Screening of Natural Inhibitors from Pleurotus ostreatus against PHGDH for Enhancing Antiviral Innate Immunity through Mendelian Randomization, Machine Learning, Molecular Docking, Molecular Dynamics Simulation, and Binding Free Energy Calculation

Abstract

The human innate immune response plays a critical role in limiting and controlling viral infections. Consequently, therapeutic strategies that promote or modulate innate antiviral immunity are regarded as promising approaches to combating viral pathogens. Recent animal studies suggest that inhibition of serine biosynthesis may enhance host antiviral innate immunity, thereby providing novel targets for drug development. This study aims to identify natural compounds derived from Pleurotus ostreatus that inhibit serine biosynthesis and consequently strengthen antiviral innate immunity using a machine learning model integrated with molecular-docking-based virtual screening. First, we identified PHGDHone of the key enzymes in the serine biosynthesis pathwayas a potential therapeutic target through Mendelian randomization (MR) analysis. The MR results provided suggestive evidence that PHGDH inhibition is associated with elevated IFN-β levels (OR = 0.862, 95% CI: 0.786–0.955, p = 0.004), supporting its immunomodulatory role and highlighting its therapeutic potential. Next, we developed a deep neural network model based on RDKit descriptors to predict the inhibitory activity of natural compounds from Pleurotus ostreatus against PHGDH, achieving a sensitivity of 96.3%, specificity of 98.2%, F1 score of 0.976, accuracy of 97.2%, and AUROC of 0.997. By combination of this predictive model with molecular-docking-based virtual screening, five compounds (FDB024162, FDB024165, FDB095731, FDB095733, and FDB095735) were prioritized as potent PHGDH inhibitors. Furthermore, molecular dynamics simulations and binding free energy calculations revealed that these compounds form stable complexes with PHGDH and bind specifically to key residues within the cofactor-binding pocket. In conclusion, the identified natural compounds exhibit a strong potential to inhibit PHGDH and thereby enhance antiviral innate immunity. Further experimental validation is warranted to fully characterize their functional contributions to innate immune activation.

1. Introduction

Infectious diseases caused by viruses have long represented a major threat to global health. The recurring emergence of viral epidemics highlights the pressing need for broad-spectrum antiviral therapies. Such agents can play a crucial role in the emergency management of emerging viral infections by rescuing critically ill patients and substantially reducing mortality, thus contributing to outbreak control. , Direct-acting antiviral agents (DAAs), which target viral proteins to disrupt the viral life cycle, are widely employed as a therapeutic strategy due to their high efficacy and specificity. However, DAAs are associated with significant limitations. Viruses possess inherently error-prone genomes that facilitate rapid mutation and the generation of viral quasispecies. Under selective pressure from DAAs, resistance-conferring mutations can arise, enabling viruses to evade inhibition. Over time, the accumulation of such mutations may lead to the emergence of drug-resistant strains, undermining the sustained efficacy of DAA-based treatments. Furthermore, DAAs require precise identification of specific viral targets, which restricts their applicability in outbreaks caused by novel or poorly characterized pathogens. In contrast, host-targeted antiviral strategiesparticularly those that modulate the host immune responseare expected to confer broader antiviral activity. By inducing a set of pleiotropic genes that collectively suppress viral replication, these approaches are less prone to driving the development of drug resistance. −

Innate immunity serves as the host’s primary defense against viral infection. Upon viral invasion, the innate immune system is rapidly activated, with type I interferons (IFNs) and their downstream signaling pathways playing a central role in antiviral defense. Type I IFNs include a single subtype of IFN-β and multiple subtypes of IFN-α, which are expressed in all nucleated cells and are essential for antiviral immunity. IFN-β activates the JAK-STAT pathway, promoting the expression of various inflammatory cytokines such as TNF-α and IL-6, thereby initiating a robust antiviral immune response. , Recent studies indicate that host metabolic processes and endogenous metabolitesincluding serine and the serine biosynthesis pathwaycan regulate antiviral innate immunity. − Serine, a nonessential amino acid, supports cell proliferation and can be synthesized de novo or absorbed from the extracellular environment. The first step in serine biosynthesis begins with 3-phosphoglycerate, an intermediate of glycolysis. Catalyzed by phosphoglycerate dehydrogenase (PHGDH), 3-phosphoglycerate is oxidized using NAD⁺ to form 3-phosphohydroxypyruvate. This intermediate is then transaminated by phosphoserine aminotransferase (PSAT), with glutamate serving as the amino donor, to yield 3-phosphoserine. Finally, phosphoserine phosphatase (PSPH) catalyzes the dephosphorylation of 3-phosphoserine to produce serine. Serine contributes to one-carbon metabolism, a network of interconnected pathways that supply one-carbon units for the biosynthesis of nucleotides, S-adenosylmethionine, NADPH, and glutathione. Notably, recent evidence shows that the serine biosynthesis pathway (SSP) suppresses IFN-β-mediated antiviral innate immunity by inhibiting the TBK1–IRF3 signaling axis. Genetic deletion or knockdown of PHGDHthe rate-limiting enzyme in serine synthesissignificantly enhances IFN-β production in response to both RNA and DNA viruses both in vitro and in vivo. Similarly, pharmacological inhibition of PHGDH with CBR-5884 or restriction of exogenous serine and glycine uptake enhances TBK1 activation, promotes nuclear translocation of phosphorylated IRF3, boosts IFN-β-mediated antiviral immunity, and protects mice from lethal viral infection. These findings suggest that inhibition of serine biosynthesis represents a promising therapeutic strategy for enhancing host antiviral innate immunity and provides novel targets for improving resistance to viral infections.

Pleurotus ostreatus, a member of the Basidiomycota phylum and the genus Pleurotus, is widely cultivated and consumed as a food source in China, Egypt, and other regions. Extensive studies have demonstrated that P. ostreatus possesses diverse pharmacological activities, including immune regulation, , antiviral effects, − anti-inflammatory properties, − antioxidant activity, − and antihypercholesterolemic effects. , Its broad bioactivity is attributed to its rich content of bioactive components such as polysaccharides, flavonoids, unsaturated fatty acids, terpenes, and phenolic compounds, which have attracted increasing scientific interest. Natural compounds derived from P. ostreatus exhibit structural diversity, chemical novelty, and unique mechanisms of immune modulation, making this species a promising source for discovering novel immunologically active agents.

