Next Article in Journal
Altered Meristem Initiation Is Associated with Increased OSHB3 Expression in a Semi-Dominant Rice Mutant
Previous Article in Journal
Advances in the Knowledge of the Reproductive Processes of the Critically Endangered Pinna nobilis Linnaeus, 1758
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Biology
Volume 15
Issue 11
10.3390/biology15110849
biology-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Circadian Transcriptomic Dynamics Identify Transferable Retina–Choroid Expression Patterns in Myopia Development via Multistage Machine Learning

by
Akarapon Watcharapalakorn
1,
Teera Poyomtip
1,
Patarakorn Tawonkasiwattanakun
1,
Putri Krishna Kumara Dewi
2,
Thotsapol Thomrongsuwannakij
3,4 and
Tanakamol Mahawan
3,4,5,*
1
Faculty of Optometry, Ramkhamhaeng University, Bangkok 10240, Thailand
2
Medical Biochemistry Division, Department of Biomedical Science, Faculty of Medicine, Universitas Pendidikan Ganesha, Singaraja 81116, Indonesia
3
Akkhraratchakumari Veterinary College, Walailak University, Nakhon Si Thammarat 80160, Thailand
4
One Health Research Center, Walailak University, Nakhon Si Thammarat 80160, Thailand
5
Research Center for Theoretical Simulation and Applied Research in Bioscience and Sensing, Walailak University, Nakhon Si Thammarat 80160, Thailand
*
Author to whom correspondence should be addressed.
Biology 2026, 15(11), 849; https://doi.org/10.3390/biology15110849 (registering DOI)
Submission received: 14 April 2026 / Revised: 11 May 2026 / Accepted: 25 May 2026 / Published: 29 May 2026
Browse Figures
Versions Notes

Simple Summary

Myopia, or short-sightedness, is becoming increasingly common worldwide and can increase the risk of serious eye disease later in life. Although light exposure and daily body rhythms are thought to influence eye growth, the biological link between time of day and myopia development remains unclear. In this study, we reanalyzed publicly available gene activity data from the retina and choroid, two eye tissues involved in controlling eye growth, using computer-based prediction methods. The study confirmed previous findings that samples collected during the middle to late light period show a distinct pattern of gene activity. This pattern was observed across eye tissues and stages of myopia development and was partly supported in an external transcriptomic dataset. These findings suggest that the timing of biological processes in the eye may be related to myopia development. The study provides a useful framework for future research on how daily rhythms affect eye growth and may help guide future studies on the timing of myopia prevention or treatment strategies.

Abstract

Circadian regulation has emerged as an important modulator of ocular growth; however, its role in organizing retina–choroid transcriptomic responses during myopia development remains incompletely understood. In this study, we reanalyzed publicly available retinal and choroidal RNA-seq datasets from chick models of form-deprivation myopia using a multistage machine learning framework. A biologically motivated ZT8/12 circadian window was defined from prior published time-of-day transcriptomic evidence and evaluated using feature selection, cross-tissue and cross-stage validation, and external validation in an independent retinal dataset. Machine learning models classified the ZT8/12 window with high performance across onset and progression datasets, and control analyses indicated that this signal reflects a broad transcriptome-wide temporal state rather than a pattern unique to the 53-gene signature. The final gene signature is therefore interpreted as a stable representative subset of the ZT8/12-associated expression state. Cross-species functional enrichment and ortholog mapping suggested hypothesis-generating functional relationships between chicken genes and human orthologs. Overall, this work provides a computational framework for evaluating time-associated expression patterns in myopia and highlights circadian timing as a candidate component of retina–choroid biology requiring further functional validation.

1. Introduction

Myopia (nearsightedness) is a major global public health concern, with projections suggesting that nearly half of the world’s population will be affected by 2050 [1]. In addition to the burden of refractive error, myopia—particularly high myopia—is associated with an increased risk of sight-threatening complications, including retinal detachment, myopic maculopathy, glaucoma, and choroidal neovascularization [2,3]. Despite extensive research, the biological mechanisms underlying abnormal ocular elongation remain incompletely understood.
Refractive development is a visually guided process in which the retina detects optical defocus and initiates signaling cascades that propagate through the choroid to the sclera, ultimately regulating eye growth [4]. This retina–choroid–sclera signaling axis integrates visual input, neuromodulation, metabolic regulation, and extracellular matrix remodeling [5]. Notably, pathological changes are often first observed in the retina, supporting its central role in initiating growth-related responses [6].
Circadian biology has increasingly been recognized as an important regulator of ocular growth [7]. In chicks and mice, manipulating circadian timing or light–dark cycles alter ocular growth. Similarly, in humans, reduced daylight exposure and irregular sleep patterns increase the risk of myopia onset, suggesting the involvement of daytime light and circadian entrainment in myopia development [8,9]. However, the molecular mechanisms linking circadian regulation to refractive development remain poorly defined.
The chick (Gallus gallus domesticus) is a well-established model for visually induced myopia [10]. Form-deprivation myopia (FDM) can be reliably induced using a translucent diffuser, producing rapid axial elongation and retinal and choroidal changes that resemble human myopia. Importantly, chicks exhibit robust circadian rhythms in retinal physiology, making them suitable for investigating time-of-day-dependent mechanisms [11].
Recent transcriptomic studies have characterized diurnal gene expression in the chick retina and choroid during myopia onset and progression, revealing pronounced time-of-day-dependent differences between occluded and control eyes [12,13]. These published analyses reported particularly strong transcriptional activity during the mid-to-late light phase, including ZT8/12. In the present study, we treat this interval as a biologically motivated analytical window rather than as a window newly discovered from model performance. This distinction is important because the primary datasets used here are derived from the same published studies that first described the temporal transcriptomic patterns.
Machine learning (ML) approaches have been increasingly used in myopia research to identify molecular biomarkers and model complex gene expression patterns associated with ocular growth [14,15]. By capturing nonlinear and multivariate relationships in high-dimensional data, ML enables robust classification and cross-dataset validation of myopia-related molecular signatures, typically using differential expression and feature selection methods. While these approaches have identified candidate biomarkers, they are generally based on single datasets and do not incorporate temporal or circadian structure [16]. More recent models integrating clinical and metabolomic features have shown promising predictive performance, but they do not account for circadian dynamics [17]. Thus, a key gap remains in applying integrative, multi-stage ML frameworks to circadian transcriptomic data in myopia.
In this study, we integrate publicly available retinal and choroidal RNA-seq datasets using a multistage machine learning framework to evaluate whether the biologically motivated ZT8/12 interval can be represented by a stable and transferable multigene expression pattern. We assess feature-selection stability, cross-tissue and cross-stage transferability, external reproducibility, and control analyses designed to distinguish time-associated expression structure from arbitrary label assignment or non-specific transcriptomic variation. We further use cross-species functional enrichment as a hypothesis-generating analysis rather than as direct evidence of conserved causal mechanisms.

