proteinopathy), their classifier achieved an overall accuracy of 82% for a PD phenotype²⁰. SHAP analysis revealed that mitochondrial features, particularly TMRM-based texture descriptors, dominated prediction across most subtypes, whereas lysosomal features were most prominent in SNCA triplication²⁰. Similarly, Vuidel et al. reported that LRRK2 G2019S dopaminergic neurons display elevated α-synuclein levels, reduced dendritic complexity, abnormal mitochondrial morphology (reduced membrane potential and increased compactness), and altered lysosomal features²⁹. Using machine-learning classifiers, these cells were separated from isogenic and unrelated controls, and drug interventions (LRRK2 inhibitors, PKC agonists) shifted morphologies towards controls. While these studies established proof-of-concept that morphomics can detect genotype-specific phenotypes, they relied on only one or two patient lines per genotype, limiting their generalisability. In contrast, our analysis of 13 iPS cell-derived neuronal lines across PRKN, LRRK2, SNCA, and matched controls demonstrates that organelle-specific signatures are robust across diverse donors and more sensitive than bulk functional assays²⁷. Moreover, by validating morphomic results in two independent laboratories using different mitochondrial dyes and imaging conditions, we show that morphomic predictions are stable despite methodological variation. Together, these advances move morphomics beyond proof-of-concept to a reproducible and scalable framework.

Morphomic approaches have also been applied beyond neurons. Schiff et al. (2022) profiled fibroblasts from 46 PD patients (sporadic, LRRK2, GBA) and 45 controls using a Cell Painting assay ( > 3000 features across multiple organelle and cellular compartments)³⁰. PD fibroblasts were distinguished from controls with AUCs of 0.76–0.89, and donor-specific morphologies remained stable across biopsies collected 3-6 years apart, indicating persistent, reproducible signatures. Antony et al. (2020) analysed fibroblasts from 41 idiopathic PD patients and 21 controls and found increased mitochondrial fragmentation and reduced branching; classification improved under FCCP-induced stress (AUC 0.87 vs 0.83), showing how stress can unmask latent defects³¹. These fibroblast studies demonstrate the broad applicability of morphomics and the stability of donor-specific profiles. However, fibroblasts provide only an indirect view of PD. By focusing on iPS cell-derived neurons, our study interrogates disease-relevant cell types and links genotype-specific differences directly to mitochondrial and lysosomal alterations at single-cell resolution.

Notably, while each genotype displayed distinct organelle-level biases, the presence of both mitochondrial and lysosomal alterations across all groups reinforces the concept of shared pathogenic pathways in PD. As discussed in the introduction and supported by our findings, these genotypes converge on similar organellar dysfunctions but differ in the relative extent to which each pathway is affected. Identifying the dominant contributor within each genotype, whether mitochondrial or lysosomal, may therefore provide a rational basis for selecting and testing pathway-specific therapeutic interventions. For example, compounds enhancing mitophagy or mitochondrial biogenesis may be particularly effective in PRKN-related disease, whereas strategies targeting lysosomal function or autophagic flux may hold greater promise for LRRK2- or SNCA-related disease. Combination approaches targeting both pathways may further enhance therapeutic efficacy. Moreover, this morphomics-based framework could offer a path to identifying cellular pathology in idiopathic patients lacking defined genetic information, thereby extending cell pathology targeted strategies to the majority of PD cases. This framework could guide drug screening and precision therapeutic design by aligning molecular pathology with targeted rescue mechanisms.

Our work has certain limitations. An important consideration is that this study focused specifically on mitochondrial and lysosomal organelles. Accordingly, the contributions reported here reflect the relative contribution of features within these compartments, rather than across all or most cellular structures. Additional cellular structures, including the cytoskeleton, endoplasmic reticulum, Golgi apparatus, synaptic compartments, and proteinopathy-related markers such as α-synuclein, were not included in the current analysis and may provide additional discriminatory information. Expanding this framework to incorporate these components will be an important next step.

In addition, SNCA neurons showed a relatively greater mitochondrial than lysosomal feature contribution in the present analysis, despite the recognised importance of lysosomal dysfunction in α-synuclein-related pathology³². However, this is consistent with our previous functional analyses²⁷, where SNCA cortical neurons exhibited reduced mitochondrial respiration together with significant impairment of autophagy (increased p62), while direct lysosome functional deficits (glucocerebrosidase activity) were not prominently altered at baseline. Future studies using lysosomal perturbation or α-synuclein challenge will be important to determine whether lysosomal signatures can be further accentuated.

