Optimisation of peptides targeting reverse transcriptase HIV-1 using QSAR, machine learning, and computational approaches Authors: Fachrur Rizal Mahendra, Indira Prakoso, Alfa Marzelino, Muhamad Rizqy Fadhillah, Syukriyansyah, Muhammad Marsha Azzami Hasibuan, Juniza Firdha Suparningtyas, Mikael Kristiadi, Aprijal Ghiyas Setiawan, Faris Izzatur Rahman, Anissa Nofita Sari, Nauval Rajwaa Raysendria, Ilham Kurniawan, Wawaimuli Arozal, Kusmardi Kusmardi Published: 2025 ABSTRACT The emergence of drug resistance and adverse side effects associated with current HIV-1 reverse transcriptase (RT) inhibitors underscores the need for novel therapeutic strategies. Peptide-based drugs offer high specificity and lower toxicity, but their development is challenged by the vast combinatorial space of possible sequences and formulation issues. This study aims to identify potent tripeptide-based inhibitors targeting HIV-1 RT through an integrated computational pipeline combining machine learning, QSAR modeling, and in silico validation techniques. From 2,197 screened tripeptides, three candidates, namely FHW, HFW, and HHW, emerged with superior predicted affinity, stability, and drug-like properties. Among them, FHW exhibited the strongest interaction with HIV-1 RT, with a binding energy of −63.50 kcal/mol using MM/GBSA, outperforming the reference drug Nevirapine. FHW peptide also shows four same residues with Nevirapine, including Leu100, Val106, Tyr181, and Tyr188 by hydrophobic contacts. DFT analysis revealed favorable electronic properties, including a low HOMO-LUMO gap (4.73 eV) and high electrophilicity index (13.60). Based on these findings, the FHW peptide demonstrates the highest electrophilicity index among the four ligands, indicating superior electrophilic character relative to Nevirapine. PerMM simulations further indicated consistently negative energy profiles for FHW during membrane translocation and reflecting favorable interactions with the lipid environment. Molecular dynamics simulations confirmed the structural stability of the FHW–RT complex over a 100 ns trajectory. Collectively, these findings identify FHW as the most promising tripeptide-based RT inhibitor with potential for development into next-generation HIV therapeutics, with HFW and HHW as alternative candidates. Introduction Human Immunodeficiency Virus (HIV) continues to pose a significant global health burden more than 4 decades after it first emerged as a pandemic. According to the Joint United Nations Programme on HIV/AIDS (UNAIDS, 2024), an estimated 39.9 million people were living with HIV globally in 2023. Despite substantial progress in antiretroviral therapy (ART), approximately 1.3 million individuals acquired new HIV infections in the same year. Two major classes of antiretroviral drugs, nucleoside reverse transcriptase inhibitors (NRTIs) and non-nucleoside reverse transcriptase inhibitors (NNRTIs), constitute the backbone of HIV treatment. These drugs act by targeting the HIV reverse transcriptase (HIV-RT), an RNA- and DNA-dependent DNA polymerase responsible for the reverse transcription of the viral RNA genome into DNA. Inhibiting HIV-RT is a critical therapeutic strategy, as it prevents the virus from converting its RNA into DNA, thereby disrupting its ability to integrate into the host genome and replicate. However, the clinical use of NRTIs and NNRTIs is often limited by a range of adverse effects, including pruritus, fatigue, nausea, myalgia, neuropathy and lactic acidosis, as well as the emergence of drug-resistant viral strains (Margolis et al., 2014; Shi et al., 2016). A study by de los Rios et al. (2021) revealed that 43.6% of individuals on ART experienced side effects, with 65.8% reporting gastrointestinal symptoms such as stomach upset, pain, and diarrhea. These adverse effects not only affect quality of life but also contribute to non-adherence, with many patients missing one or more doses of ART in the past month to avoid discomfort. The failures of several first-generation and second-generation small molecule drug-based anti-HIV therapies in various stages of clinical trials are an indication that there is a need for a paradigm shift in the future designs of anti-HIV therapeutics. The limitations and failures of several first- and second-generation small-molecule anti-HIV therapies in clinical trials underscore the urgent need for a paradigm shift in the design of future HIV therapy. Peptide-based therapeutics have emerged as a promising alternative to small-molecule drugs which offer several notable advantages, including high specificity, potency, and reduced side effects (Fosgerau and Hoffmann, 2015). Composed of short amino acid sequences, peptides can interact with target proteins more effectively than small molecules due to their larger binding surfaces and the ability to form stronger and more specific interactions (Craik et al., 2013). This specificity contributes to a lower likelihood of off-target effects, reduced drug–drug interactions, and minimal toxicity, as peptides are typically metabolized into non-toxic amino acids (Olmez et al., 2012). Additionally, peptide therapeutics are generally less susceptible to the development of drug resistance. The high degree of structural specificity required for effective peptide binding means that viruses must undergo more substantial structural mutations to evade inhibition, making resistance less likely to emerge. Despite these advantages, peptide-based therapies have several limitations, such as poor in vivo stability and limited oral bioavailability, which complicate formulation and delivery (Choonara et al., 2014). Currently, two peptide-based drugs are approved for HIV treatment. The first is Enfuvirtide, a fusion inhibitor, which interferes with the conformational changes in the viral envelope protein gp41, thereby preventing membrane fusion. The second is Maraviroc, an entry inhibitor that blocks the interaction between gp120 and the CCR5 co-receptor, thereby preventing viral entry into host cells (Shi et al., 2016). However, to date, there are no approved peptide-based therapeutics targeting HIV-RT, highlighting a gap in therapeutic development and an opportunity for future research. The design of peptide-based drugs presents significant challenges, primarily due to the immense combinatorial space of potential peptide sequences. For instance, a simple tripeptide constructed from the 20 standard amino acids results in 8,000 unique combinations which needs experimental screening that both time-consuming and cost-prohibitive. This complexity underscores the need for efficient