Machine Learning–Based Classification of Active and Latent Phases of Inherited Retinal Dystrophies Using Synthetic Proteomic Data: A Pathway-Based Application Exercise

Abstract

Objectives

Inherited retinal dystrophies are characterized by high genetic and phenotypic heterogeneity, and their clinical progression may alternate between latent and active phases. Identifying the onset of the active phase may support earlier intervention for inflammatory retinal degeneration. Plasma proteomics has shown potential for characterizing predictive biomarkers in retinal diseases, but its application remains experimental. This study aimed to develop a methodological simulation exercise to evaluate the performance of machine learning (ML) models in distinguishing active and latent phases using an artificially generated dataset.

Methods

An artificial dataset of 500 samples was created, and plasma proteomic profiles were generated for each sample using arbitrary values. Sample classification was based on a pathway activation score. Four ML models were tested: support vector machine, random forest, logistic regression, and extreme gradient boosting. Each model was trained across a range of hyperparameters.

Results

Logistic regression achieved the best performance, with an accuracy of 0.73, precision of 0.73, and F1-score of 0.73.

Conclusions

This simulation study showed that synthetic proteomic datasets can be used to evaluate ML approaches for distinguishing active and latent phases of retinal dystrophies when real data are scarce. Synthetic data can support the creation of targeted datasets for proteins associated with retinal dystrophies, helping to address the limited availability of suitable open-source data.

I. Introduction

Inherited retinal dystrophies (IRDs) represent a highly heterogeneous group of diseases characterized by progressive degeneration of photoreceptors and the retinal pigment epithelium (RPE). More than 250 mutated genes involved in these diseases have been identified, and these genes are responsible for conditions such as retinitis pigmentosa, Stargardt disease, and Usher syndrome [1]. However, the clinical course can be modulated by environmental and metabolic factors, including light exposure, chronic inflammation, oxidative stress, latent viral infections, excessive caloric intake, and micronutrient deficiencies [2–4]. For instance, exposure to high-energy blue light accelerates photoreceptor damage in patients with ABCA4 mutations [5], whereas chronic neuroinflammation can amplify retinal injury through persistent activation of the complement system [6–13]. Several experimental and theoretical models describe IRD progression as a process that can be conceptualized as alternating latent and active phases, in which degeneration accelerates after critical molecular thresholds regulating retinal homeostasis are exceeded. According to this model, IRDs are initially characterized by a latent phase resulting from genetic or environmental damage and subsequently progress to a secondary active phase, during which prolonged oxidative stress progressively impairs photoreceptors [14,15]. During the latent phase, cellular compensatory mechanisms preserve visual function despite ongoing genetic and metabolic damage. The retina activates protective strategies, including the secretion of neurotrophic factors such as BDNF and CNTF, which promote photoreceptor survival and Müller cell metabolism [16–19]. Inflammation is another important factor in IRD progression. Activation of the complement cascade, mediated by C3a and C5a, accelerates retinal damage and contributes to neuronal degeneration [6]. One of the most important events in the active phase is breakdown of the blood–retinal barrier, which increases vascular permeability, contributes to the onset of cystoid macular edema, and accelerates decline in visual acuity [20–22]. Finally, lipofuscin accumulation in the RPE promotes the release of inflammatory cytokines, such as IL-1β and TNF-α, increasing blood–retinal barrier permeability and facilitating transition to the active phase of the disease [22]. Plasma proteomics has been explored in retinal diseases as a means of identifying systemic molecular alterations associated with disease progression. However, variability in cohort size, disease heterogeneity, and data dimensionality complicates robust predictive modeling [23–26]. IRD progression has been mathematically modeled through several approaches that integrate molecular, genetic, and epidemiological data [27–29]. Computational analyses of retinal degeneration in retinitis pigmentosa have led to models that characterize the spatiotemporal dynamics of photoreceptor loss and the underlying metabolic interactions. Roberts et al. [27] proposed a partial differential equation–based model to describe the diffusion of photoreceptor degeneration by integrating retinal oxygen concentration with cell density and representing nonlinear dynamics with sigmoidal reaction terms. Similarly, Camacho et al. [29] developed a multiscale ordinary differential equation–based model to describe metabolic coupling between aerobic glycolysis and oxidative stress in cones. Their use of Hill functions with cooperativity coefficients enabled the identification of bistable metabolic states, providing a possible explanation for sudden transitions between quiescence and metabolic collapse. Experimentally, Collin et al. [28] analyzed an extensive set of monogenic murine models using nonlinear regression methods to identify degeneration kinetics characterized by latent phases followed by exponential decline, revealing sequential activation of distinct apoptotic pathways. Building on these examples from the literature, this study aimed to develop a machine learning (ML)-based classification model to differentiate between the active and latent phases of IRDs using synthetic proteomic data designed to approximate the type of information expected from real proteomic datasets. Because the progression of these diseases has been characterized as involving alternating phases, it is methodologically relevant to explore whether computational models can distinguish between predefined disease states under controlled conditions. Our framework incorporated variables representing proteomic features mapped to predefined pathway structures. This preliminary theoretical modeling exercise should be applied to real-world data to assess its applicability and to support future characterization of the progressive states of the disease. Synthetic data enabled the creation of a targeted dataset containing proteins known to be associated with retinal dystrophies. Such datasets are often difficult to identify in open-source repositories with a sufficiently large volume of high-quality data. By simulating testing and training datasets, investigators can iteratively evaluate the performance of diverse ML models across a broad parameter space that would be computationally or ethically impractical with real-world observations. This approach facilitates identification of objective functions and model architectures with the highest predictive accuracy within the synthetic framework. Consequently, this preliminary optimization may support more efficient deployment of refined algorithms on empirical proteomic data while minimizing the risk of overfitting to noise.

II. Methods

1. Generation of the Artificial Dataset

The experimental framework used a synthetic dataset of 500 samples to simulate the topological characteristics of clinical proteomic profiles, with each observation mapped as a discrete point within a 317-dimensional Euclidean space. Demographic covariates, including age and sex, were generated as independent variables; age followed a continuous uniform distribution across the interval from 40 to 60, and sex was represented as a Bernoulli-distributed discrete variable. The biochemical feature space was defined by a 315-dimensional manifold of protein abundance levels, based on publicly available resources (https://retnet.org/). Protein abundance values were generated using random sampling strategies inspired by concentration ranges reported in public proteomic databases, including plasma protein measurements available in the Human Protein Atlas (https://www.proteinatlas.org/). These databases were used solely as qualitative references for the typical order of magnitude and distribution shape of proteomic concentrations obtained by mass spectrometry. The resulting values did not correspond to real physiological concentrations and cannot be directly interpreted as biologically meaningful. Because the dataset was entirely synthetic and based on random generation, its purpose was limited to providing a high-dimensional numerical input space with dimensionality and numerical variability comparable to those of proteomic datasets commonly used in computational analyses. The synthetic dataset was generated using Python 3.8. Random sampling and numerical operations were performed using NumPy functions, and data structuring and export were handled with pandas DataFrame functions.

2. Classification of Proteins in Biological Pathways

To select representative proteins and approximate missingness in a real dataset, only proteins with an expression frequency of at least 50% across samples were considered in subsequent analysis steps. The selected proteins were categorized into ten major biological pathways associated with the pathogenesis of retinal dystrophies on the basis of literature and database searches: inflammation, oxidative stress, immunity, apoptosis, aging, visual perception, ciliopathies, lipid metabolism, cellular transport, and signal transduction (https://string-db.org/) (Table 1).

Table 1.

Assignment of selected proteins to pathway categories within the synthetic classification framework

Pathway	Proteins
Inflammation	COX2, CFB, CFD, CFHR1, TIMP3, VEGFA, LTB4R, LTE4
Oxidative stress	APOD, APOE, ARSG, HMCN1
Immunity	C2, C3
Apoptosis	CAPN5, TIMP3, RB1
Aging	APOE, ARMS2, PRPF4, GRK1
Visual perception	CNGB3, GRK1, GUCY2D, PDE6B, PDE6C, RGS9BP, SAG, RGR, RP1L1, RPGR, NRL, AIPL1, CABP4, BEST1, CLRN1, RLBP1
Ciliopathy	ARL6, BBS10, BBS2, CC2D2A, CLUAP1, IFT27, NPHP1, POC1B, RPGR, TTC8, USH2A, ALMS1, ARL2BP, CLN3, PCARE, PDZD7, WDPCP, C8orf37
Lipid metabolism	ABCC6, ACBD5, APOD, APOE, LRP5, PEX7
Cellular transport	ACBD5, AFG3L2, CHM, CLCC1, CNNM4, CYSLT1R, ESPN, HGSNAT, LCA5, LRAT, LZTFL1, MYO7A, PDE6B, PDE6C, SDCCAG8, TUBGCP4, TUBGCP6, ITM2B, KLHL7, REEP6
Signal transduction	NEK2, PRDM13, TSPAN12

