2.2. Ethics statement

All animal experimental procedures were approved by the Institutional Animal Care and Use Committee of Peking University Health Science Center. All experiments were performed in accordance with relevant guidelines and regulations.

3.1. mRNA expression analysis of FOLH1

To compare the expression levels of FOLH1 between normal tissues and pan-cancer tissues in TCGA database [25], the TIMER 2.0 online tool (http://timer.comp-genomics.org/) was utilized. We downloaded the uniformly standardized pan-cancer dataset from the UCSC database (https://xenabrowser.net/), extracted FOLH1 gene expression data across samples, and excluded cancer types with fewer than three samples. Using R software (version 3.6.4), the expression differences between normal and tumor samples were calculated for each cancer type. The unpaired Wilcoxon Rank Sum and Signed Rank Tests were applied for significance analysis (P < 0.05 was considered statistically significant), and results were visualized using the Sangerbox online tool (http://sangerbox.com/index.html). Meanwhile, the single-cell distribution of FOLH1 was analyzed using the HPA database.

3.2. Expression correlation analysis

The standardized TCGA Pan-Cancer dataset was downloaded from UCSC. After data filtering and matching, the correlation between FOLH1 expression and patient prognosis under different stratification criteria was analyzed. Pearson correlation coefficients were calculated for each cancer type using R software.

3.3. Correlation of FOLH1 expression with immune cell infiltration and immune-related genes

The online platform ASSISTANT for Clinical Bioinformatics (https://www.aclbi.com/static/index.html) was employed to assess tumor immune cell infiltration levels using six algorithms: TIMER, CIBERSORT, xCell, MCP-Counter, EPIC, and QUANTISEQ. The correlation between FOLH1 expression and infiltration levels of immune cell types, including B cells, CD8⁺ T cells, CD4⁺ T cells, monocytes, neutrophils, MDSCs, mast cells, macrophages, cancer-associated fibroblasts, Tregs, myeloid dendritic cells, and NK cells, was analyzed. Additionally, SangerBox was used to evaluate Pearson correlations between FOLH1 and 150 immune pathway marker genes across five categories: chemokines (41), receptors (18), MHC molecules (21), immunoinhibitors (24), and immunostimulators (46).

3.4. FOLH1 analysis in prostate cancer

Given the pronounced upregulation of FOLH1 in prostate cancer, immunohistochemical staining images of FOLH1 in normal and tumor tissues were obtained from the Human Protein Atlas. Prostate adenocarcinoma (PRAD) data and clinical information were downloaded from TCGA. R packages (ggplot2 and immunedeconv (v2.1.3)) in R software were used to analyze FOLH1 expression in lymph node metastasis subgroups and corresponding immune infiltration. The corresponding single-cell data in.h5 format and annotation results were downloaded from TISCH [26]. The R packages MAESTRO and Seurat (v4.4.0) were utilized to process and analyze the single-cell data. Subsequently, the cells were re-clustered using the t-SNE method. When utilizing Seurat for t-SNE clustering, we meticulously debug and configure the key parameters. During the data preprocessing stage, we optimized the data normalization method to align with the algorithmic characteristics of MAESTRO, thereby enhancing its compatibility with t-SNE clustering and improving the overall clustering results. For cell type annotation, we first identified cell-specific markers based on authoritative literature and databases. For instance, in the context of T cells, we focused on key markers such as CD3, CD4, and CD8, while for macrophages, we emphasized markers like CD68 and CD163, fibroblasts (COL1A1 + , COL3A1+), epithelial cells (EPCAM + , KRT8 + , KRT18+) [26]. Cell clusters were validated through comparison with reference signatures from the PanglaoDB and CellMarker databases. Only cells expressing ≥3 canonical markers for a specific cell type with expression levels >1.5-fold compared to other cell types were definitively annotated. The ESTIMATE package (v1.0.13) calculated stromal, immune, and ESTIMATE scores in PRAD, while the psych package (v2.1.6) determined Pearson correlations between FOLH1 expression and immune infiltration scores.

4.1. Data collection and preprocessing in drug compound activity prediction using machine learning

To develop a robust computational framework for identifying drug candidates with therapeutic potential against prostate cancer, we employed a comprehensive machine learning workflow focused on predicting the activity (expressed as pIC₅₀) of ligands associated with FOLH1, a target protein linked to prostate cancer. Only a limited number of FOLH1 ligands with quantitative activity annotations were available in public databases (420 compounds), the FOLH1 dataset alone was insufficient for robust pIC₅₀ prediction model training. An alternative data expansion strategy was adopted for target selection, we shifted our focus to target proteins exhibiting strong correlations with prostate cancer. These proteins were selected from the ProteinAtlas database [27], a large public repository of human protein data, were selected based on rigorous multi-dimensional criteria: (1) Differential expression in prostate cancer versus normal prostate tissue (fold change ≥ 2.0, p < 0.01); (2) Significant association with clinical outcomes (survival, recurrence, or metastasis) based on Kaplan-Meier analyses and multivariate Cox regression models to identify proteins with prognostic relevance; (3) Confirmed protein-level overexpression by immunohistochemistry in at least 70% of examined prostate cancer tissue samples, providing visual validation of expression patterns across tumor grades and stages; (4) Druggability potential based on structural features and well-characterized structural properties amenable to computational drug discovery, including the presence of defined binding pockets, surface accessibility, and functional domains that can interact with small molecules. These proteins were further filtered based on their protein evidence scores derived from multiple independent sources, categorized into four levels: (1) evidence at the protein level, (2) evidence at the transcript level, (3) no evidence, or (4) not analyzed. For this study, we prioritized target proteins with robust evidence (level 1) of association with prostate cancer to ensure biological relevance and increase the translational potential of our computational predictions.

Ligand data for these selected target proteins were aggregated from multiple chemical databases, including ChEMBL [28], PubChem [29], ChemSpider [30], ZINC [31], BindingDB [32], PDBbind [33], and STITCH [34]. Ligand data, stored in a CSV file containing Simplified Molecular Input Line Entry System (SMILES) strings [35] and corresponding IC₅₀ values, were imported using the Pandas library. Data quality was ensured by employing the Chem.MolFromSmiles function from the RDKit library [36] to filter out invalid SMILES strings and entries with missing or anomalous IC₅₀ values.