computational strategies to prioritize peptide candidates with the highest therapeutic potential. Computational approaches such as Quantitative Structure–Activity Relationship (QSAR) modeling and machine learning (ML) have become important to modern peptide drug discovery. These methods enable rapid in silico prediction of bioactivity, facilitating the selection of promising candidates from large peptide libraries and significantly accelerating the drug development pipeline. QSAR modeling is a ligand-based approach that predicts biological activity by correlating structural features with functional outcomes. Early QSAR studies relied on simple physicochemical descriptors (1D-QSAR). However, modern QSAR models incorporate multidimensional molecular representations, such as 2D, 3D, and 4D descriptors, and are often combined with machine learning algorithms to improve predictive accuracy (Mao et al., 2021). These advanced models can be employed to estimate quantitative outcomes such as pIC50 values, or to classify compounds based on activity, thus providing valuable insights for rational peptide design. This study presents potential candidates of novel peptide-based inhibitors toward HIV RT which offer an alternative to current antiretroviral drugs that often cause adverse side effects and contribute to drug resistance. By integrating ML techniques with QSAR modeling, we enhance the predictive accuracy of peptide-drug interactions while efficiently screening vast molecular libraries (Dogan and Durdagi, 2021; Baassi et al., 2023). Furthermore, in silico verification methods, including molecular docking simulations, were employed to evaluate the binding affinities and interaction patterns between the peptide candidates and the HIV RT active site. To refine the selection of top candidates, Molecular Mechanics-Generalized Born Surface Area (MM/GBSA) scoring was applied, providing a more robust assessment of binding free energies and stability. This multi-tiered computational approach not only accelerates the identification of promising drug candidates but also improves the reliability of virtual screening pipelines. The findings provide a foundation for future experimental validation with peptide candidates that can be evaluated in vitro and in vivo to assess their biological efficacy and pharmacological safety. Materials and methods This study employed a multi-stage machine learning approach to identify potent tripeptides with high activity, efficiency, and low toxicity. A collection of tripeptides was evaluated using a set of ligand data databases, including datasets for interaction fingerprints, chemical descriptors, IC50 prediction, and lipophilic efficiency. Docking with AutoDock Vina and AutoDock4, followed by interaction fingerprint generation using HIPPOS-PyLIF, was designed for Interaction-based modeling. Reduced feature dimensions before training models with various algorithms (e.g., Support Vector Machine (SVM), random forest, neural networks) were done using Principal Component Analysis (PCA). Peptides with high predicted activity (probability >0.5) were selected. Separately, chemical descriptors were calculated and used to classify activity levels (excellent, good, low, inactive). Peptides predicted as “excellent” were prioritized. IC50 values were predicted using structural fingerprints based on the SVM regression algorithm, with performance evaluated via R 2 and RMSE. For lead optimization (predicted favorable ADMET properties), lipophilic-ligand efficiency (LELP) was estimated based on IC50 prediction (−10< LELP <10). Final peptide candidates were filtered based on drug-likeness (score >1.00) and absence of predicted toxicity. This workflow yielded potent tripeptides with strong predicted efficacy and favorable safety profiles (Figures 1A,B). Schematic overview of the research workflow. (A) Machine learning–based categorical assessment integrating interaction fingerprint and molecular descriptor features. (B) Machine learning–based prediction of bioactivity (IC50), ligand efficiency–lipophilicity (LELP) score, and drug-likeness properties. Data collection for machine learning workflows The first step is the training data for interaction fingerprint (IFP) analysis were taken from the Directory of Useful Decoys-Enhanced (DUD-E) database (Mysinger et al., 2012), covering active ligands that are known experimentally to inhibit the HIV type 1 Reverse Transcriptase enzyme (HIV-1-RT), as well as decoy ligands that share physical-chemical similarities but differ significantly in topology and conformation, thereby assuming an inactive role. From DUD-E, a total of 496 ligands (255 active and 241 inactive) were obtained in various formats, such as Mol2, whose activity against RT-HIV is already known. The next is a total of 9,530 ligands associated with RT-HIV were also extracted from the ChEMBL for bioactivity prediction using IC50 (Gaulton et al., 2017) database integrated into DataWarrior V6.1.0. (Sander et al., 2015). The queries were converted the data contains IC50 values in Simplified Molecular-Input Line Entry System (SMILES) format. For 2,197 tripeptide query structures (excluding residues that do not contain simple nonpolar aliphatic side chains such as glycine, alanine, leucine, isoleucine, valine, and methionine, as well as residues with rigid side chains such as proline), were generated using PeptideConstructor 0.2.1 packages in Python 3.11.0 (Tien et al., 2013) and rdkit-pypi packages using 12 L-amino acid, excluding eight non-polar aliphatic R-groups. IFP prediction The 3D IFP analysis between test ligand, the query tripeptide, and receptors using Python-based Protein–Ligand Interaction Fingerprinting (PyPLIF-HIPPOS) tools (Istyastono et al., 2020). The analysis was carried out using the output of the virtual screening Autodock Vina (Trott and Olson, 2010) in the previous step, based on the steps contained in Radifar et al. (2012). HIPPOS performs a similarity comparison between the test ligand and the reference ligand that is experimentally tethered to the PDB structure. The IFP calculation involved both training ligands from ChEMBL and query test tripeptides. The results obtained from the IFP analysis were binary data (70 bins) containing information on the interaction fingerprint patterns between ligands and the receptor, including interactions such as apolar, aromatic face-to-face, aromatic edge-to-face, hydrogen bonds (acceptor and donor), and electrostatic interactions (positive and negative). Bioactivity and LELP prediction using chemical descriptors The information retrieved from chemical property