With the accumulation of large-scale biological and chemical data, machine learning technologies have demonstrated significant potential in drug discovery and development. Various computational approaches, particularly advanced machine learning algorithms, are now widely employed to identify compounds with therapeutic potential. , Meanwhile, computer-aided drug design (CADD) enables the simulation, calculation, and prediction of receptor–ligand interactions, facilitating the identification and optimization of lead compounds and thereby improving the efficiency of drug development. The integration of machine learning models with CADD techniques has proven effective in screening candidate drugs and refining subsequent stages of the discovery pipeline, significantly reducing both time and cost. ,

The aim of this study is to discover bioactive compounds derived from P. ostreatus that target the serine biosynthesis pathway to enhance antiviral innate immunity using a combined approach of machine learning and molecular-docking-based virtual screening. First, we identified PHGDHone of the key enzymes in serine biosynthesisas a therapeutic target for enhancing antiviral innate immunity through Mendelian randomization analysis. Next, five natural compounds capable of inhibiting PHGDH were identified by integrating machine learning prediction with molecular docking. Finally, the stability and binding modes of these five compounds complexed with PHGDH were characterized by using molecular dynamics simulations and MM-PBSA binding free energy calculations.

2. Materials and Methods

2.1. Mendelian Randomization

2.1.1. Exposure Data and Genetic Instruments

A total of three human serine biosynthesis related protein-coding genes (PSAT1, PHGDH, and PSPH) were obtained from Uniprot. The expression quantitative trait locus (eQTL) data for the three genes in blood downloaded from the eQTLGen consortium (https://www.eqtlgen.org/cis-eqtls.html) were used to proxy the mRNA expression level of the three genes.

2.1.2. Outcome Data

The IFN-β level was selected to proxy host antiviral innate immunity. The summary-level genome-wide association study (GWAS) data for IFN-β level in blood were extracted from the NHGRI-EBI Catalog of human genome-wide association studies. The IFN-β discovery cohort (GCST90161320) was composed of 2935 individuals obtained from Thareja et al. The IFN-β replication cohort (GCST90241552) was composed of 3301 individuals obtained from Sun et al. The detailed information on the data sources is displayed in Table S1.

2.1.3. Mendelian Randomization Analysis

Two-sample Mendelian randomization (MR) analysis utilizing GWAS summary-level data was conducted to investigate the causal association between genes related to human serine biosynthesis and IFN-β levels. The cis-eQTL single-nucleotide polymorphisms (SNPs) significantly associated with the three genes (p < 5.0 × 10^–8) were selected as instrumental variables, ensuring that the selected SNPs exhibited linkage disequilibrium (r ² < 0.1). Furthermore, the F-statistic for each SNP exceeded 10. The inverse-variance weighted (IVW) method was employed as the principal MR analysis approach to estimate the causal effect of a 1-SD change in expression levels of the genes on the IFN-β level. The MR-Egger, weighted mode, and weighted median methods were used as supplementary analysis. To address multiple testing, p < 0.016 (0.05/3) with Bonferroni correction was considered as significance level of three genes. The presence of horizontal pleiotropy among genetic variants was assessed through both the MR-Egger intercept test and Pleiotropy Residual Sum and Outlier (MR-PRESSO) test. Heterogeneity among genetic variants was evaluated using Cochran’s Q test in conjunction with both IVW and MR-Egger methods. The directionality of causality was explored via the MR Steiger directionality test. Leave-one-out analysis served as a sensitivity analysis method aimed at enhancing the robustness of our MR findings. The MR analysis process was executed by the TwoSampleMR and mrpresso R package.

2.2. Machine Learning Model Training and Predictions

2.2.1. Data Collection

A list of compounds exhibiting inhibitory activity (IC₅₀) against human PHGDH was obtained from the BindingDB database (https://www.bindingdb.org/rwd/bind/index.jsp). Compounds with an IC₅₀ value of ≤ 10 μM were classified as "active compounds", while those with an IC₅₀ value >10 μM were designated as "inactive compounds". A total of 890 compounds were randomly divided into two data sets: a training set comprising 80% of the data and a test set consisting of the remaining 20%. The training set was utilized to develop a machine learning model capable of predicting the inhibitory activity of compounds, whereas the test set served to validate the predictive model.

2.2.2. Calculation of Molecular Fingerprints/Descriptors

Three types of molecular fingerprints including Mol2Vec fingerprints, Molecular ACCess System (MACCS) Keys, and Morgan Fingerprints and one type of molecular descriptors including RDKit descriptors were selected as features of compounds for developing the machine learning model. Calculation of molecular fingerprints/descriptors of the 890 compounds was accomplished by the DeepChem package in python.

2.2.3. Machine Learning Model Construction and Evaluation

Four machine learning algorithms including Gradient Boosting with Component-wise Linear Models (glmboost), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Deep Neutral Networks (DNN) based on Mol2Vec fingerprints, MACCS Keys, Morgan Fingerprints, and RDKit descriptors of compounds were used to develop predictive model in the training set. The parameters of each algorithm were the following:

(1) glmboost model: The glmboost models were trained using the mboost R package. The parameters of the glmboost model were set: family = AdaExp­(), and mstop = 350, and other parameters were defaults.

(2) SVM model: The SVM models were developed using the e1071 R package. The parameters of SVM model were as follows: kernel = "linear", cost = 0.25, and probability = TRUE, and other parameters were defaults.

(3) XGBoost model: The XGBoost models were trained using the xgboost R package. The parameters of the XGBoost model were max_depth = 1, eta = 0.4, nthread = 20, nrounds = 150, gamma = 0, colsample_bytree = 0.8, min_child_weight = 1, subsample = 0.75, and objective = "binary:logistic", and other parameters were defaults.

(4) DNN model: The DNN models were trained using teh h2o R package. The parameters of the DNN model were activation = "RectifierWithDropout", loss = "CrossEntropy", hidden = c­(600,400,400,400), epochs = 5000, input_dropout_ratio = 0.1, hidden_dropout_ratios = c­(0.5,0.5,0.5,0.5), and balance_classes = TRUE, and other parameters were defaults.

The predictive performance of the four machine learning models comprised accuracy, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUROC) with leave-one-out cross-validation (LOOCV) in thbe training set using caret and pROC R packages. Next, the predictive performance of the developed models was validated in the test set using accuracy, sensitivity, specificity, F1 score, and AUROC. After comparing the performance of the developed models, the model with the best performance was selected as the PHGDH inhibitory activity prediction model. Hit rate (HR) and enrichment factor (EF) were used to evaluate the screening efficiency of machine learning model based virtual screening.

2.2.4. Prediction of Compound Activity

A total of 4474 natural compounds originated from Pleurotus ostreatus were obtained from The Food Database (FOODB) (https://foodb.ca/foods/FOOD00553). The PHGDH inhibitory activity of the 4474 natural compounds was predicted by using the developed machine learning model.