4.2. Feature extraction and normalization for effective drug compound activity prediction

Molecular fingerprints were generated using Morgan fingerprints via the RDKit library. These fingerprints, based on atomic neighborhood information, efficiently capture the topological structure of each molecule. In parallel, we computed a series of molecular descriptors to characterize the physicochemical properties of the compounds. These descriptors included: MolLogP, the hydrophobicity of the molecule (partition coefficient, LogP); MolWt, molecular weight; NumRotatableBonds, number of rotatable bonds; TPSA, topological polar surface area; NumHAcceptors, number of hydrogen bond acceptors, NumHDonors, number of hydrogen bond donors. The Morgan fingerprints and molecular descriptors were concatenated to construct a comprehensive feature matrix, serving as the input for the machine learning models.

4.3. Data partitioning, model construction, and performance evaluation

To assess the generalization performance of our predictive models, 80% of the data were allocated to the training set for model development, while the remaining 20% were reserved as the test set. Our machine learning framework incorporates multiple algorithms, with selection guided by several critical criteria: (1) demonstrated benchmark performance in similar ligand-binding prediction tasks; (2) capacity to efficiently process high-dimensional chemical descriptors; (3) interpretability of feature importance; and (4) computational efficiency during both training and inference phases. The selected models include: Ridge Regression [37], a linear regression model incorporating L2 regularization to minimize overfitting by penalizing large coefficients; Linear Support Vector Regression [38] (Linear SVR), a regression model based on support vector machines with a linear kernel, designed to optimize a margin-based loss function; Extra Trees Regressor [39], an ensemble learning model utilizing randomized decision trees, which enhances generalization through random feature selection and sample subsampling; and LightGBM Regressor [40], a gradient boosting framework optimized for efficiency and scalability, particularly effective for large datasets. Additionally, we extended our model comparison to include five other established methods: XGBoost [41], which implements gradient boosting with regularization techniques to prevent overfitting; Graph Neural Networks (GNNs) [42], which directly operate on molecular graph structures to capture complex structural relationships; Convolutional Neural Networks (CNNs) [43], which extract hierarchical features from molecular representations; and Random Forests (RF) [44], which leverage ensemble decision trees with bootstrap aggregation.

To strengthen the robustness of our machine learning validation protocol, we implemented 5-fold cross-validation for all models, ensuring reliable performance assessment across different data subsets. For comprehensive evaluation, we report multiple performance metrics including accuracy, precision, recall, and F1 score, alongside the standard mean absolute error (MAE), RMSE and coefficient of determination (R²). Robustness testing was performed through perturbation analysis, where we systematically varied data partitioning schemes, feature selection thresholds, and hyperparameter settings to assess model stability under different conditions. Each model was trained on the feature matrix derived from the training set, with hyperparameters tuned through grid search to optimize performance across the selected metrics.

4.4. Drug data acquisition and preprocessing

To initiate the drug screening process, we downloaded drug-related data from DrugBank [48], a widely recognized repository of approved and investigational pharmaceuticals. Subsequently, SMILES strings, which provide a text-based representation of molecular structures, were retrieved for drug compounds from multiple chemical databases, including PubChem, ChemSpider, and ChEMBL. To eliminate structurally redundant compounds and prioritize molecular diversity, we applied the Uniform Manifold Approximation and Projection (UMAP) algorithm for dimensionality reduction. Following this screening, the dataset was refined to candidate drugs. After the initial screening, similarity calculations were performed using ChemFusionSimilarity model architecture.

4.5. Molecular docking simulations with AutoDock Vina

The drug molecule exhibiting the highest similarity score to the FOLH1 ligand was selected for molecular docking studies to assess its potential binding affinity with the FOLH1 protein, using AutoDock Vina [49], a robust and widely utilized tool for predicting ligand-protein binding poses and affinities. The docking workflow encompassed the following key stages: (1) Molecular Structure Preparation: Ligand Optimization, initial 3D conformations of the selected drug molecules were generated using the MMFF94 force field, a well-established method for molecular mechanics calculations. These structures underwent energy minimization via the conjugate gradient method, with a convergence threshold of 0.01 kcal/mol. For each drug, the conformer with the lowest energy was chosen as the starting structure for docking, ensuring an energetically favorable configuration. (2) Protein Preparation: The FOLH1 protein structure (PDB: 3LUT) was obtained from the Protein Data Bank (PDB) [50] and processed to optimize its suitability for docking. This preparation involved removing crystallographic water molecules, adding hydrogen atoms, and adjusting the charge distribution using Discovery Studio Visualizer [50] (version 19.1), thereby ensuring compatibility with the docking algorithm. (3) Docking Procedure, a semi-flexible docking strategy was employed, wherein the ligand was allowed conformational flexibility while the protein receptor remained rigid.

The docking site was defined based on the binding pocket of the co-crystallized ligand within the FOLH1 receptor, ensuring biological relevance. AutoDock Vina was configured with a grid box centered at (x = 15.23, y = -8.45, z = 22.10) Å with dimensions of 20 × 20 × 20 Å, an exhaustiveness value of 16, and 9 binding poses were generated for each ligand. Docking outcomes were visualized using Discovery Studio Visualizer and UCSF Chimera [51], allowing for an in-depth examination of ligand-protein interactions. Specific criteria were established to identify hydrogen bonds, including a maximum donor-acceptor distance of 2.50 Å (e.g., O-H), a minimum angle of 120°, and consideration of bimolecular base-induced 1,2-elimination (E2) reactions within the crystallographic context. Binding affinities were quantified by calculating interaction energies, which accounted for molecular conformation, charge distribution, bond angles, and hydrogen bonding contributions. The resulting poses were ranked by their interaction energies, and top-performing conformations were filtered to pinpoint drug candidates with the highest potential for effective binding to FOLH1.

4.6. Clustering analysis of drug molecules using K-means

To investigate the structural and activity-related patterns among drug molecules matched to ligands, we applied the K-means clustering algorithm to a combined dataset of Morgan fingerprints and standardized pIC₅₀ values. Groups data points into clusters based on feature similarity, enabling the identification of molecular subsets with shared structural and functional properties. The optimal number of clusters was determined using two established techniques: Elbow Method [52] and the Silhouette Score Method [53]. We calculated the Sum of Squared Errors (SSE) for a range of cluster numbers. For each candidate cluster number, the Silhouette Score was computed to assess cluster quality. The pIC₅₀ values were standardized (mean = 0, standard deviation = 1) to normalize their scale relative to the Morgan fingerprint features, ensuring equitable contribution to the clustering outcome. PCA projects the data into a lower-dimensional space by maximizing variance retention, providing a view of the drug compounds’ structural and activity-based characteristics.