descriptors using DataWarrior V6.4.2, which involves ligand as train and tripeptide as test queries, where initially the ligands train were classified based on the IC50 (µM) range into four classes: (UNAIDS, 2024): excellent (<1 µM), (Margolis et al., 2014), good activity (1–100 µM), (Shi et al., 2016), low activity (100–200 µM), and (de los Rios et al., 2021) inactive (>200 µM). The chemical property descriptors calculated include 48 numerical features, which contain druglikeness, atom counts, ring counts, functional groups, and 3D parameters (excluding categorical parameters). The LELP data is generated from IC50 from subsequent queries. Machine learning test Machine learning model build In this study, two machine learning approaches were applied to predict the activity of tripeptides as active ligands against RT-HIV as the protein target, based on IFP and chemical properties (See Figure 1). A machine learning model was built using Visual Code Studio (Python Version 11) and Orange software 3.37.0 (Demšar et al., 2013). A total of 2,196 tripeptide candidates were examined, and 496 training ligands (255 active and 241 inactive) were taken from the ChEMBL database. For the IFP-based approach, an initial evaluation was conducted using Autodock Vina to obtain the binding affinity data (kcal/mol) and interaction patterns, which were then analyzed using PyPLIF-HIPPOS and converted into 70 binary fingerprints. This data was then processed through PCA, considering the first 30 components. Next, the data was sampled using bootstrapping and data stratification, then used for model training and testing using cross-validation (10-fold CV). Eight algorithms were used to build the model, including AdaBoost, Artificial Neural Networks, Random Forest, Gradient Boosting, Decision Trees, SVM, Naïve Bayes, and k-NN. The models were evaluated based on Area Under the Curve (AUC), Classification Accuracy (CA), F1-score, precision, Matthews Correlation Coefficient (MCC), and sensitivity (recall). The best model was used to predict the activity class of tripeptides, and peptides with an active classification probability above 0.5 were filtered for further IFP-based prediction. For the chemical property-based approach, 9,530 training ligands from ChEMBL were classified based on IC50 values into four classes: Excellent (4,698), Good activity (4,304), Low activity (151), and Inactive (377). All ligands were converted into 48 numerical chemical property descriptors. The data was then processed using PCA (first 15 principal components) and sampled as in the previous IFP approach. Model training and testing procedures, as well as performance evaluation, were performed using the same methods and algorithms. The best model was then used to predict the activity class of tripeptides based on their chemical properties, and peptides with the highest probability in the “Excellent” category were selected as priority candidates. Machine learning model assessment To evaluate the statistical significance of the model and prevent random correlations, a Y-randomization test was performed using 100 random permutations of the response variable. For each permutation, the dependent variable (Y) was randomly shuffled while keeping the feature matrix (X) unchanged. The both models (IFP and chemical descriptor) was trained and evaluated using the same data split and evaluation metrics (accuracy, AUC, precision, recall, F1, and MCC). The statistical significance of the observed model was quantified using the following permutation-based p-value: p=Npermscoreperm≥scoreorig+1Nperm+1 Where N perm denotes the number of random permutations. A low p-value (<0.05) indicates that the model performs significantly better than random chance To enhance model transparency and component most responsible for the observed predictions, SHapley Additive ex Planations (SHAP) analysis was performed using the TreeExplainer algorithm (Lundberg et al., 2017). This approach quantifies the marginal contribution of each input feature to the model’s output based on cooperative game theory, enabling both global and local interpretability of predictions. Since the classification task involved multiclass outputs, SHAP values were computed for each class and then aggregated by calculating the mean absolute SHAP value across all classes to obtain the overall feature importance. To ensure the reliability and generalizability of the developed machine learning model, the applicability domain (AD) was systematically evaluated using both leverage and Mahalanobis distance based metrics. This dual approach is widely recognized in QSAR studies to define the region of chemical space within which model predictions are considered reliable (Roy et al., 2015; Sahigara et al., 2012; Gramatica, 2007). The leverage (hi) of each compound was computed from the hat matrix as follows: hi=xiTXTX−1xi where xi the descriptor vector of the ith compound, and X is the descriptor matrix of the training set. A critical leverage value (h*) was defined as: h*=3p+1n where p is the number of model descriptors, and n is the number of training compounds. Compounds with hi > h* were considered outside the structural applicability region of the model. To complement the leverage-based AD, Mahalanobis distance (MD) was also calculated for each test compound as: MDi=xi−μT Σ−1xi−μ Where μ is the mean descriptor vector of the training set, and Σ is the covariance matrix. A compound was considered outside the AD if its Mahalanobis distance exceeded the empirical threshold: MDthreshold=meanMD+3 x SDMD QSAR model From 9,530 ligands associated with experiments on HIV-1 reverse transcriptase enzymes from DataWarrior V6.4.2, training and testing data were divided at a ratio of 20:80 to predict tripeptide properties based on IC50 values, LELP, drug similarity, and fragment toxicity potential. First, tripeptides in the “excellent” activity category were analyzed alongside 3,313 training ligand data from ChEMBL using SkeletonSpheres descriptors (1024-bin) to capture molecular structure backbond. Structural clustering was then performed based on similarity with a threshold of 0.9. The clustered data were then divided into training data (2,178; 66%) and test data (1,135; 34%), and their pIC50 values were predicted using a SVM model. Model evaluation was performed using linear regression with R 2 and RMSE metrics. The best model prediction results were then used to predict the IC50 values of the candidate tripeptides. Based on this IC50 value, the actual (training data) and predicted (test data) LELP can be calculated based on (Hopkins et al., 2014) using