2. Materials and Methods

2.1. Data Collection

This study reanalyzed publicly available RNA-seq datasets from previously published chick myopia studies (Stone et al., 2024 [12,13]), including GSE227724 (myopia onset) and GSE261232 (myopia progression). Both datasets were generated from Cornell K strain White Leghorn chicks subjected to form-deprivation myopia under 12 h light/12 h dark conditions. Retina and choroid tissues were collected at six Zeitgeber time points (ZT0-ZT20, 4 h intervals; n = 6 chicks per time point). Count matrices and metadata were inspected before normalization. Count values of zero were treated as observed low or absent expression values rather than missing data and were not imputed. No missing expression values remained after preprocessing. Raw sequencing data and expression matrices were obtained from Gene Expression Omnibus and used for downstream analyses.

2.2. Circadian Critical Window Identification

The ZT8/12 classification target was defined as a biologically motivated analytical contrast based on previously reported time-of-day transcriptomic patterns in the source datasets, rather than as a de novo discovery derived from model performance in the present study. Because circadian transcriptomes are periodic, this binary grouping does not fully capture temporal continuity or phase relationships across the 24 h cycle. Instead, the aim was to test whether the previously reported mid-to-late light-phase signal could be represented as a transferable multigene expression signature across tissues, disease stages, and an external dataset.
Samples collected at ZT8 and ZT12 were grouped as ZT_8/12 to form the primary analytical class, while the remaining time points were grouped as ZT_other. This classification framework should therefore be interpreted as a targeted validation of a biologically motivated time window, rather than as a comprehensive rhythmicity model. Dedicated rhythmicity models will be required in future work to estimate phase, amplitude, and temporal autocorrelation more directly.

2.3. Machine Learning Analysis

The machine learning framework consisted of four stages (Figure 1): (i) feature discovery in the retina onset dataset, (ii) cross-tissue validation in the choroid onset dataset, (iii) cross-stage validation using the progression dataset, and (iv) external validation using an independent retinal dataset. All analyses were conducted in R version 4.5.1 [18]. The complete analysis scripts, processed metadata, model outputs, and R session information are archived on Zenodo (https://doi.org/10.5281/zenodo.19953551; accessed on 1 May 2026). Key packages included edgeR v4.6.3 [19], Boruta v9.0.0 [20], caret v7.0.1 [21], clusterProfiler v4.16.0 [22], rrvgo v1.20.0 [23], alluvial v12.5.1 [24], and iml v0.11.4 [25]. Model training and validation were performed on a Windows 11 workstation (Microsoft Corporation, Redmond, WA, USA) with a 13th Gen Intel(R) Core (TM) i5-13420H CPU and 16 GB RAM (Intel Corporation, Santa Clara, CA, USA). Random seeds were fixed for all resampling, feature-selection, and control-analysis steps.
This study follows the TRIPOD + AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis—Artificial Intelligence extension) guidelines [26], ensuring transparent and reproducible reporting of the machine learning models used for predictive analysis in myopia. The TRIPOD + AI checklist, outlining the study’s adherence to these guidelines, can be found in Supplementary Data S6.

2.3.1. Stage 1: Primary Discovery Model (Retina—Onset Dataset)

RNA-seq count matrices and sample metadata were retrieved from GSE227724. Raw counts were processed using edgeR v4.6.3 [19]. Low-expression genes were filtered using filterByExpr function, library sizes were normalized using the Trimmed Mean of M-values (TMM) method [27], and expression values were transformed to log2 counts per million (logCPM). Zeitgeber time was encoded as sine and cosine variables for exploratory circadian visualization and metadata-aware description of sampling. To avoid target leakage, these time-encoding variables were not included in Boruta feature selection, LASSO feature selection, or supervised classifier training. All predictive models were trained using gene-expression features only.
To avoid data leakage, feature selection and model development were restricted to the training partitions. Feature selection was performed using Boruta v9.0.0 [20]. Boruta was repeated 50 times using different random seeds, with maxRuns = 100 for each run. Only genes selected in all 50 runs were retained as the final consensus signature. LASSO was used as a parsimonious linear comparator, whereas Boruta was used as the primary downstream signature because it is an all-relevant selection method capable of retaining correlated and potentially interacting genes.
Classification models were developed using Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), and regularized Logistic Regression (LR) within caret v7.0.1 [21]. These algorithms were selected to represent complementary model assumptions: RF as a nonlinear ensemble method, SVM as a margin-based classifier for high-dimensional data, NB as a low-complexity probabilistic baseline, and LR as an interpretable linear benchmark with regularization. Deep learning models were not used because the sample size was modest, and the risk of overfitting was high. Feature-selection stability was assessed using Boruta selection-frequency plots, and SHAP analysis was performed using the final RF classifier trained on the 53-gene signature by iml v0.11.4 [25]. Mean absolute SHAP values were used to summarize global feature importance, and SHAP summary plots were generated for the top contributing genes.
Importantly, explicit oversampling techniques (e.g., SMOTE) were not used to artificially balance the classes before training. Instead, the models were evaluated for their ability to handle the natural class distributions, with performance metrics calculated on the 20% hold-out test sets. These metrics included Accuracy, Precision, Recall, F1-score, Area Under the Curve (AUC), and Matthew’s correlation coefficient (MCC). Recall, F1-score, and MCC were prioritized to assess the models’ ability to classify the data accurately in the presence of class imbalance, without relying on synthetic data augmentation. Summary statistics and 95% confidence intervals were calculated for all metrics across the 50 outer iterations. Gene signatures derived from this retina onset dataset were then evaluated for cross-tissue and cross-stage transferability.

2.3.2. Stage 2: Cross-Tissue Validation (Retina → Choroid)

To evaluate transferability across tissues, the gene signatures derived from the retina onset dataset were applied to choroid samples from the same study. Models were trained using the same nested cross-validation framework, incorporating 5-fold inner cross-validation for hyperparameter tuning and 50 outer iterations of stratified resampling. The same four algorithms (RF, SVM, NB, LR) were employed. Performance was assessed using the same stringent evaluation metrics (Accuracy, Precision, Recall, F1-score, AUC, MCC) to ensure robustness across tissues.

2.3.3. Stage 3: Cross-Stage Validation (Onset → Progression)

To assess temporal robustness, the retina-derived gene signatures were applied to the independent progression dataset. Models were trained using the same algorithms and circadian classification criteria (ZT_8/12 vs. ZT_other). Classification was performed using the established 50-iteration nested cross-validation procedure. Model performance was subsequently quantified using the previously defined metrics, with particular emphasis on Recall, F1-score, and MCC, to determine whether the chronobiological signatures identified during the onset phase remained predictive and resilient to class imbalance during disease progression.

2.3.4. Stage 4: External Validation (Independent Dataset)

External validation was performed using the independent retinal RNA-seq dataset GSE203604 (n = 42). Raw count data were processed in R, duplicate gene identifiers were collapsed by averaging, and lowly expressed genes were filtered using a threshold of counts ≥10 in at least three samples. Expression values were log2-transformed, and the predefined 53-gene Boruta signature was intersected with genes detected in the external dataset, resulting in 32 available genes. Genes absent from the external dataset were excluded rather than imputed. Gene-wise z-score normalization was applied before downstream analysis. Because GSE203604 differs from the primary form-deprivation datasets and does not include the same six Zeitgeber sampling points, it was used to evaluate whether the predefined signature retained multivariate structure in an independent retinal myopia dataset rather than to repeat the ZT8/12 classification directly. Classification performance was evaluated using nested leave-one-out cross-validation with regularized Logistic Regression and a linear SVM. Statistical significance was assessed using permutation testing, and 95% confidence intervals were estimated by bootstrap resampling. A composite signature score was calculated as the mean gene-wise z-score across detected signature genes, and temporal trends across D1, D3, and D6 were evaluated using linear models including day, group, and interaction terms.