5.1 Cell culture

LNCaP cells (MeisenCTCC, CTCC-400–0229, China) were maintained in specialized cell culture medium (Procell, CM-0143, China) at 37°C with 5% CO₂. Medium was replaced every 2–3 days, and cells were passaged at 80% confluence. Log-phase cells were used for subsequent experiments.

5.2 Cell proliferation assay

Cells (2,000/well) were seeded in 96-well plates and incubated for 24 hours. Melatonin (MedChemExpress, HY-B0075) was added at varying concentrations. After 24 hours, CCK-8 reagent (Solarbio, CA1210, China) was applied, and absorbance at 450 nm was measured using a microplate reader. Data were analyzed with Excel (Microsoft Office LTSC 2021) and GraphPad Prism (v10.0.2).

5.3. Confocal imaging

Log-phase LNCaP cells were seeded on poly-L-lysine (0.1 mg/mL, Procell, PB180523, China)-coated confocal dishes (NEST, 801001, China). After 24 hours, PBS (Servicebio, G4202-500ml, China) was added to the control group, while melatonin (0, 0.25, 0.5, or 1.0 mM) was added to the experimental group. After 24 hours, fixation with 4% paraformaldehyde (Solarbio, P1110, China), cells were stained with Anti-FOLH1 antibody (abcam, ab76104, USA) and Goat Anti-Rabbit IgG H&L (Alexa Fluor 488, ab150077, USA), then imaged using a Leica TCS-SP8 (Leica) confocal microscope.

5.4. Flow cytometry

Cells were divided into treatment and control groups. The treatment group received melatonin, while the control group received an equivalent volume of PBS, and after 24 hours, the cells were labeled with PE anti-human CD81 (Biolegend, 349505, USA) and APC anti-human FOLH1 (Biolegend, 342507. USA). Isotype controls included PE anti-human CD81 (Biolegend, 349505, USA) and APC Mouse IgG1, κ Isotype Ctrl (Biolegend, 400121, USA). Mean fluorescence intensity (MFI) of FOLH1 was analyzed using a BD Calibur2 flow cytometer.

5.5. Transwell migration and invasion assays

For invasion assays, matrix gel (Beyotime, C0383, China) diluted 1:8 in serum-free medium was added to Transwell chambers (Corning, 3422, USA) and incubated at 37°C for 3 hours. Lower chambers contained 600 µL of 20% FBS medium, while upper chambers received 100 µL of serum-free cell suspension. After 24 hours, cells were fixed with 4% paraformaldehyde, stained with crystal violet, and counted using ImageJ. The Transwell migration assay is similar to the invasion assay, except that the step of spreading the basement membrane gel is omitted. The rest of the methods are the same.

As for the control group, only PBS was added.

5.6. In Vivo Xenograft model and serum melatonin measurement

We randomly assigned nude mice to different experimental groups using a random number table. Nine 5-week-old male nude mice (purchased from the Department of Animal Science of Peking University Health Science Center) were divided into three groups under controlled lighting: normal circadian rhythm (12 h–12 h light–dark, lights on at 06:00 daily), constant darkness (12 h–12 h dark–dark), and disrupted rhythm (random light/dark cycles, placing the animals in altered light-cycle conditions with an 8-h light advance every 2–3 days), with food and water ad libitum When multiple time points were investigated simultaneously, light-tight cabinets were used to shift animals to the respective phase prior to the experiments. Treatment times correspond to Zeitgeber time (ZT) and indicate the timing relative to lights on in the animal facility such that ZT4 is 4 h after lights on, ZT10 is 10 h after lights on, ZT16 is 4 h after lights off and ZT22 is 10 h after lights off. LNCaP cells (5 × 10⁵ cells in 100 µL PBS&matrix gel (Beyotime, C0383, China) were subcutaneously injected. Tumor volume was monitored, and mice were sacrificed on day 20. We dissected the mice to obtain tumors. The tumors were then embedded in paraffin and sectioned. Following this, a primary antibody (Servicebio, GB115710–100, China) was added, along with a secondary antibody (Servicebio, GB25303, China) and DAPI (Servicebio, G1012, China). Finally, the samples were mounted (Servicebio, G1401, China) and immunofluorescence scanning was performed with 3DHISTECH (Pannoramic MIDI, Hungary). Serum melatonin levels were measured at day 18 using an ELISA kit (Meimian, KT2227-B, China). Animal studies were approved by the Peking University Institutional Animal Care Committee (DLASBE0137).

5.7. Hematoxylin and eosin (H&E) staining

The tumor from mice were fixed in 4% paraformaldehyde for 24 hours. Following fixation, the samples were subjected to paraffin embedding and subsequently sectioned into 4-μm slices. For histopathological analysis, the sections were stained using the hematoxylin-eosin staining (Servicebio, G1004, China) according to the standard protocol. The stained sections were then imaged using an optical microscope (E100, Nikon Corporation, Japan).

5.8. Immunohistochemistry

Paraffin sections were dewaxed using an environmentally friendly dewaxing solution and subsequently washed with 100% ethanol (Servicebio, 100092683, China) and distilled water. The sections were then treated with a 10 mM citric acid antigen repair solution (Servicebio, G1202, China) at high power in a microwave oven for 3 minutes to reach the boiling point, followed by treatment at low power for 15 minutes. The sections were allowed to cool naturally to room temperature. After cooling, the sections were washed with PBS solution and then blocked at room temperature for 1 hour using Tris-buffered saline containing 3% bovine serum albumin (Servicebio, GC305010, China) (Servicebio, G1206, China). Following blocking, the samples were incubated with the primary antibody (Servicebio, GB115710, China) at 4 °C overnight. The sections were then washed three times with PBS solution. Subsequently, the samples were incubated with the secondary antibody (Servicebio, GB23303, China) at room temperature for 50 minutes. The treated sections were washed again three times with PBS solution, followed by the addition of DAB chromogenic solution (Servicebio, G1212, China) and hematoxylin. After dehydration and sealing, the sections were prepared for imaging.