the following equation: LELP=LogPLE LE=1.37 x (−log10⁡ IC50µMNHeavy atoms The LELP model was evaluated using a linear regression model involving actual and predicted LELP values for 2,178 ligands from ChEMBL, which was then used to train a model that predicts LELP values for candidate tripeptides. This model was evaluated using R 2 and RMSE. Drug-likeness The drug-likeness criteria were determined based on predicted values of >1.00 using DataWarrior V6.4.2, according to (Dalafave, 2010; Sander et al., 2015). The drug-likeness of the compounds was assessed using, fragment-based approach derived from a database of ∼3,300 marketed drugs and ∼15,000 non-drug-like compounds (Fluka). The method calculates a drug-likeness score based on the presence of ∼5,300 distinct substructural fragments, each associated with a predefined score obtained from the logarithmic ratio of their occurrence in drug versus non-drug collections. The final drug-likeness value for each compound is computed as the sum of the individual fragment scores present in the molecule. Positive scores indicate the predominance of drug-associated fragments, whereas negative values reflect similarity to non-drug-like fragments. Compounds with positive drug-likeness values were considered favorable and prioritized for further pharmacokinetic and toxicity evaluation. In silico verification Molecular docking assessment Molecular docking assessment was conducted using multiple scoring from AutoDock Vina (AutoDock Vina v1.1.2 Linux x86) and AutoDock4 (AutoDock-GPU v1.6). Different docking tools was used to validate the docking of 35 tripeptide ligands to the Nevirapine binding site of the HIV-1 reverse transcriptase structure (PDB code: 3QIP) (Santos-Martins et al., 2021; Trott and Olson, 2010). The receptor was prepared and cleaned of missing atoms and inappropriate structures using PyMOL (TM) v3.1.6.1 and pdbfixer (https://github.com/openmm/pdbfixer) (GitHub - openmm/pdbfixer). The screening process began with the determination of the grid box size based on the lowest Root Mean Square Deviation (RMSD) values obtained from 50 re-docking simulations using both tools. In AutoDock Vina, the grid box was centered at X = 11.074, Y = 13.791, and Z = 17.342, with dimensions of 35 × 35 × 35 Å. This configuration yielded the lowest RMSD value of 0.118 Å and a docking affinity energy of −11.5 kcal·mol−1. In contrast, in AutoDock4, the grid box was also centered at X = 11.074, Y = 13.791, and Z = 17.342, but with dimensions of 60 × 60 × 60 Å, resulting in the lowest RMSD value of 0.85 Å and a docking affinity energy of −9.08 kcal mol−1. For ligand preparation in AutoDock Vina, the 35 peptide ligands in PDB format were converted to PDBQT format using Open Babel v3.1.0 via the terminal. In contrast, for AutoDock4, ligand conversion to PDBQT format was performed using the ADFRsuite v1.0 program and the Grid Parameter File (GPF) for each ligand was generated using MGLTools v1.5.7 and executed with AutoGrid4 (Morris et al., 2009). AutoDock Vina employs a global search optimization strategy, with the default parameters set as follows: exhaustiveness = 8, number of output poses = 20, and minimum RMSD difference between poses = 1.00 Å. In contrast, AutoDock4 utilizes the Lamarckian Genetic Algorithm (LGA) by default, with both the number of LGA runs and the population size set to 100. For each ligand, the best docking pose along with its corresponding binding affinity was selected as the output. The RMSD of each docked ligand relative to its initial pose was calculated using RDKit v2025.3.3 (Sarthak Pa). MM/GBSA scoring The MM/GBSA approach was applied to refine and validate the docking results by providing a more accurate estimation of binding free energy that accounts for solvation and entropic effects, which are often simplified or omitted in standard docking scoring functions. Compared to AutoDock4 and AutoDock Vina, MM/GBSA offers improved energetic realism and can better distinguish true binders from false positives. Validation of binding energy on ligand and receptor binding was conducted using MM/GBSA analysis based on the method of Uni-GBSA pipeline (Yang et al., 2023). Binding energy validation was carried out using the output results of *.pdb receptor and *.sdf ligand files (after conversion from pdbqt using open babel packages) in Autodock4 GPU screening. Then, the analysis was carried out using the unigbsa-pipeline tool on the Bash Script Linux. Visualization of interactions The ligand–protein complex obtained from molecular docking was first converted from PDBQT to PDB format. The protein molecule was separated from the ligand and saved again in PDB format. The isolated ligand was then converted into SDF format. Both the protein and ligand structures were uploaded separately to the ProteinPlus web server for 2D interaction analysis. The PoseEdit tool within the server was used to visualize and analyze the two-dimensional interactions between the ligand and protein. To visualize the three-dimensional form, the docked ligand–protein complex was loaded into PyMOL. Surface depth analysis was performed to assess the spatial orientation and binding depth of the ligand within the protein’s binding pocket. Pocket assessment similarity Pocket similarity analysis was performed using ProteinPlus web server (https://proteins.plus/) (Schöning-Stierand et al., 2022). The analysis used input files consisting of the protein structure in PDB format and the ligand structure in SDF format. The cavity pocket parameters evaluated included the number of hydrogen bond acceptors and donors, pocket depth (Å), hydrophobicity, presence of metal ions, number of protein heavy atoms, surface area (Å2), surface-to-volume ratio, and pocket volume (Å3). Quantum mechanical analysis Quantum chemical calculations based on (Kohn and Sham, 1965; Cao et al., 2024) were performed to determine the energies of the Highest Occupied Molecular Orbital (HOMO) and the Lowest Unoccupied Molecular Orbital (LUMO) using the Density Functional Theory (DFT) method within the Jaguar module of the Schrödinger suite 2024-4. This process employed the Becke, 3-parameter, Lee-Yang-Parr (B3LYP)/6‒31 G basis set. The results were further refined with the Becke, 3-parameter, Lee-Yang-Parr dispersion correction (B3LYP-D3)3/6-311G basis set. The calculated ELUMO and EHOMO values were then used to determine various quantum chemical properties, including the HOMO-LUMO Gap (ΔG), chemical softness and hardness, global electrophilicity index, and electronegativity. Permeability prediction Permeability of the peptide through biological