2.3. Molecular Docking

The natural compounds originated from Pleurotus ostreatus screened from ML-based virtual screening were subjected to docking with PHGDH. The 3D structure files of compounds were generated and minimized using the open Babel toolbox. The crystallographic structure of PHGDH homodimers containing A chain and B chain was downloaded from the RCSB Protein Data Bank (RCSB PDB) (ID: 5ofw, resolution: 1.50 Å) (https://www.rcsb.org/structure/5OFW). The molecular docking was accomplished using the LeDock software. The protein was prepared with removal of B chain, water molecular, and cocrystallized ligand, adding hydrogen atoms and generating a docking input file for LeDock by LePro that is an automatic protein prepared tool of LeDock. The binding pocket was generated from the cocrystallized ligand in the crystallographic structure using BIOVIA Discovery Studio Visualizer. The docking parameters were default sets as follows: binding pocket sets: X min = −41.092091, X max = −16.092091; Y min = −33.756273, Y max = −8.756273; Z min = −40.705364, Z max = −15.705364; number of binding poses sets 20; other parameters were default sets. The docking scores and docking poses were generated by LeDock. Before docking analysis, a compound set containing 577 active compounds and 313 inactive compounds downloaded from the BindingDB database was docked to PHGDH with the same docking schedule to assess the screening power. The sensitivity, specificity, F1 score, AUROC, hit rate, and enrichment factor were calculated to evaluate the screening power.

2.4. Molecular Dynamics Simulations

A Gromacs 2021.4 with NVIDIA GeForce GTX 3070 graphics processing unit acceleration was used to perform a 200 ns molecular dynamics simulations on compound-PHGDH docked complexes and apo PHGDH to explore the stability of docked complexes. The CHARMM36 force field was selected as the protein force field, while the CHARMM general force field obtained from the CHARMM General force field (CGenFF) server was used as the small-molecule force field. The TIP3P water model was used to add solvents to the each compound-PHGDH complex or apo PHGDH to establish a dodecahedron water box with a periodic boundary of 1.0 nm, and sodium ions and chloride ions were added to neutralize the charge in the system. After that, each system was subjected to 5000 steps of energy minimization using the steepest descent method. After minimization, under the NVT ensemble and NPT ensemble, two-step equilibrium procedures were performed for 1 ns, respectively. Each system is controlled to the temperature at 310 K and the pressure at 1 atm through the V-rescale thermostat and Parrinello–Rahman method, respectively. The van der Waals interactions were treated by a cutoff of 10 Å, while the long-range electrostatic interaction was treated by the particle-mesh Ewald (PME) method. Finally, a 200 ns production simulation was carried out under the NPT ensemble, and the time step was set to 2 fs. The simulation trajectory was saved at 10 ps. The simulation trajectories were analyzed through calculation of root-mean-square deviation (RMSD), radius of gyration (Rg), solvent accessible surface area (SASA), and root-mean-square fluctuation (RMSF) using the gmx rms, gmx gyrate, gmx sasa, and gmx rmsf module in GROMACS to evaluate the stability of the docked complex system. The hydrogen bond interaction between PHGDH and compound was estimated based on simulation trajectories using the gmx hbond module of GROMACS. Principal component analysis was carried out using the gmx covar and gmx anaeig module in GROMACS. Free energy landscape (FEL) was obtained by projecting the first two principal components using the gmx sham module in GROMACS. The dynamic cross-correlation matrix analysis was performed using the bio3D R package.

2.5. Binding Free Energy Calculation

The Molecular Mechanics Poisson–Boltzmann Surface Area (MM-PBSA) method was used to calculate the binding free energy of the PHGDH-compound docked complex. The binding free energy is calculated according to the following formula:

$$\mathrm{\Delta }{G}_{\mathrm{bind}}=G_{\mathrm{complex}}-(G_{\mathrm{receptor}}+G_{\mathrm{ligand}})$$

G _complex, G _ligand, and G _protein represent the free energies of the complex, ligand, and receptor. The binding free energy calculation was accomplished by teh gmx_MMPBSA tool.

3. Results and Discussion

3.1. PHGDH Identified as a Therapeutic Target for Enhancing Antiviral Innate Immunity through Mendelian Randomization

Mendelian randomization (MR) is a statistical tool that uses SNPs as instrumental variables (IVs) to infer the causal relationship between exposure and outcome. By simulating randomized controlled trials, MR can effectively infer the association of drug target exposure on clinical outcomes. − In this study, three human serine biosynthesis related protein-coding genes (PSAT1, PHGDH, and PSPH) were selected as exposure (Figure A), and their associated SNPs were selected as instrumental variables from the eQTLGen consortium. The IFN-β level was used as the outcome to proxy antiviral innate immunity, and the summary-level GWAS data for the IFN-β level were extracted from the NHGRI-EBI Catalog of GWAS. The two-sample MR analysis was performed to estimate the causal effect of expression levels of the genes on the IFN-β level (Figure B).

1.

Identification of the therapeutic target for enhancing antiviral innate immunity through Mendelian randomization. (A) Human serine biosynthesis pathway. (B) Overview of the MR study design. (i) Instrumental variables are significantly associated with the exposure. (ii) Instrumental variables are not associated with confounders. (iii) Instrumental variables are not directly associated with the outcome. (C) Forest plots displaying MR results of human serine biosynthesis related protein-coding genes with IFN-β level from the discovery cohort. (D) Sensitivity analysis result of PHGDH-IFN-β association. (E) The result of leave-one-out analysis. (F) Forest plots displaying MR results of PHGDH expression with IFN-β level from the replication cohort.

For MR analysis, one out of three serine biosynthesis related protein-coding genes passed the IVW-MR test (Figure C, Table S2). The IVW-MR analysis showed that the increase in gene expression of PHGDH was associated with a lower level of IFN-β (OR = 0.862, 95% CI = 0.786–0.955, p = 0.004), indicating that PHGDH inhibition might elevate IFN-β level.

Horizontal pleiotropy occurs when some of the genetic variations used as instrumental variables have a direct effect on the outcome rather than being mediated by exposure, so we evaluated the presence of horizontal pleiotropy using the MR-Egger test and MR-PRESSO test. We also used Cochran’s Q test with the IVW method and MR-Egger method to evaluate the heterogeneity of causal estimates between genetic variants. If the p value is less than 0.05, we will consider the existence of horizontal pleiotropic or heterogeneity. The MR-Egger intercept test and MR-PRESSO test displayed that there was no presence of horizontal pleiotropy (p > 0.05), and Cochran’s Q test displayed that heterogeneity was nonexistent between genetic variants (p > 0.05) (Figure D, Table S3). Furthermore, the Steiger directionality test was performed to confirm the causal direction between PHGDH expression and IFN-β level. The Steiger directionality test showed that the IFN-β level had no reverse causal effect on the expression of PHGDH (p < 0.05) (Figure D). In addition, to evaluate the impact of each genetic variant on the MR analysis separately, we further performed a one-by-one exclusion test using leave-one-out analysis (Figure E). The leave-one-out analysis showed that all SNPs are consistent (the same on the left side of 0), indicating that the MR results were not significantly different from the single SNP. The results of the above sensitivity analysis fully showed that MR analysis is reliable and stable.