A further consideration is the use of cortical neurons in this study. While Parkinson’s disease is classically associated with the selective vulnerability of dopaminergic neurons, key pathological processes, including mitochondrial dysfunction, lysosomal impairment, and protein aggregation are not restricted to this neuronal subtype and are observed across multiple neuronal populations, including cortical neurons. Because the present study focused on these shared disease-relevant cellular pathways, cortical neurons provided a suitable system for morphomic analysis. We recognise, however, that dopaminergic neurons represent the most disease-relevant neuronal subtype for modelling selective vulnerability in Parkinson’s disease. Recent advances in differentiation protocols now enable the rapid generation of highly enriched functional human iPSC-derived dopaminergic neurons, with one protocol reporting approximately 80% TH-positive cells by day 18³³. In parallel, CRISPR/Cas9-based TH-Cre knock-in reporter hPSC lines provide a strategy to identify, isolate and analyse live TH-expressing dopaminergic neurons³⁴. Therefore, extending this morphomic framework to dopaminergic neurons will be an important future direction to determine how the signatures identified here relate to the selective neuronal vulnerability that characterises PD.

An additional limitation is that our morphomic analysis was restricted to proximal processes, excluding distal neuronal regions. This reflects the high density and complexity of neuronal cultures, which limit accurate segmentation of single neurons. As a consequence, alterations in distal processes may not be fully captured. Improved segmentation approaches in the future may enable more comprehensive assessment of distal phenotypes.

In summary, this study advances morphomics from proof-of-concept to a reproducible and generalisable platform for PD research in iPS cell-derived neurons. By leveraging a large multi-donor cohort, cross-laboratory validation, and organelle-level feature attribution, we establish a scalable framework with potential applications in patient stratification, biomarker discovery, and drug development.

Methods

Cell lines

All 13 iPSC lines used in this study, including LRRK2 R1441G (n = 3), SNCA A53T (n = 3), PRKN loss-of-function mutations (n = 3), and neurologically normal controls (n = 4), have been described previously in our recent publication²⁷ LRRK2 R1441G (n = 3), SNCA A53T (n = 3), and two neurologically normal control lines (n = 2) were obtained from the Golub Capital iPSC Parkinson’s Progression Markers Initiative (PPMI) sub-study (https://www.ppmi-info.org/access-data-specimens/request-cell-lines). The control line KOLF2.1 J was obtained from the Jackson Laboratory (#JIPSC1000) and has been described previously³⁵. The in-house control line RM3.5 has also been described previously^36,37. PRKN mutation iPSC lines (n = 3) were reprogrammed in house from patient-derived fibroblasts, with the characterisation of one iPSC line reported previously³⁷.

All research involving human participants, human material, and human iPSC lines was performed in accordance with the Declaration of Helsinki. Experiments involving human iPSCs were approved by the Human Research Ethics Committee of the University of Sydney (approval number: 2017/094) and the Human Research Ethics Committee of the University of Melbourne (approval numbers: 2020-1956049 and 2025-33900-73031-3). Written informed consent to participate in the study and for derivation and use of biological samples for research was obtained from all participants.

Cortical neuron differentiation

Cortical neural progenitors were generated from iPS cells according to the methods of³⁸. At day 25 post differentiation, vials of cortical progenitors were cryopreserved and retained in vapour phase liquid nitrogen until required. Neural progenitor cells were then thawed in the presence of 10 µM Y27632 (Stem Cell Technologies, Cat#72304) and plated into PhenoPlate 96 well plates (Perkin Elmer) pre-coated with 15 µg/ml Poly-L-ornithine (Sigma, Cat# P4957) and 10 µg/ml Mouse Laminin (Sigma, Cat# L2020) at a density of 10,000 cells per well. Cells were initially cultured in cortical base media comprised of a 1:1 mix of Dulbecco’s Modified Eagle Medium F12 (DMEM/F12) and Neurobasal media containing 0.5x B27 supplement, 0.5x N2 supplement, 0.5x ITS-A, 0.5x non-essential amino acids, 0.5x glutamax, 50 U/ml penicillin/streptomycin and 49.5 µM β-mercaptoethanol (all from Thermofisher). The Y27632 was removed from culture media at day1 and replaced by the cortical base media only. From day3, the media was changed to cortical maturation media comprised of a 1:1 mix of DMEM/F12 and Neurobasal media containing 1x B27 supplement, 1x N2 supplement, 1x ITS-A, 1x non-essential amino acids, 0.5x glutamax and 50 U/ml penicillin/streptomycin (all from Thermofisher) and further supplemented with 40 ng/ml BDNF, 40 ng/ml GDNF (both from Stem Cell Technologies), 50 µM Dibutyryl cAMP (Stem Cell Technologies), 200 nM ascorbic acid (Sigma), 100 ng/ml Mouse Laminin (Sigma) and 10 µM of the γ-secretase inhibitor DAPT (Sigma). Cells were differentiated for 12 days in cortical maturation media, with 90% of the spent media replaced every second day with fresh media. At day 15 post-thaw, a total of 40 days of differentiation, the cells were used for experiments.