membranes was analyzed using Permeability of Molecules across Membranes (PerMM) to calculate the energy profiles (Lomize et al., 2019). Meanwhile, membrane-binding affinities were analyzed using the Black Lipid Membrane (BLM) model. All atomic molecular dynamic simulation Molecular dynamics (MD) preparations were performed by using CHARMM-GUI (Jo et al., 2008), and the simulation was performed through GROMACS (Abraham et al., 2015). In CHARMM-GUI, the solution builder feature was used for simulation preparation, including molecule preparation, ligand and ion handling, water box/solvent system, ion placement, and system assembly (Brooks et al., 2009). The parameters used are temperature 310K, NaCl 0.15% concentration, AMBER force field, and time 50 ns. The output of CHARMM-GUI is the initial topology file, along with the preparation for molecular dynamics simulation. Through GROMACS, minimization, equilibration, and simulation stages were carried out to see the trajectory and dynamics of each atom in the protein-ligand system. Visualisation was done using XMGrace tools (Osideko et al., 2025) to see the plot of root mean square deviation (RMSD), root mean square fluctuation (RMSF), and Radius of Gyration (Rg). Principal component analysis (PCA) PCA was conducted using R-studio software, specifically employing the “mktrj.pca” function from the Bio3D package to analyze MD trajectories (Skjærven et al., 2014). The PCA calculation of molecular dynamics was performed using the trajectory (xtc) that had been cleaned from water molecules along with the protein structure file (pdb). PCA functioned as a multivariate exploratory tool to assess the distribution of internal energy variations within MD simulation structures, thereby facilitating the prediction of the system’s thermodynamic stability (David and Jacobs, 2014). Additionally, eigenvalue and eigenvector computations were performed for each complex to identify the two principal components with the greatest variance (Hess, 2002). The PCA model can be mathematically expressed as: M=TKPKT+R where M is the resulting matrix product, TK represents sample correlations, PK corresponds to variable correlations, K is the number of parameters in the input, and R denotes the residual matrix. Beyond thermodynamic analysis, PCA was also utilized to investigate the directional movements of protein structures by measuring fluctuations in the Cα atoms within protein complexes (David and Jacobs, 2014). These atomic motions provide insights into the protein’s stability and functional dynamics. The PCA of Cα fluctuations begins by removing all solvent molecules and ions from the system, followed by calculating eigenvalues and eigenvectors for the top three principal components (PC1, PC2, and PC3), where these PCs represent vector values within the matrix (Gholam et al., 2024; Ezaj et al., 2022). The principal component values are derived from the covariance matrix using the formula: C=VAVT Here, C is the covariance matrix, V is the matrix of eigenvectors, and A is the diagonal matrix of eigenvalues. Dynamic cross-correlation matrix (DCCM) DCCM analysis was performed using R-studio with the “dccm” function in the Bio3D package (Grant et al., 2006). The DCCM plot is used to determine the correlation between the position of the Cα atom and other atoms in the structure, both directly and indirectly, especially the displacement that occurs in the Cα atom. The DCCM value is calculated based on the following mathematical equation: DCCMij=d→i . d→jdi2dj2 The values in and dj represent the shift from the initial position to the average position. The DCCMij value ranges from −1 to 1, where positive correlation indicates parallel interaction between atoms, while negative correlation describes antiparallel interaction between Cα atoms and other atoms in the protein structure. Free energy landscape (FEL) Free energy landscape analysis was performed by evaluating the results of Gromacs simulations through fitted trajectory files (xtc), structure topology (tpr), and index files (ndx) (Yekeen et al., 2023). Throughout the entire trajectory interval (50 ns), the gmx covar command was run to generate output in the form of an eigenvalue file (eigenval.xvg) and an eigenvector file (eigenvec.trr), which represent the covariance of the overall atomic motion in the system. Next, gmx anaeig was used to project the first two principal components (PC1 and PC2), with the output being a PCA_2dproj.xvg file that describes the conformation distribution in PCA space. The projection results are then analyzed using gmx sham to construct a Gibbs free energy map (kJ/mol) in xpm format, which is then converted into images (eps and pdf) via gmx xpm2ps and ImageMagick. The Gibbs free energy calculation formula is as follows: ΔGi=−kTPirPmaxr In this equation, k represents the Boltzmann constant, while T denotes the simulation temperature. The value of P i (r) describes the probability of the system being in a particular state i determined by the reaction coordinate r (the observed parameter), and is obtained through a histogram of MD results. Meanwhile, Pmax(r) is the probability in the bin with the highest population or the most likely state. The change in free energy for state i is then expressed as ΔG i . Steered molecular dynamic simulation Steered Molecular Dynamics (SMD) simulations were performed using CHAPERONg interfaced with GROMACS (Yekeen et al., 2023; Lemkul, 2019). The simulation was initiated by preparing the simulation parameters in the paraFile.par, which included the following settings: the protein structure (pdb format) was placed in a cubic simulation box and solvated with the TIP3P water model. Counterions (Na+ and Cl−−) were added to neutralize the system at a physiological salt concentration of 0.1 M, and the temperature was maintained at 310 K. System conditioning and atomistic parameters were defined using the CHARMM36 force field version July 2022 (http://mackerell.umaryland.edu/), In addition, using python script cgenff_charmm2gmx_py*_nx*.py from the same website are used to convert small molecule ligand file (str and mol2) topology parameters from CHARMM General Force Field (CGenFF) (https://app.cgenff.com/homepage) into a format that can be used by GROMACS. All simulations were carried out in fully automated mode, employing 2 CPU threads with non-bonded interactions calculated on the CPU. The system trajectory was processed in centered mode to remove translational artifacts, and 250 frames (400 ps) with timestep 2 fs were extracted to generate a representative molecular movie. Furthermore, molecular