2.3.5. Hyperparameter Tuning

For Stages 1–3, model performance was evaluated using repeated nested resampling. The outer loop consisted of 50 iterations of stratified resampling, in which data were split into 80% training and 20% testing partitions while preserving the ZT8/12 and ZT_other class distributions. Within each training partition, 5-fold inner cross-validation was used to select optimal hyperparameters based on AUC. Reported performance values represent the mean and 95% confidence interval across the 50 outer test-set evaluations. For Stage 4 external validation, AUC confidence intervals were estimated by bootstrap resampling.
For RF, ntree was set to 500, and mtry was tuned within the training folds. For Stages 1–3, SVM models were implemented using the radial-basis kernel through the caret svmRadial method, with centering and scaling applied within training folds. Naive Bayes was tuned using Laplace smoothing parameters, and regularized Logistic Regression was tuned across alpha and lambda values. For Stage 4 external validation, a linear SVM was used because the dataset was smaller and only 32 signature genes were available after gene-overlap filtering.

2.3.6. Control and Sensitivity Analyses

To evaluate whether classification performance depended on the ZT8/12-associated expression structure rather than arbitrary label assignment or non-specific transcriptomic variation, several control and sensitivity analyses were performed. First, a time-label permutation control was conducted by randomly shuffling the ZT8/12 and ZT_other class labels while preserving the expression matrix and the 53-gene feature set. Second, random 53-gene sets were sampled from the expression matrix and evaluated using the same Random Forest workflow. Third, six candidate housekeeping/reference genes that overlapped with the dataset were examined for Zeitgeber-time association and tested as a reference-gene control. Fourth, data-derived non-circadian genes were selected from genes with weak Zeitgeber-time association and used to generate matched 53-gene control sets. Finally, single-time-point sensitivity analyses were performed for ZT8 alone and ZT12 alone and compared with the combined ZT8/12 window.

2.4. Functional Enrichment and Translation Analyses

To investigate biological processes associated with the selected gene signature, Gene Ontology (GO) enrichment analysis was performed on Boruta-selected chicken genes and their corresponding human orthologs using clusterProfiler v4.16.0 [22]. Significant GO terms (adjusted p ≤ 0.05) were summarized and reduced for redundancy using rrvgo v1.20.0 [23].
Parent GO terms for chicken and human datasets were visualized using treemaps. Enrichment results were also summarized at the parent biological-process level and visualized using alluvial diagrams generated with the alluvial package v12.5.1 [24]. These cross-species enrichment analyses were interpreted as hypothesis-generating because GO enrichment and ortholog mapping provide indirect functional evidence and may be affected by annotation bias and incomplete cross-species translatability.

3. Results

3.1. Study Design and Sample Size

The myopia onset and progression studies used identical circadian sampling schemes. Each study included 36 chicks in total, corresponding to six Zeitgeber time points with six chicks per time point. For each chick, tissues were collected from the occluded eye and the contralateral open eye, and retina and choroid were processed separately. Thus, each study contributed 144 tissue-level samples in total: 36 chicks × 2 eyes × 2 tissues. For tissue-specific analyses, retina and choroid were analyzed separately, corresponding to 72 samples per tissue per study. At the full tissue-level dataset scale, the ZT8/12 class included 48 samples, while the ZT_other class included 96 samples. Within each tissue-specific dataset, the ZT8/12 and ZT_other classes included 24 and 48 samples, respectively, as shown in Table 1. The paired origin of occluded and contralateral eyes is considered when interpreting model transferability.

3.2. ZT8/12 Analytical Window and Transcriptomic Structure

Samples collected at ZT8 and ZT12 were grouped as the ZT_8/12 analytical window, comprising 48 tissue-level samples (33.3%) in each study. The remaining time points (ZT0, ZT4, ZT16, and ZT20) were classified as ZT_other, representing 96 samples (66.7%), and were used as the comparison class for downstream machine learning analyses. Detailed phenotype metadata, including the computational class label, are provided in Supplementary Data S1.
PCA showed that samples collected at ZT8/12 were separated from other time points, supporting the presence of a strong time-associated transcriptomic structure (Figure 2). This separation is interpreted as evidence of a broad ZT8/12-associated temporal state, not as proof that the 53-gene signature is unique or causal. Figure 2 includes the percentage of variance explained by each principal component, and confidence ellipses are shown for visualization only.

3.3. Machine Learning Classification and Control Analyses (Stage 1)

Feature selection was performed using the retina onset dataset. After repeated analysis, LASSO selected 24 genes, while Boruta identified 53 genes that were selected consistently in all 50 repeated runs (50/50). The 53 gene IDs and their ortholog annotations are provided in Supplementary Data S2. The 50/50 Boruta criterion supports feature-selection stability, although subsequent control analyses indicate that the ZT8/12 signal is broad across the transcriptome rather than restricted to these genes alone.
Supervised models trained on the retina onset dataset distinguished samples collected during the ZT8/12 window from other time points. Using the LASSO-selected gene set (n = 24), all classifiers showed high performance. Using the Boruta-selected gene set (n = 53), performance was also high across classifiers, supporting the utility of this stable consensus signature for downstream transferability testing. We therefore used the Boruta-selected genes for Stages 2–4, while retaining LASSO as a parsimonious comparator (Table 2).
Boruta feature-selection stability analysis showed that the final 53-gene signature consisted only of genes selected in all 50 repeated Boruta runs, supporting high selection consistency (Supplementary Figure S1). SHAP analysis of the final RF model identified the top contributing genes for ZT8/12 classification based on mean absolute SHAP values (Supplementary Figure S2). The SHAP summary plot further showed the direction and distribution of gene-level contributions across samples (Supplementary Figure S3). These analyses support model stability and interpretability but do not imply causal roles for individual genes.
To improve biological interpretation of the model features, the top SHAP-ranked genes were examined in relation to known retinal and circadian pathways. Several top SHAP-ranked genes were related to downstream retinal signaling, ion transport, lipid metabolism, and photoreceptor-associated processes, for example, CACNA2D4, CNGA3, KCNJ11, FADS2, DGAT2, CERS2, and ESRRB. Thus, the SHAP results suggest that the ZT8/12-associated signature reflects downstream retinal physiological and metabolic states rather than direct selection of canonical clock genes. These genes were further explained by GO enrichment visualized by Treemap (rrvgo v1.20.0 [23]) and alluvial visualizations(alluvial v12.5.1 [24]) in Section 3.7.