3.2. PHGDH Remains Significant in the Replication IFN-β Cohort

To evaluate the causal association of PHGDH expression on the IFN-β level resulting from MR analysis based on the discovery IFN-β cohort, the same MR analysis was conducted based on the replication IFN-β cohort. The causal associations between gene expression of PHGDH and IFN-β level were also observed in the replication IFN-β cohort (Figure F, Table S4). The IVW-MR analysis showed that the increased PHGDH expression was associated with reduced IFN-β level (OR = 0.939, 95% CI = 0.887–0.994, p = 0.029) in the replication IFN-β cohort. Then, we performed sensitivity analysis to ensure the robustness and reliability of the MR analysis. The MR-Egger intercept test and MR-PRESSO test showed that there was no presence of horizontal pleiotropy (p > 0.05), and Cochran’s Q test showed that heterogeneity was not present between genetic variants (p > 0.05) (Table S5). Overall, the PHGDH expression negatively associated with IFN-β level resulting from MR analysis in the replication IFN-β cohort was consistent with MR results in the discovery IFN-β cohort.

Research has demonstrated that the knockdown of PHGDH at the genetic level, or the inhibition of its activity, can enhance IFN-β-mediated antiviral innate immunity and protect mice from viral infections. Our MR analysis provides suggestive evidence that inhibiting PHGDH increases IFN-β levels, thereby augmenting antiviral innate immunity. These findings support PHGDH as a promising therapeutic target for enhancing antiviral innate immunity to combat viral infections.

3.3. Machine Learning Model Training and Predictions

Twenty predictive machine learning models were developed using five algorithmsglmboost, SVM, XGBoost, NB, and DNNcombined with four molecular representations: Mol2Vec fingerprint, MACCS Keys fingerprint, Morgan fingerprint, and RDKit Descriptor. These models were trained on a data set comprising 404 active compounds and 220 inactive compounds targeting PHGDH. During model development, the performance of each model was evaluated using leave-one-out cross-validation (LOOCV), and the results are presented in Figures and . The models were further validated on an independent test set containing 173 active and 93 inactive compounds, with performance metrics also displayed in Figures and .

2.

The performance of different machine learning models was tested by LOOCV in the training set and was validated in the test set.

3.

ROC curve of (A) Support Vector Machine model based on Mol2Vec fingerprints (svm_Mol2Vec), (B) Support Vector Machine model based on RDKit descriptors (svm_RDKit), (C) Support Vector Machine model based on Morgan Fingerprints (svm_Morgan), (D) Support Vector Machine model based on MACCS Keys (svm_MACCS), (E) Gradient Boosting with Component-wise Linear model based on Mol2Vec fingerprints (glmboost_Mol2Vec), (F) Gradient Boosting with Component-wise Linear model based on RDKit descriptors (glmboost_RDKit), (G) Gradient Boosting with Component-wise Linear model based on Morgan Fingerprints (glmboost_Morgan), (H) Gradient Boosting with Component-wise Linear model based on MACCS Keys (glmboost_MACCS), (I) Extreme Gradient Boosting model based on Mol2Vec fingerprints (XGBoost_Mol2Vec) (J) Extreme Gradient Boosting model based on RDKit descriptors (XGBoost_RDKit), (K) Extreme Gradient Boosting model based on Morgan Fingerprints (XGBoost_Morgan), (L) Extreme Gradient Boosting model based on MACCS Keys (XGBoost_MACCS), (M) Deep Neutral Networks model based on Mol2Vec fingerprints (DNN_Mol2Vec), (N) Deep Neutral Networks model based on RDKit descriptors (DNN_RDKit), (O) Deep Neutral Networks model based on Morgan Fingerprints (DNN_Morgan), and (P) Deep Neutral Networks model based on MACCS Keys (DNN_MACCS).

Among the 20 models, the DNN model based on RDKit descriptors demonstrated the highest predictive performance, achieving a sensitivity of 96.3%, specificity of 98.2%, F1 score of 0.976, accuracy of 97.2%, and AUROC of 0.997 on the training set. This superior performance was consistently observed in the test set, where the model achieved a sensitivity of 86.1%, specificity of 82.8%, F1 score of 0.895, accuracy of 84.5%, and AUROC of 0.93. Furthermore, the model exhibited strong screening power, with a hit rate of 100% at 0.5%, 1%, and 2% enrichment levels and enrichment factors (EFs) of 1.544 (EF0.5%, EF1%, EF2%) in the training set. In the test set, it maintained a high screening efficiency, yielding a hit rate of 100% across all three enrichment levels and EF values of 1.537 for EF0.5%, EF1%, and EF2%.

Given its outstanding predictive capability and robust generalization performance, the RDKit Descriptor-based DNN model was selected as the final model for predicting the PHGDH inhibitory activity of natural compounds derived from Pleurotus ostreatus. A total of 4474 natural compounds sourced from Pleurotus ostreatus were retrieved from The Food Database (FOODB). Using the trained DNN model, the potential inhibitory activities against PHGDH were predicted. Of these, approximately 1176 compounds were classified as active, and the complete prediction results are provided in Table S6.

3.4. Molecular Docking

Molecular docking analysis was performed between 1176 natural compounds derived from Pleurotus ostreatus, identified through machine-learning-based virtual screening, and PHGDH. Prior to docking, the screening performance of the docking protocol was evaluated using enrichment metrics. The results demonstrated strong discriminatory power, with a hit rate of 100% at 0.5%, 87.5% at 1%, and 86.7% at 2%, along with enrichment factors (EFs) of 1.735 (EF0.5%), 1.518 (EF1%), and 1.504 (EF2%). Based on these results, the top 0.5% of compounds were selected as potential PHGDH inhibitors. The chemical structures and docking scores of these top-ranking compoundsFDB024162, FDB095733, FDB024165, FDB095731, and FDB095735are summarized in Table . Their interactions with key residues of PHGDH were further analyzed and visualized (Table and Figure ). FDB024162 formed hydrogen bonds with GLY154, ASP175, THR207, SER212, THR214, and ASN218, as well as hydrophobic interactions with ARG155, ILE156, PRO176, ILE177, and PRO208. In the complex with FDB024165, PHGDH exhibited hydrogen bonding with GLY154, HIS206, and ALA235; hydrophobic interactions with ILE156, LEU193, LEU210, PRO211, LEU216, and ALA235; and an electrostatic interaction involving ARG236. FDB095735 established hydrogen bonds with GLY154, ASP175, PRO176, HIS206, PRO208, CYS234, and ARG236; electrostatic interactions with ASP175; and hydrophobic contacts with ILE156, PRO176, LEU210, PRO211, and ALA235. FDB095731 formed hydrogen bonds with GLY152, GLY154, HIS206, THR207, SER212, and ARG236 and hydrophobic interactions with LEU193, PRO208, LEU210, PRO211, and LEU216. FDB095733 engaged in hydrophobic interactions with ILE156, PRO176, ILE177, and PRO208; hydrogen bonds with TRP197 and SER212; and an amide-pi-stacked interaction with LEU193 and GLU194. The cofactor NAD⁺ binds to PHGDH via a cofactor-binding pocket composed of three subsites: the adenine-binding subsite (residues THR78, ALA106, CYS234, ALA235, ARG236, HIS206, THR207, PRO208, and ASP260), the nicotinamide-binding subsite (TYR174, ASP175, PRO176, THR207, PRO208, and SER212), and the phosphate linker-binding subsite (GLY154–ILE156). , Docking poses revealed that the selected compounds bind deeply within this cofactor-binding pocket. Specifically, FDB024162, FDB029254, FDB095733, and FDB095731 stably occupy the adenine, nicotinamide, and phosphate linker subsites, whereas FDB095735 primarily interacts with the nicotinamide and adenine subsites and FDB024165 binds predominantly to the nicotinamide and phosphate linker subsites. Collectively, the molecular docking results indicate that these natural compounds tightly interact with critical residues in the NAD⁺-binding site of PHGDH through hydrogen bonding, hydrophobic, and electrostatic interactions, suggesting their potential to disrupt the binding of NAD⁺ to PHGDH and thereby inhibit its enzymatic activity.