6.1. Pan-cancer expression profile of FOLH1

To investigate the expression profile of FOLH1 in tumor versus normal tissues, we initially compared FOLH1 mRNA levels between cancerous and adjacent normal tissues using TCGA database. Analysis revealed significant differences in FOLH1 expression between tumor and normal tissues across 15 cancer types, as well as differential expression between primary tumors and metastatic lesions in SKCM. Notably, FOLH1 exhibited predominant overexpression in tumor tissues across most evaluated cancers, with exceptions observed in BRCA, CHOL, GBM, KICH, KIRP, and LIHC (Fig 2A). Due to limited or absent normal tissue samples for certain cancers in TCGA, we performed expanded analyses integrating TCGA and GTEx datasets. This yielded expression data for 34 cancer types, demonstrating significant FOLH1 upregulation in 17 malignancies and downregulation in 10 (Fig 2B). Through single-cell distribution analysis of FOLH1 expression patterns across various tissues, it has been demonstrated that the primary distribution of FOLH1 shows tumor-type specificity, with predominant localization observed in glandular epithelial cells, squamous epithelial cells, specialized epithelial cells, endothelial cells, and mesenchymal cells, respectively (S1A Fig). Striking upregulation was observed in prostate cancer, underscoring the potential of FOLH1 as a tumor biomarker.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. FOLH1 mRNA Expression Profiles in Human Normal and Tumor Tissues, and Immune Infiltration Patterns.

(A) Boxplots demonstrating FOLH1 mRNA expression levels in normal vs. tumor tissues using TCGA database data. Tumor tissues are denoted by red dots and boxes, while normal tissues are represented by blue dots and boxes. (B) Violin plots illustrating FOLH1 mRNA expression in normal and tumor tissues using combined TCGA+GTEx database data. Tumor tissues are shown in orange boxes and normal tissues in blue boxes. (C) Pan-cancer analysis of Pearson correlation between FOLH1‌‌ expression and age (lollipop plot). (D) Correlation heatmap between FOLH1 expression and immune cell infiltration across cancers (XCELL algorithm-based). Symbols ns, *, **, and *** denote non-significant, P < 0.05, P < 0.01, and P < 0.001, respectively.

https://doi.org/10.1371/journal.pcbi.1014315.g002

In different stratifications, the expression pattern of FOLH1 was analyzed in relation to patient characteristics. The results indicated that FOLH1 was associated with multiple relevant factors of different tumors. Specifically, in prostate cancer, it was observed that FOLH1 expression was higher in cases with lymph node metastasis (S1B Fig), and a significant negative correlation was presented in the age stratification (Fig 2C). In other stratification scenarios, although there was no statistical difference in FOLH1 expression, a significant high expression was observed in all cases.

6.2. FOLH1 expression and immune cell infiltration

The tumor microenvironment (TME) is a complex milieu supporting tumor cell survival, composed predominantly of immune cells, stromal components, and associated intra- and extracellular molecules. Tumor immune cell infiltration represents a critical component of neoplastic ecosystems, closely linked to tumorigenesis, progression, and metastasis. While prior studies have established associations between FOLH1 expression and prognosis, its correlation with immune infiltration remains underexplored. Utilizing multiple immune prediction algorithms, we analyzed the relationship between FOLH1 expression and immune infiltration levels. Heatmap analysis revealed robust associations between FOLH1 and diverse immune cell populations across pan-cancer datasets. Specifically, FOLH1 exhibited positive correlations with immune infiltration in BRCA, HNSC, LUSC, LUAD, LIHC, TGCT, THCA, THYM, UCEC and PRAD. Conversely, negative correlations were observed in COAD, GBM, KIPR, READ and UVM (Figs 2D, S2A-S2E). Leveraging the UCSC database, we identified significant associations between immune-modulating genes and FOLH1 expression across most tumor types (S3 Fig). FOLH1 expression demonstrated widespread positive correlations with both immunosuppressive and immunostimulatory genes. However, in specific cancers including TGCT, LUSC, HNSC, ESCA and NB, FOLH1 displayed negative correlations with MHC molecule-encoding genes. In TGCT, FOLH1 inversely correlated with immunosuppressive genes, while in LUSC and GBM, it showed negative associations with immunostimulatory genes.

6.3. FOLH1 in prostate cancer: expression, single-cell and immune infiltration analysis

Given the significantly elevated expression of FOLH1 in prostate cancer and its association with multiple clinicopathological factors, we focused on prostate cancer to investigate FOLH1 expression patterns. Immunohistochemical staining demonstrated markedly higher FOLH1 expression in prostate cancer tumor tissues compared to normal counterparts (Fig 3A). Further analysis revealed differential FOLH1 expression between the G2 cohort (non-lymph node metastasis group) and G1 cohort (lymph node metastasis group), with both groups showing significantly elevated expression relative to normal controls (Fig 3B). Immune infiltration analysis demonstrated correlations between lymph node metastasis status and infiltration levels of B cells, macrophages, and mast cells (Fig 3C). UMAP visualization revealed that FOLH1 predominantly clustered within epithelial cells, which constituted the major cellular component (Fig 3D). FOLH1 expression exhibited negative correlations with all three immune infiltration scoring metrics (stromal, immune, and ESTIMATE scores) in PRAD (Fig 3E). Prognostic analysis indicated that elevated FOLH1 expression may serve as an adverse prognostic factor in PRAD.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. FOLH1 expression patterns in prostate and colon tissues.

(A) Representative immunohistochemical staining images of FOLH1 in normal prostate tissue vs. prostate tumor tissue, and normal colon tissue vs. colon tumor tissue, these comparisons illustrate the differential expression and diagnostic potential of FOLH1 across different malignancy types. (B) Composite visualization of FOLH1 expression distribution in tumor vs. normal tissues through violin plots, dot plots, and boxplots (significance determined by Kruskal-Wallis test comparing G2 [lymph node metastasis-negative group]; G1 [lymph node metastasis-positive group]). (C) Percentage abundance of tumor-infiltrating immune cells per sample (stratified by lymph node metastasis status), with significance assessed via Kruskal-Wallis test. Asterisks denote significance levels: *p < 0.05, **p < 0.01, ***p < 0.001. (D) Single-cell clustering analysis: Left panel - UMAP plot showing FOLH1 expression distribution across cell types (color-coded by cell type); Color gradient represents expression intensity (darker hues = lower expression, brighter hues = higher expression). Right panel - Bar plot illustrating FOLH1 expression abundance in different cell populations. (E) Pearson’s correlation coefficients between FOLH1 expression and three immune infiltration ‌‌scores in prostate adenocarcinoma (PRAD).

https://doi.org/10.1371/journal.pcbi.1014315.g003

7.1. Identification of target proteins and ligand dataset construction

The number of FOLH1 ligands with IC₅₀ values was limited (n = 420), this dataset alone was insufficient to support reliable pIC₅₀ model training. We identified several prostate cancer-related proteins strongly associated with FOLH1 based on ProteinAtlas evidence scores, including PBOV1 (Prostate and Breast Cancer Overexpressed 1), KLK3 (Kallikrein-Related Peptidase 3), CTAG2 (Cancer/Testis Antigen 2), PSCA (Prostate Stem Cell Antigen), ESR1 (Estrogen Receptor 1), and NKX3–1 (NK3 Homeobox 1). Ligand activity data for these correlated targets were then integrated from multiple chemical databases, yielding an expanded training set of 28,241 ligands for pIC₅₀ prediction.

7.2. Performance of machine learning models

To evaluate the predictive accuracy of ligand activity, expressed as pIC₅₀ values, we compared eight machine learning models: Ridge Regression (Fig 4Aa), Linear SVR (Fig 4Ab), Extra Trees Regressor (Fig 4Ac), LightGBM Regressor (Fig 4Ad), XGBoost, Random Forest, CNN, and GNN (S4A Fig). The Extra Trees Regressor exhibited the strongest alignment with the regression line, with a high density of points concentrated along the diagonal, indicating superior predictive accuracy and consistency. In contrast, Linear SVR displayed a widely dispersed pattern with sparse density, suggesting poor predictive capability. Ridge Regression showed moderate dispersion with a notable number of outliers, while LightGBM Regressor produced an overly clustered distribution near the regression line, albeit with scattered outliers. These observations collectively suggest that the Extra Trees Regressor outperformed the other models in capturing the relationship between experimental and predicted pIC50 values. To further dissect the Extra Trees Regressor’s performance, we examined the correlation of Morgan fingerprints (1024 bits) across the training and test sets (Fig 4B). A correlation heatmap revealed that each fingerprint bit exhibited high self-correlation (visualized in red), indicating strong consistency within individual molecular representations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Comparative Analysis of Machine Learning Models for pIC₅₀ Prediction.

Performance evaluation of machine learning models for predicting ligand activity (pIC₅₀). (A) Scatter plots with Gaussian kernel density estimation showing the correlation between experimental and predicted pIC₅₀ values for four machine learning models: Extra Trees Regressor, Linear SVR, Ridge Regression, and LightGBM Regressor, diagonal line represents perfect prediction. (B) Correlation heatmap of Morgan fingerprints (1024 bits) across training and test sets, with high self-correlation (red) indicating consistency within molecular representations and low inter-fingerprint correlation demonstrating the model’s ability to distinguish unique molecular features. (C) Quantitative performance metrics including Mean Absolute Error (MAE) and coefficient of determination (R²) for both training and test sets across all models. (D) Taylor diagram illustrating the comparative performance based on correlation with experimental data, standard deviation, and root-mean-square error (RMSE), closest to the reference point (correlation = 1.0) represents superior predictive accuracy.

https://doi.org/10.1371/journal.pcbi.1014315.g004

Model performance was quantified using two key metrics: Mean Absolute Error (MAE) and the coefficient of determination (R²). Extra Trees Regressor achieved the lowest MAE values (training set: 0.041, test set: 0.076) and the highest R² scores (training set: 0.932, test set: 0.821), demonstrating exceptional predictive accuracy and model fit across both datasets. Linear SVR recorded the highest MAE (training set: 0.183, test set: 0.189) and lowest R² (training set: 0.541, test set: 0.432), indicating the weakest performance among the models. Ridge Regression produced moderate MAE values (training set: 0.175, test set: 0.180) and R² scores (training set: 0.655, test set: 0.547), reflecting intermediate predictive capability. LightGBM Regressor showed competitive MAE (training set: 0.162, test set: 0.177) and R² (training set: 0.683, test set: 0.601), but its performance was slightly inferior to that of Extra Trees Regressor (Fig 4C). These metrics confirm that the Extra Trees Regressor consistently outperformed the other models, achieving the best tradeoff between accuracy and generalization. To facilitate candidate prioritization, we additionally evaluated model performance on a secondary binary classification task, in which pIC₅₀ values were discretized into high-activity and low-activity classes (Section 2.3.2 for thresholds). Further classification metrics revealed the Extra Trees Regressor’s superior performance, with cross-validation results showing accuracy of 0.9452, precision of 0.9374, recall of 0.9429, and F1-score of 0.9351, while test results demonstrated accuracy of 0.9126, precision of 0.9074, recall of 0.9206, and F1-score of 0.9141 (Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Comparison of Cross-Validation and Test Set Performance Metrics for Machine Learning Models in Effective Drug Compound Activity Prediction.

https://doi.org/10.1371/journal.pcbi.1014315.t001

A Taylor diagram was employed to visualize the comparative performance of the four models based on their standard deviation, correlation with experimental data, and root-mean-square error (RMSE). The Extra Trees Regressor was positioned closest to the reference point (correlation = 1.0, standard deviation normalized to 0.3), signifying its superior alignment with experimental pIC₅₀ values and minimal predictive error. LightGBM Regressor and Ridge Regression also approached the 1.0 correlation axis but were located at greater distances (LightGBM beyond 0.6, Ridge Regression between 0.3 and 0.6), indicating higher variance and error. Linear SVR, positioned between 0.3 and 0.6 but far from the 1.0 correlation line (Fig 4D), underscored its poor predictive reliability.

Perturbation analysis was conducted to assess the robustness of each model to controlled input perturbations (S4A Fig). The results revealed significant differences in model stability. Most models, including Ridge, LinearSVR, RandomForest, LGBMRegressor, XGBoost, SVM, and, where applicable, CNN and GNN, exhibited relatively large fluctuations in MAE and R² when subjected to perturbations. These fluctuations suggest sensitivity to input variations, which could impact their reliability in real-world scenarios with noisy or incomplete data. In contrast, the Extra Trees Regressor demonstrated remarkable stability. Its MAE and R² metrics showed negligible fluctuations under perturbation, with MAE rising only gradually and R² declining slowly as perturbation intensity increased. The final hyperparameter configurations for all machine learning models used in this study are detailed in Table 2, which includes both tuned parameters, optimized via grid search during 5-fold cross-validation, and default parameters.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Final Hyperparameter Configurations for Machine Learning Models in Drug Compound Activity Prediction.

https://doi.org/10.1371/journal.pcbi.1014315.t002

7.3. Clustering and dimensionality reduction of FOLH1 Ligand Dataset

To elucidate the structural and physicochemical diversity within the FOLH1 ligand dataset (n = 420, detailed in S1 File), we integrated pIC₅₀ values with molecular descriptors characterizing structure and properties. This dataset was subjected to K-means clustering, dimensionality reduction, and subsequent computational analysis. The ligands were classified into three activity categories based on pIC₅₀ thresholds: low activity (pIC₅₀ < 5), medium activity (5 ≤ pIC₅₀ ≤ 7), and high activity (pIC₅₀ > 7). Among these, high-activity ligands (pIC₅₀ > 7) were the most abundant, while low-activity ligands (pIC₅₀ < 5) were the least represented. Medium-activity ligands (5 ≤ pIC₅₀ ≤ 7) fell between these extremes in quantity. A t-SNE plot (Fig 5Aa) was generated to visualize the distribution of these ligands in a two-dimensional space. High-activity ligands exhibited a broad spread, indicating substantial structural diversity. In contrast, medium-activity ligands formed a more compact cluster, suggesting greater structural similarity within this group. Low-activity ligands were sparsely distributed, consistent with their smaller population. The majority of ligands had LogP values below 5, with the highest frequency observed for LogP < 2 (Hydrophobicity, Fig 5Ab). Only six ligands exhibited LogP > 5. The t-SNE visualization revealed that ligands with LogP < 2 and LogP between 2 and 5 formed dense clusters, whereas those with LogP > 5 were markedly dispersed, reflecting their rarity and structural divergence. Most ligands had molecular weights below 500 Da, constituting the predominant group. In the t-SNE plot, these ligands appeared tightly clustered, indicating homogeneity (MolWt, Fig 5Ac). The number of hydrogen bond donors was predominantly below 5, with a notable subset below 2 (NumHDonors, Fig 5Ad). Ligands with more than 5 hydrogen bond donors were infrequent.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Molecular Descriptor Analysis and Clustering of FOLH1 Ligands.

(A) t-SNE visualization of ligand distribution colored by: (a) activity categories (low: pIC₅₀ < 5; medium: 5 ≤ pIC₅₀ ≤ 7; high: pIC₅₀ > 7), showing high-activity ligands with broad structural diversity, medium-activity ligands forming compact clusters, and sparse distribution of low-activity ligands; (b) LogP values (< 2, 2-5, > 5), revealing dense clustering for LogP < 5 compounds and dispersion for the rare LogP > 5 compounds; (c) molecular weight (< 500 Da, > 500 Da), demonstrating tight clustering of predominant lower-weight compounds and sparse distribution of higher-weight molecules; (d) number of hydrogen bond donors (< 2, 2-5, > 5), showing dense clusters for compounds with fewer donors and sparse distribution for those with > 5 donors. (B) Pairwise relationship plots of key molecular descriptors (MolLogP, MolWt, NumRotatableBonds, TPSA, NumHAcceptors, NumHDonors) colored by K-means clusters, illustrating three distinct clusters with separation based on physicochemical properties.

https://doi.org/10.1371/journal.pcbi.1014315.g005

K-means clustering was applied to the standardized molecular descriptor data, resulting in the identification of three distinct clusters. These clusters were subsequently analyzed using pairwise relationship plots generated with Seaborn’s Pairplot functionality (Fig 5B). The plots encompassed key descriptors: MolLogP (hydrophobicity), MolWt (molecular weight), NumRotatableBonds (number of rotatable bonds), TPSA (topological polar surface area), NumHAcceptors (number of hydrogen bond acceptors), and NumHDonors (number of hydrogen bond donors). The pairwise visualizations delineated the three clusters, revealing their separation based on descriptor values. For instance, the MolLogP plot highlighted a cluster with predominantly low LogP values (< 2), while the MolWt plot distinguished a tightly grouped cluster of ligands below 500 Da from more scattered outliers. Similarly, NumHDonors and NumHAcceptors plots underscored the prevalence of ligands with fewer hydrogen-bonding groups, with distinct clustering patterns emerging across all descriptors.

7.4. Comparative analysis of ChemFusionSimilarity and Tanimoto similarity metrics

To assess the performance of the deep learning-based ChemFusionSimilarity model relative to the traditional Tanimoto similarity metric, we merged both similarity scores into a unified dataframe (S2 File). This dataset enabled a comprehensive comparison of the two approaches across multiple visualization techniques, providing insights into their distributional properties and predictive outcomes. A 3D density plot was constructed to visualize the relationship between ChemFusionSimilarity and Tanimoto similarity scores (S4B Fig). The ChemFusionSimilarity values predominantly ranged from 0.16 to 0.22, while Tanimoto similarity scores were largely concentrated between 0.10 and 0.14. Notably, the plot revealed no significant outliers or dispersion, indicating a consistent and compact distribution for both metrics. A joint distribution plot depicted the bivariate relationship between ChemFusionSimilarity and Tanimoto similarity, alongside their marginal distributions (S4C Fig). The joint plot demonstrated a strong linear correlation between the two metrics, with all data points aligning closely along the regression line and no observable outliers. The marginal distributions further confirmed the concentration of ChemFusionSimilarity scores (0.16–0.22) and Tanimoto scores (0.10–0.14), reinforcing the absence of discrete or aberrant values. This alignment highlights the consistency between the deep learning-based and traditional similarity measures, while also underscoring ChemFusionSimilarity’s tendency to assign higher similarity scores. Violin plots were employed to compare the distributional characteristics of ChemFusionSimilarity and Tanimoto similarity scores (S4D Fig). Compounds identified via ChemFusionSimilarity are consistently associated with higher pIC50 values than those retrieved using conventional structural similarity alone. Following the similarity analysis, the ChemFusionSimilarity model was used to predict IC₅₀ values for the ligand-drug pairs. The resulting distribution of predicted IC₅₀ values was visualized (S4E Fig), showing that the majority fell within the range of 300–400 nM, indicative of moderate to high potency. A smaller subset of predictions yielded IC₅₀ values below 150 nM, representing ligands with exceptionally high activity.