3.4. Control and Sensitivity Analyses for ZT8/12 Classification

Control analyses supported that the ZT8/12 classification depended on the true time-associated expression structure. In the time-label permutation analysis, the true-label RF model achieved an AUC of 1.000, whereas shuffled-label controls produced a mean AUC of 0.637 ± 0.034. None of the 100 shuffled-label permutations reached the observed AUC, corresponding to a corrected empirical p-value of 0.0099.
Random 53-gene sets also achieved high classification performance, with a mean AUC of 0.997 ± 0.003. This indicates that the ZT8/12 contrast reflects a broad transcriptome-wide temporal state rather than a signal unique to the 53 Boruta-selected genes. In contrast, data-derived non-circadian 53-gene controls showed lower and more variable performance, with a mean AUC of 0.667 ± 0.080 and mean MCC of 0.048 ± 0.256, supporting that classification depended on time-associated expression structure rather than arbitrary non-circadian gene sets.
Six candidate housekeeping/reference genes overlapped with the expression matrix. However, five of these six genes showed significant Zeitgeber-time association after false-discovery-rate correction, indicating that they were not suitable stable negative controls in this circadian retinal dataset. Single-time-point sensitivity analyses showed that ZT8-only and ZT12-only contrasts were both highly separable, while the combined ZT8/12 window showed the most stable performance across metrics. Full control and sensitivity results are provided in Supplementary Data S7.

3.5. Cross-Dataset and Cross-Tissue Validation (Stage 2 and 3)

To assess transferability of the ZT8/12-associated expression pattern, models trained on the retina onset dataset were evaluated in two validation settings: cross-stage validation using the progression dataset and cross-tissue validation using choroidal samples. In the cross-stage setting (onset to progression), SVM achieved the highest performance (accuracy = 0.972; AUC = 0.979), followed by RF (accuracy = 0.931; AUC = 0.896). In the cross-tissue setting (onset retina to choroid), RF showed strong performance (accuracy = 0.954; AUC = 0.946). These results support predictive transferability under matched experimental conditions, but they do not establish direct signaling transmission or causality between retina and choroid (Table 3).

3.6. External Validation on Independent Data (Stage 4)

Of the 53 genes in the original Boruta signature, 32 were detected in GSE203604 and used for external validation. Unsupervised analyses based on these genes showed consistent multivariate structure, with PCA indicating separation of samples across time points and experimental groups. Nested leave-one-out cross-validation showed an AUC of 0.823 for regularized Logistic Regression and 0.912 for the linear SVM. Permutation testing indicated that the linear SVM performed significantly better than chance (p = 0.015), whereas Logistic Regression did not exceed chance-level performance (p = 0.68). Bootstrap confidence intervals were 0.782–0.997 for Logistic Regression and 0.841–0.998 for the linear SVM. The composite signature score increased from D1 to D6, and PCA showed variation along PC1 with differences between control and myopia groups across time as shown in Figure 3.

3.7. Cross-Species Functional Enrichment of Myopia-Associated Gene Signatures

GO enrichment analysis of Boruta-selected genes revealed distinct but related functional profiles between chicken genes and their human orthologs. After redundancy reduction, enriched biological processes in chicken were summarized into parent programs associated with cellular metabolism, ion transport, intracellular trafficking, and post-transcriptional regulation (Supplementary Data S3). Human orthologs were enriched for categories related to neuroendocrine regulation, photoreceptor maintenance, hormonal response, and mRNA stability (Supplementary Data S4). Treemap and alluvial visualizations further highlighted these differences in enriched functional categories between chicken genes and their human orthologs (Figure 4; Supplementary Data S5).
These GO patterns provide mechanistic context for the SHAP-ranked genes and suggest that the Boruta-selected chicken genes and their human orthologs map to partially overlapping but distinct biological-process groups. The enrichment patterns are consistent with the gene-level interpretation that the ZT8/12-associated signature captures retinal signaling, photoreceptor-related function, ion transport, and metabolic regulation. However, these findings should be interpreted as functional annotation trends rather than direct evidence of conserved causal mechanisms, because GO enrichment and ortholog mapping provide indirect evidence and may be influenced by annotation bias, incomplete cross-species translatability, and the fact that enrichment does not establish causality or pathway direction.

4. Discussion

This study evaluated whether a biologically motivated ZT8/12 window, previously highlighted in published chick myopia transcriptomic studies, can be represented by a stable and transferable multigene expression pattern [12,13]. Using a multistage machine learning pipeline that produced high classification performance and cross-dataset transferability, our results indicate robust confirmatory evidence for the presence of a strong ZT8/12-associated transcriptomic state across onset and progression datasets. However, the classification framework should not be interpreted as a de novo discovery of the window or as a complete periodic time-series model. Rather, it provides a targeted test of a previously reported temporal interval. The ZT8/12 interval is biologically plausible in the context of ocular circadian physiology because it spans the mid-to-late light phase, a period associated with dynamic changes in retinal neuromodulation and transition toward the dark phase [28,29]. Nevertheless, the present computational analysis does not establish that this interval is mechanistically causal in myopia progression. It should instead be considered a hypothesis-generating temporal window for future rhythmicity-based and functional studies.
The control analyses refine the interpretation of the 53-gene signature. Although all 53 genes were selected consistently in 50 out of 50 repeated Boruta runs, random 53-gene controls also achieved high classification performance. This suggests that the ZT8/12 contrast reflects a broad transcriptome-wide temporal state rather than a signal uniquely restricted to the selected genes. Therefore, the 53-gene Boruta signature should be interpreted as a stable and compact representative subset of the ZT8/12-associated expression pattern, not as an exclusive or causal gene set. The lower performance of data-derived non-circadian gene controls further supports that the classifier depends on time-associated expression structure [30,31].
Cross-tissue validation demonstrated that a retinally derived expression signature could classify choroidal samples under matched experimental conditions. This finding supports cross-tissue predictive transferability, but it does not establish direct signaling transmission or causality between retina and choroid. Because retina and choroid samples were collected from the same animals and time points, shared timing and paired biological context may contribute to the observed performance [12,13]. Independent tissue-specific datasets and functional experiments will be required to confirm whether these transcriptomic patterns reflect coordinated retina–choroid signaling [28,32].
Cross-species enrichment analysis suggested that chicken signature genes and their human orthologs are associated with partially overlapping but distinct functional categories. Human orthologs were enriched for regulatory pathways related to neuroendocrine signaling, photoreceptor maintenance, hormonal response, and mRNA stability, a pattern broadly consistent with GWAS and functional genomic studies implicating photoreceptor-related pathways, synaptic or neuroregulatory signaling, and post-transcriptional regulation in myopia development [33,34,35]. In contrast, chicken genes were more strongly associated with cellular metabolic and regulatory processes [36]. These findings suggest partial translational convergence at the pathway level, while also indicating species-specific differences in downstream biological implementation. Such differences may reflect biological divergence and species-specific regulatory context, but they may also result from ortholog-mapping limitations, gene-annotation bias, or incomplete translatability of the chick model to human myopia [37]. Therefore, this cross-species analysis should be considered hypothesis-generating rather than direct evidence of neuroendocrine reorganization in humans. Nevertheless, the presence of partially shared functional themes supports the continued relevance of the chick model for studying fundamental mechanisms of visually induced myopia [11].
The absence of canonical core clock genes such as CLOCK and BMAL1/ARNTL among the top SHAP-ranked genes suggests that the classifier captured downstream circadian-associated retinal states rather than the central molecular clock itself. This is biologically plausible because retinal circadian clocks regulate multiple downstream processes, including gene expression, synaptic communication, neurotransmitter release, metabolism, and photoreceptor physiology [38]. In the context of myopia, the mid-to-late light phase may be particularly relevant because retinal dopamine is generally associated with light-adapted retinal signaling and inhibition of excessive ocular elongation, whereas melatonin-related pathways are more closely linked to the transition toward the dark phase [39,40]. The presence of SHAP-ranked genes related to photoreceptor signaling, including CACNA2D4 and CNGA3, ion transport through KCNJ11, lipid metabolism through FADS2, DGAT2, and CERS2, and photoreceptor maintenance through ESRRB supports the interpretation that the ZT8/12 signature may reflect a coordinated downstream retinal signaling and metabolic state [41,42,43]. Nevertheless, these associations remain correlative and require functional validation.
External validation provided partial support that the predefined 53-gene signature retains reproducible multivariate structure in an independent retinal RNA-seq dataset. Although regularized Logistic Regression showed moderate classification performance, it did not exceed chance-level performance based on permutation testing. In contrast, the linear SVM achieved higher and statistically significant performance, suggesting that the detected signature genes capture multivariate expression patterns that are partly preserved outside the primary datasets [44].These findings should be interpreted cautiously because GSE203604 had a modest sample size, differed in experimental design, lacked matched circadian sampling, and contained only 32 of the 53 original signature genes.
A further limitation is that the main modelling strategy used a binary ZT8/12 versus other-time-point classification task. This approach was useful for testing a specific biologically motivated time window, but it does not fully capture the periodic and continuous nature of circadian transcriptomic data. Future studies should incorporate dedicated rhythmicity or longitudinal models, such as harmonic regression [45], cosinor-based modelling [46], or mixed-effects time-series approaches, to estimate phase, amplitude, and temporal autocorrelation more directly.
Several additional limitations should be acknowledged. This study relies on publicly available RNA-seq datasets with modest sample sizes, reflecting the constraints of circadian experimental designs requiring terminal tissue collection. The chick model is well established for visually induced myopia, but species-specific differences may limit direct translation to human disease [11]. The analysis was restricted to transcriptomic data, and other regulatory layers, including proteomic and metabolic processes, were not assessed. Functional validation experiments were not performed. Finally, paired-eye and paired-tissue origins may contribute to some cross-dataset transferability patterns and should be addressed in future experimental designs [47].
These findings identify circadian timing and retina–choroid expression patterns as candidate mechanisms requiring functional validation. Because this study did not test interventions, the results should not be interpreted as direct evidence that manipulating circadian timing will alter myopia progression. Instead, the ZT8/12 window provides a hypothesis-generating temporal framework for future mechanistic, rhythmicity-based, and chronotherapeutic studies [48,49,50].