1. Structures and Docking Scores of the Selected Compounds.

2. List of Bonding Interactions between the Selected Compounds with PHGDH.

compound	residue	distance	category	type
FDB024162	GLY154	2.99818	hydrogen bond	conventional hydrogen bond
	ASN218	2.39067	hydrogen bond	conventional hydrogen bond
	ASN218	2.56924	hydrogen bond	conventional hydrogen bond
	SER212	2.14126	hydrogen bond	conventional hydrogen bond
	THR214	1.99603	hydrogen bond	conventional hydrogen bond
	THR207	2.61723	hydrogen bond	carbon hydrogen bond
	ASP175	2.49529	hydrogen bond	carbon hydrogen bond
	SER212	2.5742	hydrogen bond	carbon hydrogen bond
	SER212	1.94917	hydrogen bond	carbon hydrogen bond
	ARG155	4.2324	hydrophobic	alkyl
	PRO176	5.35573	hydrophobic	alkyl
	PRO176	5.08509	hydrophobic	alkyl
	PRO208	4.50432	hydrophobic	alkyl
	ILE156	4.61398	hydrophobic	alkyl
	ILE177	5.21151	hydrophobic	alkyl
FDB024165	HSD206	1.81152	hydrogen bond	conventional hydrogen bond
	ALA235	2.22486	hydrogen bond	conventional hydrogen bond
	GLY154	2.76105	hydrogen bond	carbon hydrogen bond
	HSD206	2.3537	hydrogen bond	carbon hydrogen bond
	ARG236	4.10266	electrostatic	pi-cation
	PRO211	4.53921	hydrophobic	alkyl
	PRO211	5.15571	hydrophobic	alkyl
	LEU216	5.31837	hydrophobic	alkyl
	LEU193	4.34498	hydrophobic	alkyl
	LEU210	5.12206	hydrophobic	alkyl
	ILE156	4.673	hydrophobic	pi-alkyl
	ALA235	5.00229	hydrophobic	pi-alkyl
FDB029254	GLY154	2.10388	hydrogen bond	conventional hydrogen bond
	ARG236	2.83703	hydrogen bond	conventional hydrogen bond
	ASP175	1.7844	hydrogen bond	conventional hydrogen bond
	HSD206	1.83685	hydrogen bond	conventional hydrogen bond
	THR207	2.09734	hydrogen bond	conventional hydrogen bond
	PRO176	2.32361	hydrogen bond	carbon hydrogen bond
	HSD206	2.51825	hydrogen bond	carbon hydrogen bond
	THR207	2.27351	hydrogen bond	carbon hydrogen bond
	PRO208	2.53716	hydrogen bond	carbon hydrogen bond
	SER212	2.71156	hydrogen bond	carbon hydrogen bond
	LEU193	4.37931	hydrophobic	alkyl
	PRO176	4.4417	hydrophobic	alkyl
	TYR174	4.9013	hydrophobic	pi-alkyl
	TRP197	4.95956	hydrophobic	pi-alkyl
FDB095733	TRP197	1.66701	hydrogen bond	conventional hydrogen bond
	SER212	2.81955	hydrogen bond	carbon hydrogen bond
	LEU193	4.81431	hydrophobic	amide-Pi stacked
	PRO176	4.55735	hydrophobic	alkyl
	PRO208	4.75257	hydrophobic	alkyl
	PRO208	4.14428	hydrophobic	alkyl
	ILE177	5.3606	hydrophobic	alkyl
	ILE156	5.16247	hydrophobic	alkyl
FDB095735	LEU193	1.89765	hydrogen bond	conventional hydrogen bond
	GLY152	2.42822	hydrogen bond	carbon hydrogen bond
	PRO176	2.53621	hydrogen bond	carbon hydrogen bond
	PRO192	3.04942	hydrogen bond	carbon hydrogen bond
	PRO208	2.802	hydrogen bond	carbon hydrogen bond
	TYR174	2.15616	hydrogen bond	carbon hydrogen bond
	PRO208	4.23119	hydrophobic	alkyl
	PRO208	5.18339	hydrophobic	alkyl
	ILE177	4.66645	hydrophobic	alkyl
	ILE177	4.84617	hydrophobic	alkyl
FDB095731	GLY154	2.24914	hydrogen bond	conventional hydrogen bond
	SER212	2.3585	hydrogen bond	conventional hydrogen bond
	ARG236	2.7534	hydrogen bond	conventional hydrogen bond
	HSD206	1.86491	hydrogen bond	conventional hydrogen bond
	GLY152	2.75124	hydrogen bond	carbon hydrogen bond
	GLY154	2.57179	hydrogen bond	carbon hydrogen bond
	THR207	2.88845	hydrogen bond	carbon hydrogen bond
	PRO211	4.60534	hydrophobic	alkyl
	PRO211	5.08814	hydrophobic	alkyl
	LEU210	4.84654	hydrophobic	alkyl
	PRO208	5.29053	hydrophobic	alkyl
	LEU193	5.47598	hydrophobic	alkyl
	LEU216	4.74301	hydrophobic	alkyl

4.

Three-dimensional interaction and two-dimensional interaction profile of the selected natural compound docked with PHGDH. Hydrogen bonds are displayed as dark green ball and stick, attractive charges are displayed as gold ball and stick, hydrophobic bonds are displayed as pink ball and stick, and carbon–hydrogen bonds are displayed as light green ball and stick.