3. Calculation of Pathway Scores

To estimate pathway-level activity, a pathway score was calculated for each sample as the weighted average of the expression levels of proteins belonging to each pathway. This approach aggregated individual protein measurements into pathway-level indicators that summarized the overall activity of biologically related protein groups. After pathway scores were calculated, Min–Max normalization was applied to scale the scores to a range from 0 to 1. This normalization step improved comparability across pathways and samples by reducing scale differences among the computed pathway activity values.

4. Annotation of Samples into Active and Latent Disease Phases

Pathway activity states were determined using a threshold-based approach. Pathways with normalized scores ≥0.70 were classified as active, whereas pathways with scores <0.70 were classified as latent. This classification was applied independently to each of the ten pathways for every sample. Sample-level binary labels were then assigned using a rule-based aggregation scheme. A sample was labeled active if two or more pathways were classified as active, whereas samples with zero or one active pathway were labeled latent. The resulting dataset comprised 82 protein features per sample and a binary label derived from this pathway scoring procedure. Among the 500 samples, 262 were classified as active-phase samples and 238 as latent-phase samples. Because both the data and labels were synthetically generated, these activity states represent computational categories within the proposed framework rather than clinical disease phases.

5. Development of the Machine Learning Model

To classify samples into active or latent phases, several ML algorithms were implemented using Python. The final dataset was divided into a training set (80%) and a test set (20%) for model validation. Before training, the data were scaled using MinMaxScaler to ensure uniform numerical scaling of the variables. Unlike standardization, this method preserves the relative distances of the original data distribution and can be useful for non-Gaussian datasets and sparse structures. MinMaxScaler may also enhance numerical stability by aligning input data with the numerical ranges expected by some algorithms. Four ML models were tested: support vector machine (SVM), random forest (RF), logistic regression (LR), and extreme gradient boosting (XGBoost). The best model was selected by comparing standard performance metrics, including accuracy, precision, recall, sensitivity, specificity, and F1-score, with parameters optimized through cross-validation. RandomizedSearchCV with 5-fold cross-validation was used for hyperparameter tuning. Additional details are provided in Supplementary Table S2. Based on the findings of Schroder et al. [30], using fewer hyperparameters is justified because it minimizes meta-over-fitting. Reducing the search space can lower the risk of selection-based overfitting, in which a model appears to perform well on validation data because of random variation.

III. Results

1. Exploratory Data Analysis

Although the dataset was synthetically generated, exploratory data analysis (EDA) was conducted as a quality-control step to verify the statistical properties of the simulated data and ensure that no systematic biases were introduced during data generation. EDA enabled assessment of variable distributions, dispersion patterns, and pairwise correlations within the synthetic proteomic matrix (Supplementary Figures S1–S4). Supplementary Table S1 summarizes the expression frequency, mean, and standard deviation for the most represented simulated proteins. Across the 315 variables, a broad range of values was observed, confirming that the simulation process produced heterogeneous distributions rather than uniform or degenerate patterns. The distribution of simulated spectral counts for selected proteins is shown in Figure 1. EDA was performed to confirm that the synthetic dataset exhibited sufficient variability and structure to allow meaningful evaluation of ML behavior under controlled conditions, without implying biological relevance.

Figure 1.

Distribution of simulated protein spectral counts. Proteins are shown on the x-axis, and their expression levels are shown on the y-axis.

2. Performance of Machine Learning Models

Table 2 reports model performance in terms of accuracy, precision, recall, and F1-score. Precision quantifies the proportion of positive predictions that are correct and therefore reflects the burden of false-positive predictions. In contrast, recall assesses model sensitivity by capturing the proportion of relevant instances correctly identified. The F1-score combines precision and recall through their harmonic mean, providing a summary metric that penalizes imbalances between these two measures. The LR model performed best across all measures, achieving an F1-score of 0.73, accuracy of 0.73, and precision of 0.73. Both SVM and XGBoost achieved an accuracy of 0.70 and showed similar performance across the other metrics. RF yielded the lowest performance, with an accuracy of 0.64 and an F1-score of 0.64. Among the models, SVM and XGBoost had higher sensitivity (0.75), indicating a stronger ability to identify positive instances, but lower specificity (0.64), suggesting more frequent false-positive predictions. In contrast, LR maintained balanced sensitivity and specificity values of 0.73 each, supporting more even identification of positive and negative cases. The area under the curve (AUC) was calculated and is discussed in Supplementary Figure S8.

Table 2.

Performance of the machine learning models

Model	Accuracy	Precision	Recall	F1-score	Sensitivity	Specificity	AUC
SVM	0.70	0.70	0.69	0.70	0.75	0.64	0.77
RF	0.64	0.64	0.64	0.64	0.64	0.64	0.72
LR	0.73	0.73	0.73	0.73	0.73	0.73	0.78
XGBoost	0.70	0.70	0.69	0.70	0.75	0.64	0.76

AUC: area under the curve, SVM: support vector machine, RF: random forest, LR: logistic regression, XGBoost: extreme gradient boosting.

3. Error Analysis and Confusion Matrix

Figure 2 shows the confusion matrix for the LR model on the test set of 100 samples. The model correctly identified 40 samples as active-phase and 33 samples as latent-phase. However, 15 false negatives and 12 false positives were observed. The confusion matrices for the other three tested models are provided in Supplementary Figures S5–S7.

Figure 2.

Confusion matrix for the logistic regression (LR) model.

4. Feature Contribution Analysis within the Synthetic Pathway-based Classification Framework

Feature importance analysis was performed to examine how individual protein variables influenced ML classification within the synthetic framework. The top contributing features identified by the LR model are presented in Figure 3. These importance scores indicate the relative influence of specific variables on the model’s decision-making process when distinguishing between active and latent states. Because the dataset was synthetically generated and class labels were assigned using a pathway-based scoring rule, these results should be interpreted solely within the context of the computational simulation. The identified features therefore reflect the structure of the synthetic scoring framework rather than biologically validated biomarkers for inherited retinal dystrophies.

Figure 3.

Feature importance for the logistic regression model. The 15 highest-contributing proteins are displayed.

IV. Discussion

IRDs are characterized by high genetic and phenotypic heterogeneity, with clinical progression often conceptualized as alternating between latent and active phases. As described in the literature, transitions between these states involve complex mechanisms, including oxidative stress, inflammatory signaling, metabolic imbalance, and cellular remodeling. Identifying the onset of the active phase remains a major challenge, because earlier intervention during this window may help preserve visual function. In this study, we conducted a methodological simulation exercise to evaluate the performance of ML models in distinguishing between these phases. Given the current scarcity of large-scale, open-source proteomic data for IRDs, our findings are necessarily confined to a theoretical modeling context. This work did not attempt to validate specific biological mechanisms; rather, these concepts served as the conceptual foundation for constructing a controlled, rule-based synthetic environment. The primary objective was to assess how different algorithms, including SVM, RF, LR, and XGBoost performed when class labels were generated through a predefined scoring function. LR achieved the highest performance in this simulation, with accuracy, precision, and F1-score values of 0.73. Because the labels were derived from a structured mathematical rule, this performance reflects the model’s ability to approximate that specific rule rather than its ability to uncover novel, independent biological relationships. Feature contribution analysis within the LR model identified variables with larger coefficient magnitudes. These values represent alignment with the predefined simulation parameters and should not be interpreted as validated biological biomarkers. This approach allowed full control over data dimensionality and pathway grouping, facilitating systematic exploration of model robustness. Such simulations are useful for benchmarking algorithms and understanding methodological behavior before empirical datasets are analyzed. Future research must validate this pipeline with real-world clinical data to determine its applicability to medical practice. In addition, integrating retinal imaging data, which can provide objective markers of photoreceptor degeneration and blood–retinal barrier breakdown, may enhance the predictive performance and clinical relevance of these ML approaches.

In conclusion, this study illustrates how ML models perform under conditions in which class labels are generated using predefined rules. The observed performance reflects the models’ ability to approximate the rule used to construct the labels rather than biological or clinical predictive capability. Future work will involve testing the methodological pipeline on real proteomic data to assess its practical applicability. External validation is required.

Supplementary Materials

Supplementary materials can be found via https://doi.org/10.4258/hir.2026.32.3.285.