5. Conclusions

This study shows that a biologically motivated ZT8/12 window can be represented by a transferable multigene expression pattern in chick retinal and choroidal transcriptomic datasets. The findings support cross-stage and cross-tissue predictive transferability of this time-associated expression state, with partial support from an independent retinal dataset. However, the results do not establish causality or direct retina–choroid signaling mechanisms. The 53-gene Boruta signature should be interpreted as a stable representative subset of a broader ZT8/12-associated transcriptomic state, not as an exclusive or causal gene set. These findings provide a computational framework and candidate gene set for future functional and rhythmicity-based studies of circadian regulation in myopia development.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/biology15110849/s1. Supplementary Data S1: Phenotypic metadata and ZT8/12 class labels. Supplementary Data S2: Boruta-selected genes, ortholog annotations, and feature information. Supplementary Data S3: Gene Ontology enrichment results for chicken genes. Supplementary Data S4: Gene Ontology enrichment results for human orthologs. Supplementary Data S5: Treemap and alluvial visualization data for functional enrichment results. Supplementary Data S6: TRIPOD + AI checklist [26,51]. Supplementary Data S7: Control and sensitivity analyses for ZT8/12 classification. Supplementary Figure S1: Boruta feature-selection frequency across 50 repeated runs. Supplementary Figure S2: SHAP global feature importance for the top 20 genes. Supplementary Figure S3: SHAP summary plot for the top 20 Boruta genes.

Author Contributions

Conceptualization, T.M. and A.W.; methodology, T.M. and A.W.; software, A.W.; validation, A.W., T.M. and T.P.; formal analysis, A.W. and T.M.; investigation, A.W., T.P. and P.T.; resources, T.T.; data curation, A.W.; writing—original draft preparation, A.W.; writing—review and editing, T.M., P.K.K.D. and T.T.; visualization, A.W.; supervision, T.M.; project administration, T.M.; funding acquisition, T.M. All authors have read and agreed to the published version of the manuscript.

Funding

This work was partially supported by Walailak University under the International Mobility for Publication and Collaboration scheme (Contract Number WU-CIA-07208/2025).

Institutional Review Board Statement

This study was reviewed and approved for exemption by the Walailak University Institutional Animal Care and Use Committee (WU-IACUC) under project number WU-ACUC-68045.

Informed Consent Statement

Not applicable.

Data Availability Statement

Publicly available GEO datasets GSE227724, GSE261232, and GSE203604 were used in this study. Additional supporting data and results are provided in the Supplementary Materials. The machine learning framework and R scripts used for analysis (including Stages 1 through 4) are archived and publicly available on Zenodo at https://doi.org/10.5281/zenodo.19953551 (accessed on 1 May 2026).

Acknowledgments

The authors would like to thank Richard A. Stone and his team at the University of Pennsylvania and Andrei V. Tkatchenko and his team at the Columbia University for generating and publicly sharing the RNA-sequencing data and associated metadata used in this study.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
RFRandom Forest
SVMSupport Vector Machine
NBNaïve Bayes
LRLogistic Regression
AUCArea Under the Curve
MCCMatthew’s Correlation Coefficient
PCAPrincipal Component Analysis
GOGene Ontology
TMMTrimmed Mean of M-values