3.5. Molecular Dynamics Simulations

A 200 ns molecular dynamics (MD) simulation was performed to evaluate the structural stability of the docked complexes and to investigate the dynamic conformational changes of PHGDH upon binding with selected natural compounds.

3.5.1. Root-Mean-Square Deviation (RMSD)

To assess the structural stability of PHGDH in complex with the selected natural compounds, root-mean-square deviation (RMSD) values of the Cα atoms were calculated relative to the initial frame of the simulated trajectory. As shown in Figure A, the RMSD increased steadily during the initial phase and rapidly converged, indicating that all systems reached equilibrium within the simulation time. The average RMSD values for PHGDH bound to FDB024162, FDB024165, FDB095731, FDB095733, FDB095735, and apo PHGDH were 0.368, 0.355, 0.426, 0.295, 0.296, and 0.374 nm, respectively. Binding of the selected natural compounds did not significantly alter the overall flexibility of PHGDH. Specifically, FDB024162, FDB024165, FDB095733, and FDB095735 induced a slight reduction in mean protein RMSD, whereas FDB095731 resulted in a modest increase.

5.

Conformational stability analysis of the selected compound-PHGDH docked complexes: (A) RMSD of the Cα atom, (B) Rg profile, and (C) SASA profile of the protein of selected compound-PHGDH docked complexes and apo PHGDH.

3.5.2. Radius of Gyration (Rg)

The radius of gyration (Rg) was computed throughout the simulation to evaluate the compactness of PHGDH in the presence and absence of ligands. Figure B illustrates the fluctuation patterns of Rg for both bound and unbound forms of PHGDH over the 200 ns MD trajectory. The average Rg values for the FDB024162-, FDB024165-, FDB095731-, FDB095733-, and FDB095735-bound PHGDH complexes were 1.611 1.595, 1.610, 1.588, and 1.600 nm, respectively, compared to 1.515 nm for apo PHGDH. No significant differences in Rg were observed between the ligand-bound and apo forms, suggesting that compound binding does not substantially affect the overall compactness or global conformation of PHGDH. These results confirm that the PHGDH-compound complexes maintain structural integrity and compact folding throughout the simulation.

3.5.3. Solvent Accessible Surface Area (SASA)

The solvent accessible surface area (SASA) is a key indicator of protein surface exposure to solvent and can reflect conformational changes. Therefore, SASA values were calculated to assess potential alterations in the hydrophobicity and surface accessibility of PHGDH upon binding to the screened natural compounds. As depicted in Figure C, SASA remained stable throughout the 200 ns simulation for both bound and unbound systems. The average SASA values for PHGDH in complex with FDB024162, FDB024165, FDB095731, FDB095733, FDB095735, and apo PHGDH were 94.95 nm², 91.46 nm², 93.53 nm², 93.00 nm², 92.16 nm², and 113.4 nm², respectively. The minimal variation in SASA across conditions indicates that the selected compounds do not induce significant conformational perturbations in PHGDH.

3.5.4. Root-Mean-Square Fluctuation (RMSF)

Root-mean-square fluctuation (RMSF) provides residue-level insight into local flexibility during MD simulations. Thus, RMSF values were calculated for each amino acid residue of PHGDH in both the ligand-bound and apo states to evaluate the impact of compound binding on local dynamics. As shown in Figure , the five compound-PHGDH complexes exhibit highly similar RMSF profiles to each other and to apo PHGDH, with consistent fluctuation patterns across residues. The average RMSF values for the FDB024162-, FDB024165-, FDB095731-, FDB095733-, and FDB095735-bound complexes were 0.209 0.214, 0.202, 0.223, and 0.183 nm, respectively, compared to 0.222 nm for apo PHGDH. Overall, the differences in residue fluctuations among the complexes are negligible. Notably, residues at the binding site display low RMSF values, suggesting stabilized interactions upon compound binding and favorable binding stability. This implies that these compounds stabilize the protein without inducing major conformational changes. However, in comparison to the apo form, the VAL261–PRO266 region exhibits increased fluctuation upon binding FDB095733, indicating reduced local stability in this segment when bound to this particular compound, whereas other compounds confer greater rigidity to this region.

6.

RMSF profile of the protein of selected compound-PHGDH docked complexes and apo PHGDH.

3.6. Hydrogen Bond Analysis

To evaluate major role of hydrogen bonds contributed to keeping the binding stability of protein–ligand complexes, we analyzed dynamic hydrogen bonding statistics of compound-PHGDH complexes for the entire simulation.The number of hydrogen bonds formed between PHGDH and the selected natural compounds throughout whole simulation time are shown in Figure . The median number of hydrogen bonds formed is 3, 2, 2, 1, and 2 for FDB024162, FDB024165, FDB095731, FDB095733, and FDB095735-PHGDH complexes throughout the 200 ns MD simulation, respectively. As shown in Figure , hydrogen bonds formed between PHGDH and the selected compounds were consistent over entire MD simulation times, suggesting that hydrogen bonds are conducive to the formation of a stable complex.

7.

Hydrogen bond analysis between the selected compounds and PHGDH during MD simulation. The hydrogen bond number varies with simulation time.

Next, we calculated the occupancy of residues participating in the formation of a hydrogen bond with a compound in the five complexes to explore key residues participating in hydrogen bonding as presented in Table S7. The top 10 hydrogen bonds in terms of occupancy were visualized as hydrogen bond existence maps (Figure ). In the case of the FDB024162-PHGDH complex, Asp175 and Tyr174 residues were involved in hydrogen bonding interacting with different atoms of FDB024162, and these hydrogen bonds were continuously seen over the whole simulation times (Figure A). Similarly, hydrogen bonds formed between the LEU209 residue of PHGDH and different atoms of FDB024165 were seen constantly over most of the simulation times (Figure B). For the FDB095731-PHGDH complex, hydrogen bonds formed between the Asp175 residue of PHGDH and different atoms of FDB095731 were constantly observed during 0–90 ns simulation times (Figure C). Contrary to the FDB095731-PHGDH complex, ASN218, TRP197, and GLU194 residues of PHGDH were constantly observed forming hydrogen bonds interacting with different atoms of FDB095735 during 100–200 ns simulation times (Figure E.) For the FDB095733-PHGDH complex, fleeting and less constant hydrogen bonds formed between SER212, TRY174, and TRP197 residues of PHGDH and FDB095733 were seen over simulation times (Figure D). Overall, these results revealed that hydrogen bonding has a definite contribution to the stable binding of the selected compounds and PHGDH.

8.

Hydrogen bond existence map for ligands (A) FDB024162, (B) FDB024165, (C) FDB095731, (D) FDB095733, and (E) FDB095735 over MD simulation.