dynamics parameter (MDP) files were prepared for various steps, including ionization, energy minimization, equilibration (NPT), ligand pulling simulation, and the production run. Including the setup of the pulling coordinate velocity at 0.01 nm·ps−1, the pulling simulation with a harmonic potential force constant of 1,000 kJ ·mol−1·nm−2, the V-rescale thermostat (310 K), pressure control with the isotropic Parrinello–Rahman barostat (1 bar), and the use of Particle Mesh Ewald (PME) (grid spacing 0.12 nm, order 4). Pharmacokinetics prediction Pharmacokinetic prediction of compounds was performed using the pkCSM platform (http://biosig.unimelb.edu.au/pkcsm/) (Pires et al., 2015), a graph-based signatures web server used to analyze ADME (Absorption, Distribution, Metabolism, and Excretion) properties of small molecules based on input in the form of compound structure representations in SMILES format. Each compound was transformed into SMILES notation and entered into pkCSM to obtain pharmacokinetic parameters. These prediction results were used to evaluate the potential of potential tripeptides. Toxicology prediction The potential toxicity profile was assessed using DataWarrior V6.4.2 (Sander et al., 2015), which included parameters such as mutagenicity, tumorigenicity, reproductive effects, irritation, and the presence of harmful functions (nasty functional groups) or structural patterns such as Pan-Assay Interference Compounds (PAINS). Only compounds yielding results of “not available” or “not risky” were considered acceptable for further evaluation. In addition, toxicity endpoints were further evaluated using the pkCSM platform (http://biosig.unimelb.edu.au/pkcsm/) (Pires et al., 2015), including AMES mutagenicity, maximum tolerated dose (human) (MTD), hERG I and hERG II inhibition, oral rat acute toxicity (LD50), oral rat chronic toxicity (LOAEL), and hepatotoxicity. Results Machine learning model evaluation and external data prediction The classification performance of eight ML algorithms was evaluated using six metrics: AUC, CA, F1-score, Precision (Prec), Recall, and MCC (Supplementary Figures S1 and S2). The ensuing results are illustrated in Figure 2, with Figure 2A representing the first-layer classification based on the top 30 principal components (PCs) from interaction fingerprint features, and Figure 2B representing the second-layer classification based on the top 15 PCs from chemical property descriptors. Evaluation of machine learning model based on (A) interaction fingerprints, and (B) chemical properties. Among all models tested for IFP prediction (Figure 2A), AdaBoost, Artificial Neural Networks, and Random Forests consistently showed superior performance across all metrics, particularly in AUC values approaching or exceeding 0.9, indicating high discriminative power, as well as F1 scores (approaching 0.8) compared to other models, signifying a harmonic mean between precision and recall. Among these models, AdaBoost recorded the highest MCC value, indicating strong agreement between predictions and actual class labels. Conversely, the SVM, Naïve Bayes, and k-Nearest Neighbors (kNN) models demonstrated comparatively lower performance. Notably, Naïve Bayes and k-Nearest Neighbors (kNN) performed relatively poorly, particularly in terms of F1 and MCC metrics. The Decision Tree demonstrated intermediate performance, exhibiting stable yet unremarkable scores across all metrics. According to the IFP model prediction, out of 1,106 tripeptides from 2,197 were predicted to be active using the AdaBoost algorithm with minimum probability >0.5. In the second layer model (Figure 2B), which uses chemical descriptor features, the overall classification performance is slightly better than in the first layer. The performance of eight machine learning models used to predict IC50 activity classification was also evaluated. It shows Random Forest, AdaBoost, and Neural Network models consistently outperformed others across most evaluation metrics, particularly in terms of AUC and F1-score, suggesting their robustness in handling multi-class classification tasks, maintaining their status as the top performers. Random Forest achieved the highest AUC values supported by its high MCC values, indicating its superior ability to distinguish between the different activity classes. Neural networks exhibited consistent performance, though they demonstrated a slight decrease for almost every single evaluation metric. Still, there was also a marked improvement in the AUC metric in comparison to interaction-based models. As in the initial layer, kNN, Naïve Bayes, and SVM demonstrated the least robust results, particularly in terms of MCC scores. Naïve Bayes and SVM showed comparatively lower performance, especially in MCC and F1, suggesting limitations in accurately capturing complex chemical patterns. Using chemical descriptor model prediction, among 1,106 peptides, 628 peptides have predicted excellent bioactivity with Random Forest approach with minimum probability >0.3. Model robustness and validation The robustness and predictive reliability of the developed two-layer machine learning framework were comprehensively evaluated using standard performance metrics and Y-randomization testing to verify that the observed predictive power did not arise from chance correlations. In the first-layer model, based on IFP features and constructed using the AdaBoost algorithm, the classifier demonstrated excellent predictive performance with an AUC of 0.9348, classification accuracy (CA) of 0.8732, precision of 0.8693, recall of 0.8732, F1-score of 0.8700, and MCC of 0.7317 (Table 1). TABLE 1Performance metrics of the first-layer AdaBoost model based on IFP features.MetricsValueArea under the ROC curve (AUC)0.9348Classification accuracy (CA)0.8732Precision0.8693Recall (sensitivity)0.8732F1-score0.8700Matthews correlation coefficient (MCC)0.7317 Performance metrics of the first-layer AdaBoost model based on IFP features. To confirm that this performance was not the result of random associations, a Y-randomization (permutation) test with 100 iterations was conducted using AUC as the primary evaluation metric. The randomized models yielded an average AUC of 0.50 ± 0.02, whereas the original AdaBoost model retained a substantially higher AUC of 0.9348, resulting in a permutation-based p-value of 0.0099 (p < 0.05). These results indicate that the AdaBoost classifier captures a genuine and statistically significant relationship between IFP-derived molecular patterns and peptide bioactivity. The distribution of AUC values obtained from the permuted and