References

  1. Holden, B.A.; Fricke, T.R.; Wilson, D.A.; Jong, M.; Naidoo, K.S.; Sankaridurg, P.; Wong, T.Y.; Naduvilath, T.J.; Resnikoff, S. Global Prevalence of Myopia and High Myopia and Temporal Trends from 2000 through 2050. Ophthalmology 2016, 123, 1036–1042. [Google Scholar] [CrossRef] [PubMed]
  2. Puro, D.G. Retinovascular physiology and pathophysiology: New experimental approach/new insights. Prog. Retin. Eye Res. 2012, 31, 258–270. [Google Scholar] [CrossRef]
  3. Du, Y.; Meng, J.; He, W.; Qi, J.; Lu, Y.; Zhu, X. Complications of high myopia: An update from clinical manifestations to underlying mechanisms. Adv. Ophthalmol. Pract. Res. 2024, 4, 156–163. [Google Scholar] [CrossRef]
  4. Summers, J.A.; Schaeffel, F.; Marcos, S.; Wu, H.; Tkatchenko, A.V. Functional integration of eye tissues and refractive eye development: Mechanisms and pathways. Exp. Eye Res. 2021, 209, 108693. [Google Scholar] [CrossRef]
  5. Li, C.Y.; Song, Q.Q.; Xu, W.J.; Li, X.Y.; Huang, Y.; Wang, N.L.; Li, S.M. The Light-Eye-Brain Axis: Neurobiological Links Between Mood Disorders and Myopia-A Narrative Review. Ophthalmol. Ther. 2026, 15, 901–924. [Google Scholar] [CrossRef]
  6. Zhang, J.; Shi, L.; Shen, Y. The retina: A window in which to view the pathogenesis of Alzheimer’s disease. Ageing Res. Rev. 2022, 77, 101590. [Google Scholar] [CrossRef]
  7. Zhang, X.; Jie, Y. Importance of Circadian Rhythms in the Ocular Surface. Biomolecules 2024, 14, 796. [Google Scholar] [CrossRef] [PubMed]
  8. Ashby, R.; Harb, E.N.; Ostrin, L.A.; Flitcroft, D.I.; Schaeffel, F.; Karouta, C.; Gawne, T.; Chakraborty, R.; Thomson, K.; Read, S.; et al. IMI: The Role of Light in Refractive Development and Myopia: Evidence from Animal and Human Studies. Investig. Ophthalmol. Vis. Sci. 2025, 66, 5. [Google Scholar] [CrossRef] [PubMed]
  9. Bai, W.-L.; Wang, M.-J.; Gan, J.-H.; Huang, Y.; Liu, Z.-H.; Li, C.-Y.; Wang, N.-L.; Li, S.-M. Circadian rhythm disruption induces myopia in mice. Exp. Eye Res. 2026, 264, 110823. [Google Scholar] [CrossRef]
  10. Wisely, C.E.; Sayed, J.A.; Tamez, H.; Zelinka, C.; Abdel-Rahman, M.H.; Fischer, A.J.; Cebulla, C.M. The chick eye in vision research: An excellent model for the study of ocular disease. Prog. Retin. Eye Res. 2017, 61, 72–97. [Google Scholar] [CrossRef]
  11. Morgan, I.G.; Ashby, R.S.; Nickla, D.L. Form deprivation and lens-induced myopia: Are they different? Ophthalmic Physiol. Opt. 2013, 33, 355–361. [Google Scholar] [CrossRef]
  12. Stone, R.A.; Tobias, J.W.; Wei, W.; Carlstedt, X.; Zhang, L.; Iuvone, P.M.; Nickla, D.L. Diurnal gene expression patterns in retina and choroid distinguish myopia progression from myopia onset. PLoS ONE 2024, 19, e0307091. [Google Scholar] [CrossRef]
  13. Stone, R.A.; Tobias, J.W.; Wei, W.; Schug, J.; Wang, X.; Zhang, L.; Iuvone, P.M.; Nickla, D.L. Diurnal retinal and choroidal gene expression patterns support a role for circadian biology in myopia pathogenesis. Sci. Rep. 2024, 14, 533. [Google Scholar] [CrossRef] [PubMed]
  14. Liu, N.; Li, L.; Yu, J. Application of artificial intelligence in myopia prevention and control. Pediatr. Investig. 2025, 9, 114–124. [Google Scholar] [CrossRef] [PubMed]
  15. Mu, J.; Zhong, H.; Jiang, M. Machine-learning models to predict myopia in children and adolescents. Front. Med. 2024, 11, 1482788. [Google Scholar] [CrossRef] [PubMed]
  16. Zhang, S.; Wang, T.; Wang, H.; Gao, B.; Sun, C. Identification of potential biomarkers of myopia based on machine learning algorithms. BMC Ophthalmol. 2023, 23, 388. [Google Scholar] [CrossRef]
  17. Hou, X.W.; Yang, J.L.; Li, D.L.; Tao, Y.J.; Ke, C.F.; Zhang, B.; Liu, S.; Cheng, T.Y.; Wang, T.X.; Xu, X.; et al. Machine Learning-Based Integration of Metabolomics Characterisation Predicts Progression of Myopic Retinopathy in Children and Adolescents. Metabolites 2023, 13, 301. [Google Scholar] [CrossRef]
  18. Team, R.C. R: A Language and Environment for Statistical Computing; R Foundation for Statistical Computing: Vienna, Austria, 2025. [Google Scholar]
  19. Robinson, M.D.; McCarthy, D.J.; Smyth, G.K. edgeR: A Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 2010, 26, 139–140. [Google Scholar] [CrossRef]
  20. Kursa, M.B.; Rudnicki, W.R. Feature selection with the boruta package. J. Stat. Softw. 2010, 36, 1–13. [Google Scholar] [CrossRef]
  21. Kuhn, M. Building predictive models in R using the caret package. J. Stat. Softw. 2008, 28, 1–26. [Google Scholar] [CrossRef]
  22. Wu, T.; Hu, E.; Xu, S.; Chen, M.; Guo, P.; Dai, Z.; Feng, T.; Zhou, L.; Tang, W.; Zhan, L.; et al. clusterProfiler 4.0: A universal enrichment tool for interpreting omics data. Innovation 2021, 2, 100141. [Google Scholar] [CrossRef] [PubMed]
  23. Sayols, S. rrvgo: A Bioconductor package for interpreting lists of Gene Ontology terms. microPubl. Biol. 2023, 2023, 10-17912. [Google Scholar] [CrossRef]
  24. Bojanowski, M.; Edwards, R. Alluvial: R Package for Creating Alluvial Diagrams, R Package Version 0.1; R Foundation for Statistical Computing: Vienna, Austria, 2016; Volume 2.
  25. Molnar, C.; Casalicchio, G.; Bischl, B. iml: An R package for Interpretable Machine Learning. J. Open Source Softw. 2018, 3, 786. [Google Scholar] [CrossRef]
  26. Collins, G.S.; Moons, K.G.M.; Dhiman, P.; Riley, R.D.; Beam, A.L.; Van Calster, B.; Ghassemi, M.; Liu, X.; Reitsma, J.B.; van Smeden, M.; et al. TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024, 385, e078378. [Google Scholar] [CrossRef]
  27. Robinson, M.D.; Oshlack, A. A scaling normalization method for differential expression analysis of RNA-seq data. Genome Biol. 2010, 11, R25. [Google Scholar] [CrossRef]
  28. French, A.N.; Ashby, R.S.; Morgan, I.G.; Rose, K.A. Time outdoors and the prevention of myopia. Exp. Eye Res. 2013, 114, 58–68. [Google Scholar] [CrossRef]
  29. Chakraborty, R.; Ostrin, L.A.; Nickla, D.L.; Iuvone, P.M.; Pardue, M.T.; Stone, R.A. Circadian rhythms, refractive development, and myopia. Ophthalmic Physiol. Opt. 2018, 38, 217–245. [Google Scholar] [CrossRef]
  30. Souto-Maior, C.; Serrano Negron, Y.L.; Harbison, S.T. Nonlinear expression patterns and multiple shifts in gene network interactions underlie robust phenotypic change in Drosophila melanogaster selected for night sleep duration. PLoS Comput. Biol. 2023, 19, e1011389. [Google Scholar] [CrossRef] [PubMed]
  31. Taylor, M.B.; Ehrenreich, I.M. Higher-order genetic interactions and their contribution to complex traits. Trends Genet. 2015, 31, 34–40. [Google Scholar] [CrossRef] [PubMed]
  32. Ko, G.Y. Circadian regulation in the retina: From molecules to network. Eur. J. Neurosci. 2020, 51, 194–216. [Google Scholar] [CrossRef]
  33. Meguro, A.; Yamane, T.; Takeuchi, M.; Miyake, M.; Fan, Q.; Zhao, W.; Wang, I.J.; Mizuki, Y.; Yamada, N.; Nomura, N.; et al. Genome-Wide Association Study in Asians Identifies Novel Loci for High Myopia and Highlights a Nervous System Role in Its Pathogenesis. Ophthalmology 2020, 127, 1612–1624. [Google Scholar] [CrossRef]
  34. Tedja, M.S.; Swierkowska-Janc, J.; Enthoven, C.A.; Meester-Smoor, M.A.; Hysi, P.G.; Felix, J.F.; Cowan, C.S.; Consortium, C.; Cherry, T.J.; van der Spek, P.J.; et al. A genome-wide scan of non-coding RNAs and enhancers for refractive error and myopia. Hum. Genet. 2025, 144, 67–91. [Google Scholar] [CrossRef]
  35. Tedja, M.S.; Wojciechowski, R.; Hysi, P.G.; Eriksson, N.; Furlotte, N.A.; Verhoeven, V.J.M.; Iglesias, A.I.; Meester-Smoor, M.A.; Tompson, S.W.; Fan, Q.; et al. Genome-wide association meta-analysis highlights light-induced signaling as a driver for refractive error. Nat. Genet. 2018, 50, 834–848. [Google Scholar] [CrossRef]
  36. Cherry, T.J.; Yang, M.G.; Harmin, D.A.; Tao, P.; Timms, A.E.; Bauwens, M.; Allikmets, R.; Jones, E.M.; Chen, R.; De Baere, E.; et al. Mapping the cis-regulatory architecture of the human retina reveals noncoding genetic variation in disease. Proc. Natl. Acad. Sci. USA 2020, 117, 9001–9012. [Google Scholar] [CrossRef] [PubMed]
  37. Phan, M.H.Q.; Zehnder, T.; Puntieri, F.; Magg, A.; Majchrzycka, B.; Antonovic, M.; Wieler, H.; Lo, B.W.; Baranasic, D.; Lenhard, B.; et al. Conservation of regulatory elements with highly diverged sequences across large evolutionary distances. Nat. Genet. 2025, 57, 1524–1534. [Google Scholar] [CrossRef]
  38. McMahon, D.G.; Iuvone, P.M.; Tosini, G. Circadian organization of the mammalian retina: From gene regulation to physiology and diseases. Prog. Retin. Eye Res. 2014, 39, 58–76. [Google Scholar] [CrossRef] [PubMed]
  39. Wang, Y.; Fehlhaber, K.E.; Sarria, I.; Cao, Y.; Ingram, N.T.; Guerrero-Given, D. The auxiliary calcium channel subunit α2δ4 is required for axonal elaboration, synaptic transmission, and wiring of rod photoreceptors. Neuron 2017, 93, 1359–1374.e1356. [Google Scholar] [CrossRef]
  40. Onishi, A.; Peng, G.H.; Poth, E.M.; Lee, D.A.; Chen, J.; Alexis, U.; de Melo, J.; Chen, S.; Blackshaw, S. The orphan nuclear hormone receptor ERRβ controls rod photoreceptor survival. Proc. Natl. Acad. Sci. USA 2010, 107, 11579–11584. [Google Scholar] [CrossRef]