3.7. Secondary Structure Analysis

The dictionary of secondary structure of proteins (DSSP) method was employed to investigate the secondary structural changes of PHGDH induced by binding of the selected compounds during the MD simulation. Binding of the compounds led to interconversion among different secondary structural elements in specific residue regions, resulting in localized fluctuations throughout the simulation trajectory (Figure ). As shown in Figure , compound binding induced transitions from α-helix to turn in the ALA118–ARG119 region; from coil to bend in the PRO282, HSD283, GLY285, THR288, and ALA291 region; and from bend to coil in the SER280, CYS281, LEU284, and ASP260 region. These structurally perturbed regions are located distal to the ligand binding site, suggesting that such changes do not interfere with the ligand binding. Overall, the secondary structure of PHGDH remains well-preserved, and all compound-bound complexes maintain a stable and compact conformation.

9.

Color maps to the secondary structure changes of each region of the PHGDH protein during the MD simulation time for binding of (A) FDB024162, (B) FDB024165, (C) FDB095731, (D) FDB095733, (E) FDB095735, and (F) apo protein.

3.8. Principal Component Analysis (PCA) and Free Energy Landscape (FEL) Analysis

To identify the dominant motion modes contributing to the overall dynamics of PHGDH in complex with the selected natural compounds, principal component analysis (PCA) was performed on the entire MD simulation trajectories of the compound-PHGDH complexes. First, a covariance matrix was constructed based on the Cα atomic positions of PHGDH in both ligand-bound and apo states. This matrix was then diagonalized to obtain eigenvectors and their corresponding eigenvalues. The trace values of the covariance matrix for PHGDH bound to FDB024162, FDB024165, FDB095731, FDB095733, FDB095735, and the apo protein were 15.037 21.557, 14.325, 20.739, 7.396, and 17.419 nm, respectively, indicating reduced global flexibility upon bindingexcept for FDB024165 and FDB095733, which exhibited higher dynamical behavior. The cumulative contribution of the first three principal components to the essential dynamics decreased upon compound binding, as illustrated in Figure . Specifically, the first three eigenvectors accounted for 69.2%, 81.8%, 74.0%, 76.3%, and 54.4% of the total variance in the FDB024162-, FDB024165-, FDB095731-, FDB095733-, and FDB095735-bound complexes, respectively, compared to 80.9% in apo PHGDH. Similarly, the first eigenvector alone explained 36.1%, 56.6%, 51.2%, 31.9%, and 30.3% of the variance in the respective complexes versus 58.3% in the apo form. These results indicate that compound binding reduces the dominance of the leading principal components, implying a restriction in large-scale collective motions and a decrease in the overall conformational fluctuation. This suggests that ligand occupancy at the active site may suppress the disordered movements of PHGDH, thereby enhancing system stability.

10.

Percentage of variance of eigenvectors graphs of PCA analysis for five compound-PHGDH complexes and apo PHGDH.

Subsequently, projections of the first two principal components (PC1 and PC2) onto the essential subspace were visualized as scatter plots to characterize dominant conformational motions, where each point represents a distinct protein conformation and color gradients reflect progression over simulation time (Figure ). The area coverage in the essential subspace (PC1 vs PC2) for PHGDH bound to FDB024162, FDB024165, FDB095731, FDB095733, FDB095735, and apo PHGDH was calculated as 341.8447 nm², 389.6816 nm², 365.2637 nm², 410.4165 nm², 217.8938 nm², and 394.8029 nm², respectively. Except for the FDB095733-bound complex, all other ligand-bound systems exhibited reduced conformational sampling, indicating that compound binding restricts the conformational freedom of PHGDH. This narrowing of accessible states reflects enhanced thermodynamic stability and favors the formation of tightly packed stable complexes. Collectively, these findings demonstrate that PHGDH adopts more constrained and stable conformations upon binding to the selected compounds.

11.

Essential subspace projection graphs of the first two principal components of PCA analysis for five compound-PHGDH complexes and apo PHGDH.

To visualize the energy minima and conformational landscape of PHGDH and its compound-bound complexes, we constructed free energy landscapes (FELs) by using three dimensions: PC1, PC2, and Gibbs free energy (where red indicates high energy and blue indicates low energy), presented in 3D representations (Figure ). As shown in Figure , PHGDH bound to the selected compounds sampled a broad conformational space, exhibited multiple small local minima, and displayed a reversed population distribution compared to apo PHGDH. This suggests that these compounds induced PHGDH to adopt diverse, flexible conformational ensembles during the 200 ns simulation. Specifically, PHGDH bound to FDB095731 and FDB095735 explored an extensive conformational space centered around a single global energy minimum, indicating minimal rearrangement of the overall conformation. In contrast, for PHGDH complexes with FDB024162, FDB024165, and FDB095735, more distinct energy basins emerged, which were closely spaced and widely distributed, reflecting the presence of multiple stable conformations. This implies that these compounds can accommodate various protein conformations and form stable interactions with PHGDH. Overall, compared with the free energy surface of apo PHGDH, the five compound-bound complexes exhibited significant population shifts, suggesting transitions into multiple metastable states. These findings indicate that FDB024162, FDB024165, and FDB095735 effectively modulate the conformational dynamics of PHGDH, potentially influencing its functional activity. Finally, the lowest-energy conformations of each complex were extracted (Figure ). At the energy minimum, multiple distinct structural states were observed. Superimposition of these conformations reveals the dynamic behavior of the complexes. As shown in Figure , despite differences, common features are evident across complexes, such as limited movement of the protein backbone relative to the ligand, stable positioning of the compound within the cofactor-binding pocket, and only minor conformational changes in the binding site.

12.

Free energy landscape for the (A) FDB024162-PHGDH complex, (B) FDB024165-PHGDH complex, (C) FDB095731-PHGDH complex, (D) FDB095733-PHGDH complex, (E) FDB095735-PHGDH complex, and (F) apo PHGDH. (1) Three-dimensional representation of free energy landscape. (2) The conformations of each complex that represent the lowest energy are displayed in the third column.

3.9. Dynamic Cross-Correlation Matrix (DCCM) Analysis

We then performed dynamic cross-correlation matrix (DCCM) analysis to investigate correlated motions within the PHGDH-compound complexes. In this representation, strongly positive correlations (dark blue regions) indicate residues moving in the same direction, while negative correlations (yellow regions) reflect anticorrelated motions (residues moving in opposite directions) (Figure ). As illustrated in Figure , among the five complexes, PHGDH bound to FDB024165 and FDB095731 showed increased positively correlated movements followed by those of FDB024162. These enhancements suggest enhanced intermolecular coordination and improved complex stability. In contrast, PHGDH complexes with FDB095735 and FDB095733 exhibited reduced positive correlations, implying weakened residue–residue contacts and relatively lower structural coherence. Furthermore, anticorrelated motions decreased significantly upon compound binding across all systems, indicating a general suppression of opposing domain movements. Collectively, these results demonstrate that the screened natural compounds promote a more coordinated and stable dynamic profile in PHGDH.

13.

Dynamic cross-correlation matrix (DCCM) plots for the (A) FDB024162-PHGDH complex, (B) FDB024165-PHGDH complex, (C) FDB095731-PHGDH complex, (D) FDB095733-PHGDH complex, (E) FDB095735-PHGDH complex, and (F) apo PHGDH. (1) The high positive region (dark blue) indicates the strong positively correlated movements, while the negative region (yellow) indicates the strong anticorrelated movements. (2) Three-dimensional representation of correlated motions in red and anticorrelated motions in blue are presented.

3.10. Binding Free Energy

The binding free energy of each compound-PHGDH complex was calculated using the MM/PBSA method to evaluate the binding affinity of the selected compounds toward PHGDH. The total binding free energy (ΔG _total) comprises contributions from van der Waals interactions (ΔE _vdW), electrostatic interactions (ΔE _ele), polar solvation energy (ΔG _polar), and nonpolar solvation energy (ΔG _nonpolar). As shown in Table , the calculated ΔG _total values for the FDB024162-, FDB024165-, FDB095731-, FDB095733-, and FDB095735-PHGDH complexes were −9.843, −14.298, −12.163, −11.862, and −9.715 kcal/mol, respectively. Variations in binding free energies among the five complexes arise from differences in the relative contributions of ΔE _vdW, ΔE _ele, and ΔG _polar. Notably, the van der Waals interaction exhibited the largest contribution in the FDB024162-PHGDH complex (−58.86 kcal/mol), while the electrostatic interaction contributed most significantly to the FDB095731-PHGDH complex (−38.647 kcal/mol). Across all complexes, the magnitudes of van der Waals contributions exceeded those of electrostatic and polar solvation energies, indicating that van der Waals forces play a dominant role in stabilizing the compound-PHGDH interactions. The FDB095731-PHGDH complex displayed the weakest electrostatic interaction compared with the others, which may explain its lower median number of hydrogen bonds observed in the prior hydrogen bond analysis.

3. Binding Free Energy Analysis.

energy (kcal/mol)	FDB024162	FDB024165	FDB095731	FDB095733	FDB095735
ΔG total	–9.843	–14.298	–12.163	–11.863	–9.715
ΔGnonpolar	52.781	33.987	59.572	23.358	33.841
ΔG polar	–8.319	–8.253	–7.919	–7.120	–7.019
ΔE ele	–22.220	–10.965	–38.647	–6.050	–13.363
ΔE vdW	–58.861	–52.007	–57.262	–46.557	–47.501
TΔdS	–26.777	–22.939	–32.094	–24.506	–24.328

To identify key residues contributing significantly to the binding of the selected compounds with PHGDH, residue-level decomposition of the binding free energy was performed. The energetic contribution of individual residues in each of the five complexes is illustrated in Figure . For FDB024162, major contributors included ILE177, PRO176, PRO208, LEU210, LEU216, LEU193, TYR174, LEU151, THR213, GLY152, and THR207. In the FDB024165 complex, the most influential residues were PRO208, ILE177, LEU210, PRO176, ARG155, ARG236, ILE156, LEU216, and THR213. For FDB095731, key residues included ILE177, PRO208, ARG155, PRO176, THR207, LEU210, THR213, LEU153, ILE156, and LEU216. In the FDB095733 complex, significant contributions came from PRO176, LEU210, LEU193, ILE177, PRO211, PRO208, LEU216, and THR213. For FDB095735, the primary contributors were LEU193, LEU216, PRO176, THR213, ILE177, and LEU210. It is evident that several key residuessuch as ILE177, PRO176, PRO208, LEU210, THR213, and LEU216are consistently involved across multiple complexes, suggesting their critical roles in ligand binding. Most of these shared residues are located within the adenine-binding and nicotinamide-binding subsites of the PHGDH cofactor-binding pocket, implying that interactions with PRO176 and PRO208 at this site are particularly important for compound binding. Meanwhile, residues ARG155, GLY154, and ILE156, situated in the phosphate linker-binding subsite, contributed substantially only in the FDB095731 and FDB024165 complexes, which may account for their relatively stronger overall binding free energies compared to those of the other complexes.

14.

Energy contribution in each residue of selected compound-PHGDH complexes.

4. Conclusions

Animal studies have demonstrated that inhibiting serine biosynthesis may serve as a promising therapeutic strategy for antiviral infections by enhancing host antiviral innate immunity, thereby offering novel targets for strengthening innate immune responses to viral challenges. In this study, we identified five glycerophospholipid compounds derived from Pleurotus ostreatus that inhibit PHGDHa key enzyme in the serine biosynthesis pathwayusing machine learning models and molecular-docking-based virtual screening, with the aim of enhancing antiviral innate immunity. We first selected PHGDH from protein-coding genes associated with serine biosynthesis as a potential therapeutic target for boosting antiviral innate immunity through Mendelian randomization (MR). The MR analysis yielded suggestive evidence that PHGDH inhibition is associated with elevated IFN-β levels, supporting its role in enhancing antiviral innate immunity and highlighting its potential as a therapeutic target against viral infections. Subsequently, we developed a deep neural network model based on RDKit descriptors to predict the inhibitory activity of natural compounds from Pleurotus ostreatus. By integrating this model with molecular-docking-based virtual screening, we successfully identified five glycerophospholipid compounds (FDB024162, FDB024165, FDB095731, FDB095733, and FDB095735) with strong PHGDH inhibitory potential. Furthermore, molecular dynamics simulations and binding free energy calculations revealed that these compounds form stable complexes with PHGDH, primarily through interactions with specific residues within the NAD⁺ cofactor-binding pocket. These findings suggest that the compounds effectively inhibit PHGDH, thereby potentially enhancing the innate immunity against viral infections. Future studies should include experimental validation to confirm their functional roles in modulating antiviral innate immune responses.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c13179.

• Information of eQTL and GWAS summary data; MR association between genetic gene expression and IFN-β in the discovery cohort; heterogeneity and pleiotropy tests of instrument effects in the discovery cohort; MR association between genetic gene expression and IFN-β in the replication cohort; heterogeneity and pleiotropy tests of instrument effects in the replication cohort; activity of natural compounds originated from Pleurotus ostreatus predicted by the RDKit descriptors based DNN model; and occupancy of residues participated in the formation of hydrogen bond with the compound in the five complexes (XLSX)

Shaohua Xu: Conceptualization, Methodology, Investigation, Writing–original draft preparation, Writing–reviewing and editing, Project administration. Kai Yang: Methodology, Visualization, Investigation, Software, and Validation.

This work was supported by Wuwei Science and Technology Plan Project under Grant WW25A01RQNK004.

All summary-level data used in this MR analysis were obtained from public databases. The original studies received ethical approval and obtained individual consent.

The authors declare no competing financial interest.