Immunocytochemistry for characterization

Differentiated cortical cells from all lines were stained with COUP-TF-interacting protein 2 (CTIP2) to identify layer V cells, T-box brain protein 1 (TBR1) to identify deep-layer V and VI cells, together with microtubule-associated protein 2 (MAP2) as a general neuronal marker. Cells were stained with antibodies against MAP2, TBR1 and CTIP2 to identify mature cortical neurons. Immunostaining was performed using the CytoFix/CytoPerm fixation and permeabilization kit (BD Biosciences) as previously described (PMID 34404843). Briefly, cells were fixed in Cytofix for 20 min, washed three times (5 min each), and permeabilised in Cytoperm buffer for 30 min. Primary antibodies (Supplementary Table 1) were applied for 1 h, followed by three washes (5 min each) and incubation with secondary antibodies (Supplementary Table 2) for 30 min. Hoechst nuclear dye (ThermoFisher) was included during the final 10 min of secondary antibody incubation. After three additional washes, plates were sealed with parafilm and stored at 4 °C in the dark until imaging. One-way ANOVA comparing each genotype to control did not reveal any statistically significant differences in the expression of TBR1 and CTIP2 markers (Supplementary fig. 2).

Live-cell assays

Live-cell imaging assays were performed to assess mitochondrial and lysosomal phenotypes. 96 well-plates were imaged on the Opera Phenix high-content imager (PerkinElmer). Working dilutions of live-cell reagents were prepared in culture media and added to wells after removal of old media. Details of all reagents used are mentioned in Supplementary Table 3

For the mitochondrial and lysosomal assays described in Figs. 1–4, cells were incubated with Mitotracker Deep Red (Thermo Fisher, M22426), lysosomal staining reagent Cytopainter (Abcam, ab176827) and Cytopainter Green (abcam, 138891) dyes for 60 min. Staining with Hoechst 33342 (Thermo Scientific, 62249) to label nuclei was then performed for 10 min and cells washed with media and imaged immediately.

For mitochondrial TMRM phenotyping described in Fig. 5, cells were labelled with Calcein (Invitrogen, C3100MP) to identify viable cells, TMRM (Invitrogen, T-668) to assess mitochondrial membrane potential, and Hoechst 33342 (Thermo Scientific, 62249). Cells were incubated with probes for 30 min, washed twice with media, and imaged immediately.

High content imaging

Images were acquired on the Opera Phenix PLUS high-content screening system (Revvity) using either 20× or 40× water-immersion Zeiss Plan Apochromat objectives (20×, NA 1.0, WD 1.7 mm; 40×, NA 1.1, WD 0.62 mm). Detection was performed using a monochrome sCMOS camera (2160 × 2160 pixels, 6.5 μm pixel size, quantum efficiency >80%, read noise <1 e⁻, dynamic range ~33,000:1). The system was equipped with a microlens-enhanced dual-layer spinning disk Nipkow confocal device (50 μm pinhole size, 555 μm inter-pinhole distance), dual excitation paths to minimise cross-talk, and laser-based illumination. All images were acquired in 16-bit TIFF format.

Image acquisition and downstream analysis were performed using Harmony software (v5.2, Revvity). For each experiment, details including objective magnification, average number of cells imaged, fluorescence channels, number of fields of view, Z-step size, and number of planes per Z-stack are provided in Supplementary Table 4. For live-cell imaging assays, the number of Z-steps per stack was minimised to enable completion of each experiment within 30 min, ensuring consistency across experimental conditions.

Image data extraction

Image segmentation and feature extraction were performed using Harmony High-Content Analysis software version 5.2 (Revvity) following a Cell Painting style workflow²¹. Nuclei were segmented from the Hoechst channel, and cytoplasmic boundaries were defined using Calcein (for live-cell assays) or Cytopainter (for fixed-cell assays). To reduce segmentation errors from overlapping neuronal processes, cytoplasmic masks were restricted to include only the soma and proximal neurites. Feature extraction was performed on the mitochondrial (Mitotracker or TMRM) and lysosomal (Lysotracker Deep Red) channels using the “Calculate Cell Painting Properties” module with a standard SER scale of 1. Each organelle contributed 660 single-cell features, grouped into four main categories: intensity (mean, SD, quantiles, contrast), morphology (area, shape, length, perimeter), texture (SER, Haralick, and Gabor descriptors capturing spot, edge, and granularity patterns), and spatial (radial distribution, symmetry, compactness, and perinuclear enrichment). Features were quantified across five concentric cellular regions - nuclear, inner, middle, cell, and outer capturing the radial organisation of organelles (Supplementary fig. 3). Border-touching objects were excluded, and the resulting datasets were exported as tab-delimited text files with embedded metadata (cell line, plate, and batch) for downstream machine-learning analysis.

Data analysis

Highly correlated features (Spearman’s ρ > 0.9) were removed, reducing the dataset from 1320 features to 627 (325 Lysosomal, 302 Mitochondrial). Cells with missing feature values and wells with cell counts below 200 were excluded from analysis. In total 9% and 18% of cells were removed from laboratories 1 and 2, respectively, due to missing values and low well cell counts

The dataset was then split into training (60%), validation (20%) and test (20%) datasets. To account for batch effects features were scaled to training set controls in each batch separately using a standard scaler. To maintain class balance for training the classification models, the training dataset was randomly down sampled to the minority class for each prediction model (Supplementary Table 5). ML predictions were made using a Light Gradient-Boosting Machine (LGBM) classifier with Python using the LightGBM package. Classifiers were trained in the training datasets and n_estimators were optimised in the validation set by enabling 20 rounds of early stopping, using the balanced log loss in the validation sets as the evaluation metric. All other LGBM hyperparameters used default values. The final results are the predictions in the test datasets. Graphs were generated in GraphPad Prism and Python. Multiclass receiver operator curves were generated using the interpolated average of the receiver operator curves for a given pairwise comparison. AUC-ROC and accuracy confidence intervals were calculated using bootstrapping resampling.

SHAP-weighted organelle attribution

To determine the relative contribution of mitochondria and lysosomes to classification, we analysed SHAP (SHapley Additive exPlanations) values from each trained model. SHAP values were calculated on class balanced datasets. For each comparison (e.g., controls vs. PD, controls vs. PRKN), we extracted the absolute mean SHAP value for all features. Features were ranked by SHAP values, and the top 10 features were used for organelle-level attribution.

The weighted contribution of each organelle was calculated as the sum of SHAP values for features belonging to that organelle, divided by the total SHAP values across the top 10 features. Mitochondrial and lysosomal contributions, therefore, represent the proportion of explanatory power attributable to features from each compartment. This approach accounts for both the number of features and their weighted influence on the classifier’s decision, in contrast to simple feature counts.

Sources of variation Variation in the cell morphomics datasets can arise from multiple sources, including biological replicates (cell lines), experimental batches and technical replicates (wells within plates). Biological variability was addressed by including multiple independent cell lines per group (n = 3 per PD genotype and n = 4 controls). Technical variability was addressed by including multiple wells per cell line (n = 3 wells per line per plate). Experimental batch effects were captured by including multiple independent differentiation batches (n = 3). To account for batch effects, features were scaled relative to control wells within each batch using a standard scaler applied separately to each batch. To determine the generalisability of morphomics workflow to cell lines unseen during classifier training, a leave one cell line out cross validation was performed using the Control vs PD machine-learning model. Mean prediction probability for each cell line produced an AUC 0.81 for Controls (4 cell lines) vs PD (9 cell lines). These results indicate that the morphomic machine learning framework retains predictive capability in unseen cell lines, although further validation in larger cohorts will be required to confirm robustness and explore genotype specific predictions (Supplementary fig. 4).

Author contributions

Conceived the idea and secured funding-D.K., G.M.H., C.S., L.T., C.P., J.J. Planned the experiments-G.W., N.D., C.S. Generation and Q.C. of iPS cell lines-D.O., R.D., D.S., C.S., C.P. Performed the experimental work-Y.L., J.C., G.W., A.L.G., M.Z., D.A.B., C.P., T.F. Data analysis and figures-G.W., M.P., R.S., N.D., J.C. Drafted the manuscript-G.W., M.P., Y.L., N.D., C.S., G.M.H. Edited and approved the final version-all authors.

Data availability

The data, protocols, and key lab materials used and generated in this study are listed in a Key Resources Table alongside their persistent identifiers. All data and analysis code generated in this study and the Key Resources Table are publicly available via Zenodo repository: [10.5281/zenodo.18323330](https:/doi.org/10.5281/zenodo.18323330). The Zenodo repository includes single-cell feature datasets derived from high-content imaging, provided as spreadsheet files. These include raw feature tables from individual experimental batches and laboratories, as well as batch-normalised and combined datasets used for downstream analysis. Data preprocessing, statistical analysis, machine-learning classification, and SHAP-based feature attribution were performed using Jupyter Notebook scripts deposited in the same Zenodo record. A README file is included to describe the contents of the repository, file structure, and the purpose of each dataset and analysis notebook. A feature name reference file is also provided to map original feature names generated by the imaging software to the harmonised feature names used in this study. Raw microscopy image files are stored separately due to their size and are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41531-026-01458-w.