  41. Gerhardt, M.J.; Priglinger, S.; Biel, M.; Michalakis, S. Biology, pathobiology and gene therapy of CNG channel-related retinopathies. Biomedicines 2023, 11, 269. [Google Scholar] [CrossRef]
  42. Beverley, K.M.; Pattnaik, B.R. Inward rectifier potassium (Kir) channels in the retina: Living our vision. Am. J. Physiol.-Cell Physiol. 2022, 323, C772–C782. [Google Scholar] [CrossRef] [PubMed]
  43. Fliesler, S.J.; Anderson, R.E. Chemistry and metabolism of lipids in the vertebrate retina. Prog. Lipid Res. 1983, 22, 79–131. [Google Scholar] [CrossRef]
  44. Glazko, G.V.; Emmert-Streib, F. Unite and conquer: Univariate and multivariate approaches for finding differentially expressed gene sets. Bioinformatics 2009, 25, 2348–2354. [Google Scholar] [CrossRef]
  45. Yang, R.; Su, Z. Analyzing circadian expression data by harmonic regression based on autoregressive spectral estimation. Bioinformatics 2010, 26, i168–i174. [Google Scholar] [CrossRef]
  46. Cornelissen, G. Cosinor-based rhythmometry. Theor. Biol. Med. Model. 2014, 11, 16. [Google Scholar] [CrossRef]
  47. Remyes, D.; Nasef, D.; Remyes, S.; Tawfellos, J.; Sher, M.; Nasef, D.; Toma, M. Clinical Applicability and Cross-Dataset Validation of Machine Learning Models for Binary Glaucoma Detection. Information 2025, 16, 432. [Google Scholar] [CrossRef]
  48. Chierigo, A.; Ferro Desideri, L.; Traverso, C.E.; Vagge, A. The Role of Atropine in Preventing Myopia Progression: An Update. Pharmaceutics 2022, 14, 900. [Google Scholar] [CrossRef] [PubMed]
  49. Yam, J.C.; Jiang, Y.; Tang, S.M.; Law, A.K.P.; Chan, J.J.; Wong, E.; Ko, S.T.; Young, A.L.; Tham, C.C.; Chen, L.J.; et al. Low-Concentration Atropine for Myopia Progression (LAMP) Study: A Randomized, Double-Blinded, Placebo-Controlled Trial of 0.05%, 0.025%, and 0.01% Atropine Eye Drops in Myopia Control. Ophthalmology 2019, 126, 113–124. [Google Scholar] [CrossRef]
  50. Kaskal, M.; Sevim, M.; Ulker, G.; Keles, C.; Bebitoglu, B.T. The clinical impact of chronopharmacology on current medicine. Naunyn Schmiedeberg’s Arch. Pharmacol. 2025, 398, 6179–6191. [Google Scholar] [CrossRef] [PubMed]
  51. Debray, T.P.; Collins, G.S.; Riley, R.D.; Snell, K.I.; Van Calster, B.; Reitsma, J.B.; Moons, K.G. Transparent reporting of multivariable prediction models developed or validated using clustered data. BMJ 2023, 380, e071058. [Google Scholar] [CrossRef]
Biology 15 00849 g001
Figure 1. Multistage machine learning framework for evaluating time-associated expression patterns in myopia development. RNA-seq data from onset and progression studies were normalized and analyzed through a four-stage pipeline. Stage 1 (S1) to Stage 4 (S4). Downstream functional analyses provide insights into circadian regulation and retina–choroid interactions underlying myopia development.
Figure 1. Multistage machine learning framework for evaluating time-associated expression patterns in myopia development. RNA-seq data from onset and progression studies were normalized and analyzed through a four-stage pipeline. Stage 1 (S1) to Stage 4 (S4). Downstream functional analyses provide insights into circadian regulation and retina–choroid interactions underlying myopia development.
Biology 15 00849 g001
Biology 15 00849 g002
Figure 2. PCA of retinal and choroidal transcriptomic profiles. (A) Full retinal transcriptome (14,484 genes) by ZT point. (B) Full retina grouped by target window (ZT8/12 vs. others). (C) Retinal 53-gene Boruta signature. (D) Choroidal validation of the 53-gene signature. Axis labels indicate the percentage of variance explained by each principal component. Confidence ellipses indicate group-level dispersion and are shown for visualization only; they were not used for classifier training.
Figure 2. PCA of retinal and choroidal transcriptomic profiles. (A) Full retinal transcriptome (14,484 genes) by ZT point. (B) Full retina grouped by target window (ZT8/12 vs. others). (C) Retinal 53-gene Boruta signature. (D) Choroidal validation of the 53-gene signature. Axis labels indicate the percentage of variance explained by each principal component. Confidence ellipses indicate group-level dispersion and are shown for visualization only; they were not used for classifier training.
Biology 15 00849 g002
Biology 15 00849 g003
Figure 3. External validation of the 53-gene signature in GSE203604. (a) Receiver operating characteristic curves from nested leave-one-out cross-validation for regularized Logistic Regression and linear SVM models. AUC values and 95% bootstrap confidence intervals are shown. (b) PCA based on the 32 detected genes from the predefined 53-gene signature, showing sample distribution by experimental group and time point. (c) Temporal trend of PC1 scores derived from signature-gene expression. (d) Temporal trend of the composite signature score, calculated as the mean gene-wise z-score across detected signature genes. Error bars indicate standard error.
Figure 3. External validation of the 53-gene signature in GSE203604. (a) Receiver operating characteristic curves from nested leave-one-out cross-validation for regularized Logistic Regression and linear SVM models. AUC values and 95% bootstrap confidence intervals are shown. (b) PCA based on the 32 detected genes from the predefined 53-gene signature, showing sample distribution by experimental group and time point. (c) Temporal trend of PC1 scores derived from signature-gene expression. (d) Temporal trend of the composite signature score, calculated as the mean gene-wise z-score across detected signature genes. Error bars indicate standard error.
Biology 15 00849 g003
Biology 15 00849 g004
Figure 4. Functional Translation of Myopia Gene Programs. Major cellular programs in chicken are mapped to regulatory pathways in humans.
Figure 4. Functional Translation of Myopia Gene Programs. Major cellular programs in chicken are mapped to regulatory pathways in humans.
Biology 15 00849 g004
Table 1. Experimental design, sampling scheme, and circadian classification.
Table 1. Experimental design, sampling scheme, and circadian classification.
StudyDurationDayRep/ZTTotal ChicksSamples per TissueTotal Tissue-Level SamplesClassn (%)n per Tissue
Onset1 cycle163672144ZT_8/1248 (33.3)24
ZT_other96 (66.7)48
Progression4 cycles563672144ZT_8/1248 (33.3)24
ZT_other96 (66.7)48
Note: Total tissue-level samples were calculated as 36 chicks × 2 eyes × 2 tissues = 144 samples per study. Tissue-specific analyses were conducted separately for retina and choroid, resulting in 72 samples per tissue per study. ZT8/12 included samples from ZT8/12, whereas ZT_other included ZT0, ZT4, ZT16, and ZT20.
Table 2. Performance of machine learning classifiers for identifying the ZT8/12 analytical window in the retina onset dataset. Values are mean ± 95% confidence interval across 50 outer resampling iterations.
Table 2. Performance of machine learning classifiers for identifying the ZT8/12 analytical window in the retina onset dataset. Values are mean ± 95% confidence interval across 50 outer resampling iterations.
MetricRFSVMNBLR
LASSO (n = 24)
Accuracy0.909 ± 0.0810.920 ± 0.0770.831 ± 0.1060.943 ± 0.066
Precision0.946 ± 0.0640.951 ± 0.0610.967 ± 0.0510.949 ± 0.062
Recall0.924 ± 0.0750.938 ± 0.0680.789 ± 0.1150.973 ± 0.046
F1-Score0.933 ± 0.0710.940 ± 0.0670.860 ± 0.0980.960 ± 0.055
AUC0.952 ± 0.0600.950 ± 0.0620.957 ± 0.0570.969 ± 0.049
MCC0.802 ± 0.1130.828 ± 0.1070.684 ± 0.1310.868 ± 0.096
BORUTA (n = 53)
Accuracy0.957 ± 0.0570.932 ± 0.0710.888 ± 0.0890.938 ± 0.068
Precision0.971 ± 0.0470.971 ± 0.0480.965 ± 0.0520.944 ± 0.065
Recall0.969 ± 0.0490.933 ± 0.0710.873 ± 0.0940.973 ± 0.046
F1-Score0.969 ± 0.0490.948 ± 0.0630.911 ± 0.0800.957 ± 0.057
AUC0.964 ± 0.0530.950 ± 0.0620.957 ± 0.0580.958 ± 0.057
MCC0.905 ± 0.0830.860 ± 0.0980.776 ± 0.1180.858 ± 0.099
Abbreviations: RF, Random Forest; SVM, Support Vector Machine; NB, Naïve Bayes; LR, Logistic Regression.
Table 3. Cross-stage and cross-tissue validation performance of machine learning classifiers.
Table 3. Cross-stage and cross-tissue validation performance of machine learning classifiers.
MetricCross-Stage Validation (Onset → Progression)Cross-Tissue Validation (Onset: Retina → Choroid)
RFSVMNBLRRFSVMNBLR
Accuracy0.9310.9720.8470.4030.9540.8980.8260.618
Precision1.0000.9231.0000.3580.9370.8460.6960.428
Recall0.7921.0000.5421.0000.9250.8550.9250.615
F10.8840.9600.7030.5270.9250.8370.7740.512
AUC0.8960.9790.7710.4480.9460.8860.8270.630
MCC0.8470.9410.6640.1930.8980.7770.6840.226
Abbreviations: RF, Random Forest; SVM, Support Vector Machine; NB, Naïve Bayes; LR, Logistic Regression.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Watcharapalakorn, A.; Poyomtip, T.; Tawonkasiwattanakun, P.; Dewi, P.K.K.; Thomrongsuwannakij, T.; Mahawan, T. Circadian Transcriptomic Dynamics Identify Transferable Retina–Choroid Expression Patterns in Myopia Development via Multistage Machine Learning. Biology 2026, 15, 849. https://doi.org/10.3390/biology15110849