original models is depicted in Figure 3A, where the grey histogram bars represent the randomized models and the red dashed line denotes the original model’s AUC. Y-randomization validation of (A) the first-layer AdaBoost model using IFP features, and (B) the second-layer Random Forest model using chemical descriptor features. In the second-layer model, based on chemical descriptor features and developed using the Random Forest algorithm, the classification performance also remained robust. The model achieved an AUC of 0.8802, CA of 0.7964, precision of 0.7973, recall of 0.7964, F1-score of 0.7964, and an MCC of 0.5934 (Table 2). A similar Y-randomization test (100 permutations) yielded an average AUC of 0.5039 ± 0.02, whereas the original model maintained an AUC of 0.8802, yielding a permutation-based p-value of 0.0099 (p < 0.05). TABLE 2Performance metrics of the second-layer Random Forest model based on chemical descriptor features.MetricsValueArea under the ROC curve (AUC)0.8802Classification accuracy (CA)0.7964Precision0.7973Recall (sensitivity)0.7964F1-score0.7964Matthews correlation coefficient (MCC)0.5934 Performance metrics of the second-layer Random Forest model based on chemical descriptor features. This finding confirms that the Random Forest model effectively learns true structure activity relationships (SARs) within the chemical descriptor feature space rather than spurious patterns. The comparative performance distribution between the original and randomized models is shown in Figure 3B, where the red dashed line marks the original model’s AUC. Overall, both classification layers demonstrated statistically significant and reproducible predictive performance, establishing the robustness of the proposed multi-layer ML framework in modeling peptide bioactivity based on IFP and chemical descriptor features. Having confirmed the statistical robustness and validity of both classification layers, further analyses were carried out to explore the interpretability of each model using SHAP-based feature attribution, aiming to uncover the most influential structural and physicochemical factors driving the observed predictive behavior Model explainability and validation (SHAP analysis) The SHAP dependence analysis provided a comprehensive interpretation of how each principal component (PC) contributed to the AdaBoost classification model (Figure 4A). Among the 30 PCA-derived features, PC3, PC25, and PC11 exhibited the highest average SHAP magnitudes (0.0232, 0.0231, and 0.0184, respectively), suggesting that these latent molecular descriptors exerted the strongest influence on the predicted activity outcomes. The dependence plots demonstrate nonlinear and asymmetric relationships between individual PC values and SHAP impacts, indicating that changes in specific principal components led to either positive or negative shifts in predicted activity probability depending on their structural information. SHAP dependence plots for (A) all 30 PCA components used in the AdaBoost model (first-layer), and (B) all 15 PCA components used in the Random Forest model (second-layer). For example, PC3 showed a monotonic increase in SHAP values with increasing component score, implying a direct relationship between this latent feature and model confidence toward active class prediction. Meanwhile, PC25 and PC11 exhibited more complex, nonlinear effects, suggesting interacting latent features that jointly modulate model output. Intermediate components such as PC17, PC29, and PC1 also showed moderate SHAP contributions (|SHAP| ≈ 0.015–0.013), reinforcing that the predictive signal is distributed across multiple latent variables, not limited to a few dominant axes. Lower-ranked components (e.g., PC16–PC30) exhibited near-zero SHAP values, indicating minimal influence on classification decisions. Overall, the SHAP analysis confirms that the AdaBoost model learned chemically meaningful and statistically consistent patterns, while PCA successfully preserved essential structure–activity information rather than introducing noise or redundancy. The SHAP dependence analysis of the Random Forest model provided detailed insights into the relative importance and interaction behavior of PCA-derived chemical descriptor features (Figure 4B). Among the 15 principal components used as inputs, PC14, PC10, and PC4 exhibited the highest mean absolute SHAP values (0.0335, 0.0228, and 0.0222, respectively), indicating that these latent variables contributed the most to model predictions of IC50 activity classes. The dependence plots revealed several distinct nonlinear and asymmetric patterns, showing that changes in certain principal components could either increase or decrease SHAP values depending on their underlying chemical variance. For instance, PC14 demonstrated a strong positive correlation between component score and SHAP value, implying that molecules with higher PC14 scores were more likely to be predicted as highly active. By contrast, PC10 and PC4 displayed non-monotonic behaviors, where mid-range component values produced maximum SHAP impacts, suggesting localized feature interactions in chemical descriptor space. Lower-ranked components (e.g., PC7–PC15) showed smaller yet interpretable SHAP contributions, reinforcing that the Random Forest model learned from a distributed set of latent molecular features rather than relying on a single dominant variable. This result also supports the effectiveness of PCA in dimensionality reduction retaining chemically informative components while minimizing redundancy. Overall, the SHAP analysis confirms that the second-layer Random Forest model identified statistically robust and chemically meaningful structure–activity relationships, capturing generalizable descriptors that explain variations in peptide IC50 values beyond noise or random correlation. Beyond overall performance, SHAP-based interpretability analysis was conducted to elucidate which PCA-derived features most strongly influenced model predictions in both classification layers (Figure 4). The resulting dependence plots revealed nonlinear and chemically meaningful feature–response relationships, confirming that the models learned interpretable structure–activity patterns rather than noise-driven correlations. To further interpret the chemical significance of the most influential components identified by SHAP, PCA loading analyses were conducted separately for each classification layer. This step links latent PCs to tangible chemical or structural features, providing mechanistic insight into model behavior. The loading plots (Figure 5A) revealed that PC3, PC25, and PC11 were dominated by fingerprint bits such as bit_7, bit_36, and bit_67, which correspond to distinct molecular substructures associated with hydrogen bonding, aromatic stacking, and peptide–target interactions. Positive loadings (e.g., bit_7 = +0.42) were related to active peptide patterns, whereas negative loadings (e.g., bit_49 = −0.31) indicated structural motifs linked to inactivity. These findings confirm that AdaBoost captured specific interaction-level information that distinguishes active from inactive peptides. Principal component loading plots and heatmaps of feature contributions in the two-layer ML models. (A) PCA loading plots of top components (PC3, PC25, PC11) in the first-layer AdaBoost model based on interaction fingerprints. (B) PCA loading plots of top components (PC14, PC10, PC4) in the second-layer Random Forest model based on physicochemical descriptors. (C and D) Heatmaps of PCA loading coefficients showing clusters of co-varying features across interaction fingerprints (C) and physicochemical descriptors (D). The global relationship among all fingerprint bits is visualized in the heatmap (Figure 5C), which shows that correlated fingerprints cluster together, suggesting the model recognizes co-occurring substructural motifs rather than isolated features. For the second-layer Random Forest model, the loading plots (Figure 5B) identified PC14, PC10, and PC4 as key components dominated by global molecular descriptors such as Symmetric Atoms (+0.46), Globularity SVD (−0.45), Hetero-Rings (+0.39), and Druglikeness (+0.30). These descriptors highlight that peptides with higher symmetry, moderate hetero-ring presence, and favorable topological complexity tend to show greater bioactivity probabilities. In contrast, more globular and irregularly structured molecules contributed negatively to model confidence. The accompanying heatmap (Figure 5D) visualizes the overall distribution of descriptor loadings across components. Clear feature clusters are visible for ring-related and aromaticity-based descriptors, indicating that Random Forest learned chemically coherent feature groupings that reflect underlying structure–activity relationships. Applicability domain (AD) and model generalization The applicability domain (AD) analysis was conducted to evaluate the structural reliability and generalization capability of both classification layers. The AD was examined using Williams plots, which integrate leverage (h*) and Mahalanobis distance (MD) criteria to identify potential outliers or structurally influential compounds. The quantitative summary of AD metrics for each model is presented in Table 3, while the graphical representations are illustrated in Figure 6. TABLE 3Application Domain (AD) metrics for both classification models.Model layerAlgorithmFeature typeLeverage threshold (h*)Mahalanobis distance threshold (mean + 3SD)Outliers detectedInterpretationFirst layerAdaBoostPCA-transformed interaction fingerprint (IFP) features0.19Mean + 3SD1 compoundModel operates within a well-defined applicability domain; one compound identified as structurally distinct from the training domain.Second layerRandom forestPCA-transformed chemical descriptor features0.01Mean + 3SDNoneAll compounds fall within the AD boundaries; predictions are chemically consistent and generalizable. Application Domain (AD) metrics for both classification models. Williams plot of: (A) The first-layer AdaBoost model based on PCA-transformed IFP features. The plot shows leverage (x-axis) versus Mahalanobis distance (y-axis). The red dashed line indicates the leverage threshold (h* = 0.19), and the green dashed line marks the Mahalanobis distance limit (mean + 3SD). Most compounds are within both boundaries, with only one structural outlier detected. (B) The second-layer Random Forest model based on PCA-transformed chemical descriptor features. The red dashed line represents the leverage threshold (h* = 0.01), and the green dashed line denotes the Mahalanobis distance threshold (mean + 3SD). All compounds fall within the AD region, confirming that predictions are made within a chemically reliable and generalizable domain. In the first-layer AdaBoost model, which utilized PCA-transformed IFP features, the leverage threshold (h*) was 0.19. As shown in Figure 6A, the vast majority of compounds were located within the defined AD boundaries, indicating that predictions were made on structurally representative samples. Only one compound exceeded both the leverage and Mahalanobis distance thresholds, suggesting that it represents a structurally unique tripeptide distinct from the main chemical domain. Overall, the model operated within a well-defined and statistically robust applicability domain, ensuring reliable predictions for most compounds. For the second-layer Random Forest model, based on PCA-transformed chemical descriptor features, the leverage threshold (h*) was 0.01, and no compounds exceeded the predefined boundaries (Figure 6B). All predictions were made within the valid chemical domain, demonstrating that the model generalizes well to unseen data and avoids extrapolation beyond the trained chemical space. Collectively, these results confirm that both classification layers exhibit structural consistency, robustness, and generalizability, reinforcing confidence in their predictive outputs. Overall, the Y-randomization, SHAP interpretability, and AD analyses collectively validate that the proposed multilayer AdaBoost–Random Forest framework achieves high predictive accuracy while maintaining interpretability and domain reliability across the chemical space. QSAR model evaluation and external data prediction The regression-based QSAR models showed strong predictive performance for both bioactivity and lipophilic efficiency. The IC50 prediction model achieved an R 2 of 0.955 and RMSE of 0.231, with a regression equation (y = 0.9329x + 0.0007) based 1,655 training ligands, closely matching the ideal line, indicating reliable prediction of inhibitory potency across diverse tripeptides (Supplementary Figure S2A). Similarly, the LELP values calculated from IC50 prediction yielded an excellent model performance, with an R 2 of 0.994 and RMSE of 0.469, showing a strong linear fit (y = 0.9742x + 0.22) based on 2,178 training ligands, suggesting accurate estimation of lipophilicity-related efficiency (Supplementary Figure S2B). Together, these models support the early identification of peptide candidates with both high potency and favorable pharmacokinetic profiles. Within 628 peptides, 88 peptides showing ideal features of LELP (−5