We examined the relationship between Topological Polar Surface Area (TPSA) and drug compound activity (pIC₅₀) as shown in S4F Fig. Compounds with TPSA values greater than 125 Ų consistently demonstrated lower activity, with pIC₅₀ values predominantly falling between 5 and 6. In contrast, compounds with TPSA values below 125 Ų exhibited substantially higher activity, with pIC₅₀ values typically ranging from 7 to 9. This pronounced threshold effect suggests that excessive polar surface area may hinder membrane permeability or target binding, highlighting TPSA as a crucial parameter for optimizing drug compound potency. To quantify the relative importance of various molecular descriptors in predicting compound activity, we conducted feature importance analysis (S4G Fig). The results identified Complementary Property Index (CPI) as the most influential descriptor with a coefficient magnitude of 3.24, indicating its substantial positive correlation with compound potency. In contrast, LogP exhibited the lowest impact with a coefficient magnitude of -0.41, suggesting a mild negative correlation between lipophilicity and activity within our dataset. Hydrogen bond donor count also demonstrated a negative correlation with activity, with a coefficient magnitude of -0.4, emphasizing the molecular properties that most significantly influence compound activity against the target of interest.

7.5. Drug candidate screening and similarity matching

12,457 drug compounds from the DrugBank database was refined to 1,819 candidate drugs (S3 File). Among the 1,819 candidates, 36 exhibited similarity scores exceeding 0.8 with FOLH1 ligands (S4 File), indicating strong structural resemblance. From this subset, four drugs with the highest similarity scores were prioritized for detailed investigation: Glutathione, Methotrexate, Melatonin, and Estrone (Fig 6A). The docking results provided quantitative binding affinities (expressed in kcal/mol) and detailed insights into the intermolecular interactions stabilizing each ligand-protein complex (Fig 6B). Glutathione (Binding Affinity: -11.078 kcal/mol) exhibited the lowest predicted binding energy among the tested candidates, indicating a favorable theoretical interaction. The complex featured one hydrophobic contact, seven hydrogen bonds, and two salt bridges. The predominance of hydrogen bonds and ionic salt bridges suggests a highly stable interaction profile, likely driven by polar and charged residues within the FOLH1 binding pocket. Methotrexate (Binding Affinity: -10.236 kcal/mol) also demonstrated a comparable predicted binding affinity. This complex included one hydrophobic contact, nine hydrogen bonds, one pi-stacking interaction, and one pi-cation interaction. The extensive hydrogen bonding network, complemented by aromatic and cationic interactions, underscores Methotrexate’s capacity for robust and multifaceted engagement with FOLH1. Melatonin (Binding Affinity: -9.681 kcal/mol), indicative of favorable binding. The docking revealed six hydrogen bonds as the primary stabilizing forces. Estrone (Binding Affinity: -9.670 kcal/mol), closely comparable to Melatonin. This complex exhibited four hydrophobic contacts, five hydrogen bonds, and one pi-stacking interaction. The balanced contribution of hydrophobic and hydrogen-bonding interactions, enhanced by aromatic stacking, indicates a versatile binding mechanism tailored to the FOLH1 active site. It is important to note that the differences in predicted binding affinities among these top four candidates (ranging from -9.670 to -11.078 kcal/mol) are relatively small and fall within the expected uncertainty margin of standard molecular docking algorithms (typically ±1–2 kcal/mol).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Top Drug Candidates and Molecular Docking Analysis with FOLH1.

(A) Chemical structures of the four highest-scoring drug candidates: Glutathione, Methotrexate, Melatonin, and Estrone, selected based on their exceptional similarity scores with known FOLH1 ligands. (B) Molecular docking results visualized using Discovery Studio Visualizer and UCSF Chimera, showing binding poses and interaction networks for each candidate with FOLH1. The visualization highlights the spatial orientation of each ligand within the FOLH1 binding pocket and specific residues involved in key intermolecular interactions.

https://doi.org/10.1371/journal.pcbi.1014315.g006

7.6. Clustering analysis of candidate drugs

The optimal number of clusters was determined using two complementary methods, the Elbow Method and the Silhouette Score Method. The Elbow Method plotted the SSE against a range of cluster numbers, identifying an inflection point at three clusters with an SSE of 450, indicating a balance between clustering tightness and model complexity. The Silhouette Score Method corroborated this finding, with the highest score of 0.2 achieved for three clusters (Fig 7A). Principal Component Analysis (PCA) was employed to reduce the dimensionality of the clustered data into a two-dimensional space (Fig 7B). Cluster 2 was predominantly distributed in the upper-right quadrant and included three noteworthy drugs: Glutathione, Melatonin, and Estrone, suggesting shared structural or activity traits. Cluster 1 was concentrated in the left region, encompassing a broader set of candidates, while Cluster 3 consisted solely of Methotrexate, positioned distinctly from the others. This separation highlights Methotrexate’s unique molecular profile compared to the other clusters. A heatmap of Morgan fingerprint bits (1024-bit representation) revealed cluster-specific patterns (Fig 7C). Cluster 1 exhibited the highest frequency at bit 80, indicating a dominant structural motif. Cluster 2 showed frequent activation across multiple bits (80, 272, 416, 656, 880), suggesting greater structural diversity. Cluster 3, comprising only Methotrexate, was most frequent at bit 688. Common high-frequency bits across all clusters (80, 144, 384, 800, 896) indicate shared structural features, while cluster-specific patterns underscore their distinctiveness. Cluster 1 exhibited predominantly medium biological activity, with pIC₅₀ values ranging from 0.5 to 0.7, and molecular weights (MolWt) between 200 and 350 Da. This cluster was further defined by a higher number of hydrogen bond donors, typically ranging from 3 to 7, and LogP values spanning -3–0, indicative of greater hydrophilicity. Additionally, the number of hydrogen bond acceptors varied from 5 to 7 (Fig 7D). Cluster 2 comprised compact, potent drugs with pIC50 values exceeding 7, indicating strong binding affinity (typically corresponding to IC50 values < 100 nM). The molecular weights of drugs in this cluster ranged from 200 to 350 Da, aligning well with the desirable range for oral bioavailability as per Lipinski’s Rule of Five. The number of hydrogen bond donors varied from 3 to 7, while hydrogen bond acceptors ranged from 5 to 7, suggesting a balanced capacity for polar interactions critical for target binding. LogP values spanned from -3–0, indicating moderate hydrophilicity, which is advantageous for solubility and membrane permeability in drug-like molecules. Notable drugs in Cluster 2 included Glutathione (307.08 Da, 6 HBD, 6 HBA, LogP -2.21), Melatonin (149.05 Da, 2 HBD, 3 HBA, LogP 0.15), and Estrone (270.16 Da, 1 HBD, 2 HBA, LogP 3.82). These molecules, despite their structural diversity, shared compact scaffolds and high potency, underscoring the cluster’s enrichment with drug-like candidates. Cluster 3, represented solely by Methotrexate, also displayed high activity (pIC₅₀ > 7) but was distinguished by a significantly larger molecular weight of 450 Da. This cluster featured a moderate number of hydrogen bond donors (NumHDonors = 5) and a balanced LogP value of approximately 0, suggesting an equilibrium between hydrophilicity and hydrophobicity. With 10 hydrogen bond acceptors, the highest among the clusters, Cluster 3 highlighted Methotrexate as a larger, highly active compound with an extensive capacity for hydrogen-bond interactions (Fig 7D).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Clustering Analysis and Structural Characterization of FOLH1-Targeted Drug Candidates.

(A) Cluster optimization using the Elbow Method (Sum of Squared Errors) and Silhouette Score Method. (B) PCA showing the distribution of the three clusters in two-dimensional space: Cluster 2 (upper-right quadrant) containing Glutathione, Melatonin, and Estrone; Cluster 1 (left region) encompassing the majority of candidates; and Cluster 3 (distinct position) consisting solely of Methotrexate. (C) Heatmap visualization of Morgan fingerprint bits (1024-bit representation) revealing cluster-specific structural patterns. (D) Physicochemical property analysis by cluster. (E) Adverse Drug Reaction risk analysis showing predominantly Low risk classification across all 36 candidates, with a small fraction categorized as Medium risk and none as High risk. (F) Maximum Common Substructure (MCS) analysis revealing a carbonyl group (C = O) as a common scaffold among moderately active drugs, while high-activity drugs consistently contained a hydroxyl group (OH), suggesting its importance for enhanced potency.

https://doi.org/10.1371/journal.pcbi.1014315.g007

An Adverse Drug Reaction (ADR) risk analysis was conducted for all 36 candidates (Fig 7E). The majority of drugs were classified as Low risk, with only a small fraction categorized as medium risk. No candidates exhibited High risk, suggesting a generally favorable safety profile across the dataset. This assessment enhances the therapeutic potential of the identified drugs, particularly the high-activity candidates in Clusters 2 and 3. The MCS was computed for each cluster. The MCS featured a substructure with an oxygen atom doubly bonded to a carbonyl group, reflecting a common scaffold among moderately active drug compounds. High-activity drugs consistently included a hydroxyl group (OH) within their MCS, suggesting its role as a critical functional group for enhanced potency (Fig 7F).

8.1. Melatonin suppresses FOLH1 expression and inhibits invasion and metastasis of LNCaP cells in vitro

To determine the optimal melatonin concentration for subsequent experiments, CCK-8 assays determined an IC₅₀ of 1.063 mM melatonin for LNCaP cells (S5A Fig). Based on these results, the maximum melatonin concentration applied in experiments was set at 1.0 mM. The effects of melatonin on FOLH1 expression were evaluated using confocal imaging and flow cytometry. Cells were treated with melatonin at concentrations of 0, 0.25, 0.5, or 1.0 mM. Fig 8A and 8B demonstrate a concentration-dependent, albeit modest, reduction in cellular fluorescence intensity, with the highest tested concentration of 1.0 mM melatonin inducing an 11.98% decrease. Flow cytometric analysis (Fig 8C and 8D) revealed a progressive decline in both the proportion of FOLH1-positive cells and the MFI with increasing melatonin concentrations. Furthermore, Transwell assays demonstrated dose-dependent suppression of LNCaP cell invasion and migration by melatonin (Fig 8E). Invasion and migration rates were quantified using ImageJ software and plotted in Fig 8F (n = 3, p < 0.05).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Melatonin suppresses FOLH1 expression and inhibits LNCaP metastasis/invasion.

(A) Confocal images of LNCaP cells treated with different melatonin concentrations (0; 0.25; 0.5; 1.0 mM (a-d)), DAPI = blue, FOLH1 = green. (B) Statistical analysis of FOLH1 fluorescence intensity across concentrations (n = 5). C, D. Flow cytometric quantification of FOLH1 signal intensity (n = 3): (C) Left to right - isotype control, 0; 0.25; 0.5; 1.0 mM melatonin treatments. E, F. Invasion/migration assays: Control vs. melatonin-treated cells were seeded in Transwell upper chambers for 24H. Migrated cells were fixed with 4% paraformaldehyde, imaged, and counted from 5 random fields (bar plots). G-I. Circadian rhythm-dependent melatonin variations and corresponding tumor volumes. J, K, L. FOLH1 immunofluorescence staining, Pathological sections (HE staining), immunohistochemistry and intensity quantification in prostate cancer under different circadian conditions. Data presented as mean ± SD. *p < 0.05, **p < 0.01, ***p < 0.001.

https://doi.org/10.1371/journal.pcbi.1014315.g008

8.2. Circadian Rhythm Disruption Alters Melatonin Levels and FOLH1 Expression

To further validate the effects of melatonin on FOLH1 expression and tumor progression in vivo, we established xenograft models in nude mice subjected to distinct light-dark cycles and performed melatonin rhythm assessments. The experimental design comprised three groups of nude mice with varying photoperiods. LNCaP cells were implanted subcutaneously into these mice, and serum melatonin levels were measured at designated timepoints alongside tumor volume monitoring. Finally, we euthanized the mice and harvested tumor tissues at ZT16 (Zeitgeber Time 16). Altered light conditions significantly modulated melatonin levels: the normal circadian group exhibited robust diurnal melatonin fluctuations with higher overall levels, whereas the constant darkness group displayed reduced melatonin levels (though retaining residual rhythmicity). The disrupted rhythm group showed intermediate melatonin levels with abolished circadian patterns (Figs 8G, S5B).

Correspondingly, tumor growth was slowest in the circadian group, accompanied by the lowest FOLH1 fluorescence intensity. Conversely, the disrupted rhythm group demonstrated accelerated tumor growth and the highest FOLH1 fluorescence intensity (Fig 8H-8L). While limited by the small sample size (n = 3 per group), these preliminary findings suggest a potential association between endogenous melatonin levels, tumor growth, and FOLH1 expression dynamics.