AMA Style

Watcharapalakorn A, Poyomtip T, Tawonkasiwattanakun P, Dewi PKK, Thomrongsuwannakij T, Mahawan T. Circadian Transcriptomic Dynamics Identify Transferable Retina–Choroid Expression Patterns in Myopia Development via Multistage Machine Learning. Biology. 2026; 15(11):849. https://doi.org/10.3390/biology15110849

Chicago/Turabian Style

Watcharapalakorn, Akarapon, Teera Poyomtip, Patarakorn Tawonkasiwattanakun, Putri Krishna Kumara Dewi, Thotsapol Thomrongsuwannakij, and Tanakamol Mahawan. 2026. "Circadian Transcriptomic Dynamics Identify Transferable Retina–Choroid Expression Patterns in Myopia Development via Multistage Machine Learning" Biology 15, no. 11: 849. https://doi.org/10.3390/biology15110849

APA Style

Watcharapalakorn, A., Poyomtip, T., Tawonkasiwattanakun, P., Dewi, P. K. K., Thomrongsuwannakij, T., & Mahawan, T. (2026). Circadian Transcriptomic Dynamics Identify Transferable Retina–Choroid Expression Patterns in Myopia Development via Multistage Machine Learning. Biology, 15(11), 849. https://doi.org/10.3390/biology15110849

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop