AI-assisted discovery of potent FGFR1 inhibitors via virtual screening and in silico analysis

Abstract

Fibroblast growth factor receptor 1 (FGFR1) is recognized as an oncogene that fosters tumor development, playing a vital role in cancer progression. This has established it as a promising target for cancer drug development. However, existing FGFR1 inhibitors are often limited by drug resistance and lack of specificity, emphasizing the need for more selective and potent alternatives. To address this challenge, the present study employed an AI-driven virtual screening approach, integrating molecular docking (MD) and molecular dynamics simulations (MDS) to discover novel FGFR1 inhibitors. A voting classifier integrating three machine learning classifiers was utilized to screen 10 million compounds from the eMolecules database, leading to 44 promising candidates with a prediction probability exceeding 80%. MD identified compound with PubChem Compound Identifier (CID) 165426608 (−10.8 kcal/mol) as the highest-scoring ligand, while compounds with CID 145940129 (−9.8 kcal/mol), CID 131910163 (−9.4 kcal/mol), CID 155915988 (−9.2 kcal/mol), and CID 132423733 (−9.1 kcal/mol), exhibited binding affinities comparable to or slightly lower than that of the native ligand (−10.4 kcal/mol). MDS further revealed that all these compounds, except CID 131910163, maintained structural stability with time. Thermodynamic stability assessment confirmed the spontaneity and feasibility of their complex formation reactions with negative ΔG_BFE values ranging from −21.87 to −12.76 kcal/mol. Decomposition of binding free energy change further provided key stabilizing residues. The heatmaps and histograms of the interaction over the full 200 ns simulation period highlighted the prominent interaction profiles. Structural similarity analysis of the four MDS-stable compounds displayed the dice similarity scores of 0.200000 to 0.452830 with known FGFR1 inhibitors. Additionally, the pIC₅₀ prediction using a voting regressor indicated promising pIC₅₀ values (7.07 to 7.47), highlighting their potential as hit candidates for further structural optimization and therapeutic development. Further, this study underscores the efficiency of machine learning-based virtual screening and in silico analysis as a cost-effective and reliable strategy for accelerating hit drug discovery from large datasets, even with limited resources and time.

1. Introduction

Cancer represents a critical global health challenge, accounting for one in every six deaths worldwide. In 2022, around 20 million individuals were newly diagnosed with cancer, and approximately 9.7 million deaths were attributed to cancer-related diseases [1]. Despite significant advancements in different therapeutic approaches such as chemotherapy, hormone therapy, and immunotherapy, cancer mortality rates remain alarmingly high. This is mainly because of the complex genetic and phenotypic diversity of cancer, and the emergence of drug-resistant phenotypes [2].

The fibroblast growth factor receptor (FGFR) signaling axis plays a crucial role in transducing signals that govern various cellular processes, including proliferation, angiogenesis, differentiation, embryonic development, migration, organogenesis, and survival [3]. Members of the fibroblast growth factor receptor family, including FGFR1, frequently undergo genomic alterations such as mutations, amplifications, and gene fusions across various cancer types [4]. FGFR1, in particular, has been extensively studied and recognized as an oncogene that fosters tumor development, underscoring its critical role in cancer progression [5]. Overexpression of FGFR1 has been observed in cancers such as breast [6,7], lung [8], ovarian [9,10], bladder [11], prostate [12,13], and gastric cancers [14], among others. Consequently, targeting FGFR1 for cancer therapy has become an appealing therapeutic strategy [15]. To date, drugs such as Regorafenib, Nintedanib, Sorafenib, Lenvatinib, Erdafitinib, Pemigatinib, Infigratinib, and Futibatinib have received FDA approval for FGFR1 inhibition [16–18]. These inhibitors work by reducing FGFR1 activity, which is often overexpressed in certain cancers, thereby impeding tumor growth. However, the efficacy of these inhibitors is limited due to challenges like drug resistance and lack of specificity [19]. Therefore, the development of novel inhibitors with enhanced effectiveness and reduced side effects remains a significant challenge.

Traditional drug discovery methods rely heavily on in vivo experiments and in vitro screening, which are both costly and labor-intensive [20]. Preclinical drug discovery constitutes approximately one-third of the drug development expenses and usually takes nearly five and a half years [21,22]. The high failure rate during drug development further exacerbates the costs. As a result, methodologies that can reliably predict success at early stages are critically valuable. Computer-aided drug design (CADD) has emerged as a transformative approach in this domain [23]. By employing in silico techniques, CADD accelerates drug discovery and reduces the time required for identifying leads and introducing new drugs. These methods also enable the prediction of biological activity for chemical compounds against specific targets [24]. In this study, an Artificial Intelligence (AI)-driven virtual screening approach of millions of molecules was adopted to advance FGFR1-targeted drug discovery [25,26]. By leveraging AI’s ability to analyze vast datasets and predict drug efficacy in a relatively short time span, this study aims to streamline the drug development process and increase the likelihood of identifying alternate treatments [27]. Fig 1 outlines the detailed workflow adopted in this study.

Fig 1. Detailed workflow of the study.

2. Materials and methods

2.1. Data collection and curation

The Chemical European Molecular Biology Laboratory (ChEMBL) database was used to obtain the Simplified Molecular Input Line Entry System (SMILES) representations and half-maximal inhibitory concentration (IC₅₀) values for 2,153 FGFR1 inhibitors [28]. The IC₅₀ value represents the concentration of a compound required to inhibit a specific biological process or activity by 50% which serves as a preliminary guide for selecting efficient and biologically active molecules. After filtering entries without IC₅₀ values, retaining bioactivity data measured in nanomolar (nM), and removing duplicates, 1876 data points remained. The IC₅₀ values were transformed into pIC₅₀ values using negative logarithms to standardize the data. Lipinski’s Rule of Five (RO5) was applied to assess drug-likeness and exclude less potent compounds [29,30], resulting in 1523 data points for model training. Radar plots depicting the physicochemical properties of the filtered dataset are shown in Fig 2.

Fig 2. Physicochemical radar plots of the compound in the dataset that (a) fulfill RO5 and (b) violate RO5.

2.2. Model building and database screening

Molecular fingerprints [31], encoded as numerical vectors or bit-strings, facilitate rapid similarity evaluations critical for virtual screening [31,32], structure-activity relation studies, and chemical space mapping [33]. Using the RDKit toolkit [34], fingerprints from SMILES entries were computed, and the dataset was classified into 813 active and 710 inactive compounds (1523 total) using a pIC₅₀ threshold of 7.0, as a cut-off ranging from 5 to 7 has been recommended [26]. Based on the Morgan3 protocol, which employs 2048 bits as a circular fingerprint [35], machine learning models were constructed using Scikit-learn and XGBoost. The Morgan3 fingerprints (radius = 3) encode molecular features extending up to three bonds from each atom, allowing the capture of broader substructural patterns that may play a vital role in determining biological activity. Scikit-learn is a versatile machine-learning library that provides a diverse range of algorithms for classification, regression, clustering, and dimensionality reduction tasks [36]. XGBoost is an advanced library tailored for the fast and scalable execution of gradient-boosting algorithms [37]. Twenty classification models were trained, and the best-performing models were fine-tuned to create a voting classifier, amplifying accuracy and robustness in comparison with the individual models [38]. A similar approach was applied for building a voting regressor. The voting classifier was then employed to screen 10 million compounds from the eMolecules database [39,40]. Compounds with invalid SMILES, those violating the Rule of Five (RO5), and Pan-Assay Interference Compounds (PAINS) were excluded prior to screening.

Classification models’ performance was evaluated using accuracy, precision, sensitivity, specificity, and Area Under Curve (AUC) metrics, calculated based on the confusion (error) matrix [41]. Regressors were assessed according to mean absolute error (MAE), root-mean-squared error (RMSE), and R2 scores [42]. Multiple steps were used systematically for model training and database screening to identify potential molecules.

2.3. Molecular docking calculations

The 44 potential inhibitors with prediction probabilities above 80% obtained from the screening of 10 million compounds from the eMolecules database were selected as hit ligands. The 3D structures of 35 compounds available in the PubChem database (https://pubchem.ncbi.nlm.nih.gov/) [43] were retrieved in SDF format, while the remaining 9 compounds were drawn using their SMILES strings. The molecular formulas were verified using the Avogadro program (v1.2.0) [44] after adding the hydrogen atoms. Energy minimization was carried out using the UFF force field with 5000 steps employing a conjugate gradient algorithm, ensuring energy convergence at 1.0 × 10⁻⁸ kcal/mol. This process was repeated until the global minima was reached. The bond orders, including double bond positions, were examined, and steric hindrance or stress was removed. Finally, the optimized ligands were converted to PDBQT format with Gasteiger charges using AutoDock Tools [45].

The 3D crystal structure of FGFR1 (PDB ID: 4ZSA, DOI: https://doi.org/10.2210/pdb4ZSA/pdb) with X-ray crystallographic resolution of 2.00 Å was obtained from the RCSB database (https://www.rcsb.org/) [46]. Missing amino acid residues were repaired using the SwissModeling server (https://swissmodel.expasy.org/) [47], where model_01 of template 5B7V.1.A (global model quality estimate: 0.88, qualitative model energy analysis with distance constraints: 0.83 ± 0.05) was selected due to its 100% sequence identity. The finalized protein structure was converted to PDBQT format with the addition of polar hydrogens and Kollman charges using the AutoDock Tools. The apo form of the protein was then utilized as the target for computational analyses.

The molecular docking calculations of the ligands with the FGFR1 protein were done using the user-friendly software AutoDock Vina [48], with the same protocol as outlined by Phunyal et al. [49], with slight modifications of parameters. A grid box size of (50, 50, 50) ų, the grid center at (x: 5.060, y: − 0.501, z: 16.013), an energy range of 4, and 20 number of modes were used with an exhaustiveness (converged) of 64. The five protein-ligand complexes with the top binding affinities were saved in pdb format and utilized for MDS. The binding interaction between the protein and ligand was visualized using the PyMOL [50] program and the protein-ligand interaction profiler (PLIP) [51] web server.

2.4. Molecular dynamics simulations

The GROMACS (version 2021.2) software [52] was used to simulate the protein-ligand complex, with the CHARMM36 force field [53] applied to the receptor, while the ligand parameters were derived from the SwissParam server [54]. The system was solvated using the TIP3P water model in a triclinic box with a 12 Å spacing at the sides, to prevent any unwanted effects caused by repetition of the simulation box. Neutralization was achieved by adding counter ions, followed by the inclusion of an isotonic NaCl solution (0.15 M). Equilibration was carried out in four steps with each of 1 ns at 310 K and 1 bar, with the first two using the NVT ensemble and the last two using NPT. The V-rescale thermostat, which is a modified version of the Berendsen method, was used for the temperature coupling, while pressure coupling was applied through the isotropic Parrinello-Rahman approach. The final 200 ns production run, with a 2 fs step size, was performed without constraints on the protein-ligand complex. Additional parameter details for different system setups can be found in the literature [55–58]. The complex was centered and analyzed using GROMACS built-in modules to obtain geometric parameters such as snapshots, root mean square deviation (RMSD) of the ligand and protein backbone, root mean square fluctuation (RMSF), radial pair distribution function (RPDF), radius of gyration (R_g), and solvent-accessible surface area (SASA).

The thermodynamic stability of the protein-ligand complexes was evaluated using the Molecular Mechanics Poisson-Boltzmann Surface Area (MMPBSA) binding free energy calculations [59], following the parameters as utilized by Shrestha et al. [60]. A 20 ns equilibrated segment from the 200 ns molecular dynamics trajectory was used for this purpose. Calculations were carried out using the MMPBSA module [61], which applies the Poisson–Boltzmann solvation model. Binding free energy was computed at 100 ps intervals to assess the stability, spontaneity, and feasibility of complex formation over time. The spontaneity and viability of the forward reaction were evaluated based on the sign of the free energy changes. To investigate the contribution of individual amino acid residues to binding free energy change, decomposition analysis was performed on the same 20 ns equilibrated segment using the g_mmpbsa tool [62], which allowed the decomposition of the total binding energy to identify key stabilizing and destabilizing residues. The associated gmx_MMPBSA_ana subprogram was used for data analysis and visualization.

Additionally, to understand residue-level interaction dynamics over the entire course of the simulation, amino acid interaction heatmaps and histograms were generated using the full 200 ns trajectory (20,000 frames).

2.5. Similarity analysis

Based on the principle that structurally similar compounds often share chemical and biological properties, a similarity analysis between potential and known FGFR1 inhibitors [63] was conducted. The relevance of this analysis is closely tied to the nature of structure-activity relationships (SARs) that define biologically active molecules, serving as key factors for the success of ligand-based virtual screening, regardless of the methods employed [64]. Using Morgan2 fingerprints, the similarity maps based on the dice similarity metric were generated, highlighting structural features influencing biological activity [65,66]. Morgan2 fingerprints (radius = 2) encode molecular features extending up to two bonds, emphasizing localized structural variations. In similarity maps, Morgan2 is beneficial for highlighting key local substructures that impact activity, enhancing the interpretability of the visualization.

2.6. Computational resources

All the calculations, including plot generation, were executed on high-performance multiprocessor systems. The machine learning computations were conducted on a system featuring 96 cores, 256 GB of RAM, a 16 GB GPU accelerator, and running Ubuntu 20.04 LTS. Meanwhile, MD and MDS were performed on a system with 24 cores, a 24 GB GPU accelerator, and running Ubuntu 20.04 LTS. The analysis and the visualization of the data were done on a personal computer with Windows 11 operating system.

3. Results and discussion

3.1. Model evaluation and screening results

Twenty classification models were evaluated, and their performance metrics are summarized in Table 1.

Table 1. General performance of 20 different classification models.

Models	Accuracy	AUC
SVC	0.88	0.88
ExtraTreesClassifier	0.88	0.88
NuSVC	0.88	0.87
RandomForestClassifier	0.88	0.87
XGBClassifier	0.88	0.87
LGBMClassifier	0.85	0.85
CalibratedClassifierCV	0.85	0.85
BaggingClassifier	0.85	0.85
LogisticRegeression	0.85	0.85
Perceptron	0.84	0.84
PassiveAggressiveClassifier	0.84	0.84
KNeighborsClassifier	0.83	0.83
DecisionTreeClassifier	0.82	0.82
SGDClassifier	0.82	0.82
ExtraTreeClassifier	0.82	0.82
LinearSVC	0.81	0.81
AdaBoostClassifier	0.81	0.81
BernouliiNB	0.81	0.81
NearestCentroid	0.80	0.80
RidgeClassifierCV	0.80	0.80

The Support Vector Classifier (SVC), ExtraTreesClassifier (ET), and Extreme Gradient Boosting Classifier (XGB) demonstrated superior accuracy and AUC scores, leading to their integration into a voting classifier with a soft voting mechanism. The soft voting mechanism is an ensemble learning technique that predicts the final class by averaging the probability estimates from multiple models and selecting the class with the highest mean probability [67]. For SVC, the parameters were set to C = ‘2.0’ and probability = ‘True’; for ET, n_estimators = ‘400’, criterion = ‘log_loss’, and max_features = ‘log2’; for XGB, n_estimators = ‘1000’, max_depth = ‘5’, and learning_rate = ‘0.04’; all other parameters were set to their default values.

Fig 3 presents confusion matrices for individual and voting classifiers, while Table 2 summarizes the five-fold cross-validation results.

Fig 3. Confusion matrix for (a) SVC; (b) ExtraTreesClassifier; (c) XGB Classifier; (d) Voting Classifier.

Table 2. Five-fold cross-validation of individual classifiers and voting classifier using five parameters.

Model	Accuracy	Precision	Sensitivity	Specificity	AUC
SVC	0.88 ± 0.01	0.88 ± 0.01	0.90 ± 0.02	0.86 ± 0.02	0.95 ± 0.01
ET	0.88 ± 0.01	0.88 ± 0.02	0.90 ± 0.03	0.86 ± 0.03	0.95 ± 0.01
XGB	0.88 ± 0.01	0.88 ± 0.02	0.90 ± 0.01	0.86 ± 0.03	0.94 ± 0.01
VC	0.89 ± 0.01	0.89 ± 0.02	0.91 ± 0.01	0.87 ± 0.02	0.95 ± 0.01

ROC curves in Fig 4 illustrate the excellent discrimination ability of these models.

Fig 4. ROC curves of the individual models and the voting classifier.

The AUC scores show that all the classifiers have excellent discrimination abilities between active and inactive compounds. Additionally, a voting classifier was tested on an external test set, which consisted of FDA-approved selective inhibitors of FGFR1- Erdafitinib, Pemigatinib, Infigratinib, and Futibatinib. The model classified all these drugs as active with high prediction probabilities (> 90%), further demonstrating the reliability of our model. Based on these results, the voting classifier was used to screen the eMolecules database, identifying 44 compounds with prediction probabilities above 80% as potential active inhibitors of FGFR1 protein.

3.2. Docking score comparison and interaction analysis

The molecular docking protocol was validated by docking the native ligand into the apo protein’s active site. The pose of the docked native ligand from the molecular docking was superimposed on its pose from the crystal structure, as shown in Fig 5, resulting in a heavy atom RMSD of 0.397 Å (< 2 Å) [68]. This confirmed the parameters, algorithm, and ligand poses, justifying the numerical method adapted with the capability of reproducing the natural process.

Fig 5. Superimposed image of native ligand (blue) from the crystal structure with the docked native ligand (magenta) from molecular docking calculations (a heavy atom RMSD = 0.397 Å).

The binding affinities and poses of ligands interacting with the FGFR1 protein were obtained from molecular docking calculations. The 44 ligands obtained by the virtual screening of 10 million compounds were subjected to molecular docking calculations against the FGFR1 receptor to assess their potential for competitive inhibition. Ligand M34 exhibited the highest binding affinity (−10.8 kcal/mol), outperforming the native ligand (−10.4 kcal/mol). Additionally, four ligands, M29 (−9.8 kcal/mol), M26 (−9.4 kcal/mol), M32 (−9.2 kcal/mol), and M28 (−9.1 kcal/mol), demonstrated binding affinities comparable to or slightly lower than that of the native ligand. The binding affinity of 44 ligands is presented in Table 3 along with their molecular formula, PubChem chemical identifier (CID), and parent ID.

Table 3. Binding affinities for 44 ligands with the FGFR1 protein, along with that of the native ligand. (Bold-faced was chosen for MDS).

Compounds		PubChem CID	Parent ID	Molecular Formula	Binding Affinities (kcal/mol)
Ligands	M34	165426608	338098121	C 36 H 41 N 5 O 6	−10.8
	M29	145940129	323961687	C 38 H 47 FN 6 O 9	−9.8
	M26	131910163	290025896	C 21 H 25 N 5 O 2	−9.4
	M32	155915988	331142671	C 19 H 22 N 4 O 3	−9.2
	M28	132423733	316221773	C 24 H 28 FN 5 O	−9.1
	M33	163358414	337753544	C24H25N5O2	−9.0
	M5	15944647	22060461	C21H18Cl2N4O2	−8.9
	M17	75410789	48548835	C22H22FN3O5	−8.9
	M12	75410713	48548681	C21H22FN3O3	−8.8
	M19	75410793	48548843	C22H21F4N3O3	−8.8
	M30	145936598	323969661	C41H48FN7O8	−8.8
	M35	NR	338535583	C26H35FN4O2	−8.8
	M36	NR	342779808	C24H25F3N4O4	−8.8
	M4	45160386	20467665	C26H32N4O5	−8.7
	M13	75410731	48548717	C22H25FN4O2	−8.6
	M16	75410782	48548821	C22H22FN3O5	−8.6
	M20	75410795	48548847	C24H25FN4O3	−8.6
	M11	75410711	48548677	C21H24FN3O2	−8.5
	M14	75410740	48548735	C20H23FN4O2	−8.5
	M8	75410670	48548595	C20H20F4N4O2	−8.4
	M7	75410654	48548563	C19H20F2N4O2	−8.3
	M9	75410675	48548605	C21H22FN3O4	−8.2
	M24	75410803	48548863	C21H23FN4O4	−8.1
	M6	42453839	25992257	C26H37N5O3S	−8.0
	M15	75410742	48548739	C18H22FN5O	−8.0
	M2	45177117	20428857	C24H35N5O4	−7.9
	M22	75410800	48548857	C22H24FN3O5S	−7.9
	M1	45168627	20392252	C29H36N4O4	−7.8
	M18	75410790	48548837	C20H21FN4O3	−7.8
	M21	75410797	48548851	C19H22FN5O3	−7.7
	M44	168463113	345870006	C20H31N7O3	−7.7
	M25	75410804	48548865	C19H22FN5O3	−7.6
	M3	45177479	20430493	C26H34N4O3	−7.5
	M31	145942272	324002419	C20H23N7O3	−7.5
	M10	75410691	48548637	C18H22FN5O2	−7.4
	M27	155880121	313775290	C20H28N4O3	−7.4
	M43	NR	345224621	C23H33N5O4S	−7.1
	M39	NR	344364342	C24H35N5O3	−7.0
	M23	75410801	48548859	C18H19FN4O3S	−6.9
	M38	NR	344358704	C26H39N5O3	−6.8
	M40	NR	344365110	C26H39N5O3	−6.6
	M42	NR	344373816	C25H36ClN5O3	−6.5
	M37	NR	344352563	C24H35N5O3	−6.1
	M41	NR	344373304	C23H32ClN5O3	−6.0
Native ligand	LWJ	66555680	–	C20H16N6O	−10.4

Note: NR = Not Recorded

The mode of interactions was studied for the complexes of M26, M28, M29, M32, and M34 with the protein. The 3D representations of the complexes, as shown in Fig 6, demonstrated that all top five ligands were bound at the catalytic site of the protein, suggesting the competitive inhibitors.

Fig 6. 3D representations (left) of docked ligands at the catalytic pocket of the protein in complexes and the molecular structures (right) of the ligands.

To better understand the details at the molecular level, the bonding interactions between the top five docked ligands and key amino acid residues, along with the distances, were studied (Table 4) and the interaction profiles are presented in Fig 7. The interaction analysis revealed several key hydrophobic interactions between the ligands and the protein’s amino acid residues, along with hydrogen bonds.

Table 4. Types of interactions between the top five docked ligands and key amino acid residues in the protein-ligand complexes.

Complexes	Types of interactions	Amino acid residues (Distance, Å)
Protein-M26 complex	Hydrogen bond	Asp184 (3.51, 3.53, 3.53), Phe185 (3.93), Gly186 (4.08)
	Hydrophobic	Leu27 (3.62, 3.67, 3.78), Val35 (3.30, 3.57, 3.93), Tyr106 (3.75, 3.88), Leu173 (3.72), Phe185 (3.31)
Protein-M28 complex	Halogen bond (Fluorine)	Ile88 (3.51)
	Hydrophobic	Val35 (3.58, 3.79), Ile88 (3.58), Val104 (3.67, 3.83), Leu173 (3.71), Ala183 (3.80)
Protein-M29 complex	Hydrogen-bond	Lys57 (3.06), Asp195 (2.90)
	Pi-stacking	Phe32 (4.58)
	Hydrophobic	Ala31 (3.69), Ile194 (3.88), Lys198 (3.76), Thr200 (3.73), Pro206 (3.32, 3.87)
Protein-M32 complex	Hydrogen-bond	Asp184 (3.00, 4.03)
	Hydrophobic	Leu27 (3.52, 3.71), Val35 (3.79), Val104 (3.83), Leu173 (3.79)
Protein-M34 complex	Hydrogen bond	Leu27 (3.75)
	Hydrophobic	Pro26 (3.55), Leu27 (3.76), Val35 (3.47, 3.59), Val104 (3.58), Leu173 (3.97), Phe185 (3.60), Thr201 (3.76)
Protein-Native complex	Hydrogen-bond	Glu105 (2.80), Ala107 (2.87, 2.95)
	Hydrophobic	Leu27 (3.53, 3.92), Lys57 (3.81), Ile88 (3.78), Val104 (3.65), Leu173 (3.71), Phe185 (3.66)

Fig 7. Molecular-level interaction profiles of top docked complexes along with that of the native complex.

The molecular interaction analysis revealed that all ligand-protein complexes (M26-M34) were primarily stabilized by hydrophobic interactions, especially with residues Leu27, Val35, Val104, Leu173, and Phe185, which appeared frequently across the top complexes and the native complex, except M29-complex. Among these, residues Leu27, Leu173, and Phe185 were the most consistently conserved hydrophobic residues, underlining their importance in ligand anchoring. Notably, residue Asp184 (3–4 Å) contributed to hydrogen bonding in M26 and M32, while M26 further exhibited additional hydrogen bonds with Phe185 (3.93 Å) and Gly186 (4.08 Å), suggesting a stabilizing role of Asp184 that was not observed in the native complex.

In contrast, the native complex formed distinct hydrogen bonds with Glu105 (2.80 Å) and Ala107 (2.87, 2.98 Å), interactions not replicated by the ligand-bound complexes. Despite this, the native complex shared hydrophobic contacts with residues Leu27, Val104, Leu173, and Phe185, which were also involved in complexes of M26, M28, M32, and M34, indicating partial overlap in binding site occupation. M29-complex exhibited a unique interaction pattern, including hydrogen bonds (Lys57: 3.06 Å, Asp195: 2.90 Å), a π-stacking interaction with Phe32, and distinct hydrophobic residues, implying a different binding orientation. Additionally, M28 uniquely formed a halogen bond with residue Ile88 (3.51 Å) via a fluorine atom.

Overall, hydrophobic interactions emerged as the key driving force for ligand binding, with residues Leu27, Val104, Leu173, and Phe185 serving as conserved contributors across multiple complexes. In addition, hydrogen bond distances ranging from 2.90 to 4.03 Å indicated strong to moderate binding affinity, as shorter bond lengths generally correlate with stronger interaction [69]. While most protein-ligand complexes exhibited similar hydrophobic interaction patterns with the native complex, M29 displayed a distinct binding profile, which may influence its receptor modulation potential.

3.3. Adduct stability with time (spatial and energetic)

Understanding the spatial and energetic stability of the adduct is crucial for evaluating the inhibitory potential of ligands on the FGFR1 protein. To achieve this, MDS was performed for 200 ns. The structural and interaction stability of the protein-ligand complexes for the top five ligands was analyzed by examining various time-dependent parameters. The spontaneity and feasibility of the complex formation reactions for the top five ligands were determined in terms of changes in binding free energy. Both geometric and thermodynamic parameters are discussed in the following sections.

3.3.1. Structural stability assessment.

The binding of the ligand to the protein can induce structural alterations in both the protein and the ligand, which may affect the stability of the complex and hence the inhibition mechanism [70]. The stability of the top five ligand complexes was evaluated by analyzing various computational metrics from MDS trajectories, which provided insight into the structural integrity of the protein-ligand complexes. The metrics include the study of ligand pose (snapshot), RMS deviation of ligands and protein backbone relative to protein backbone, RMS fluctuation of the α-carbon atoms, RPDF, R_g, and SASA which are discussed next.

Dynamic insights into ligand behavior at the protein active site through MDS: Snapshots were taken at various time intervals during the MDS to examine the orientation and position of the docked ligands, providing insight into the stability of the complex’s geometry over time. Detailed images of the top four complexes at five distinct instances are presented in Fig 8.

Fig 8. Ligand’s positions and orientations within the protein’s active site across four different complexes at five distinct instances, as extracted from the MDS trajectories.

Snapshots taken at 1, 50, 100, 150, and 200 ns showed that most ligands remained at the same location but with variations in orientation, except for a few cases. For the M28-complex, the ligand exhibited distinct rotational motion starting at 1 ns, accompanied by a slight upward position shift from 100 ns till the end due to translational movement. The protein backbone displayed some motion, particularly in the α-helix and loops (on the right side) from 50 ns onward, with subtle movement in the central β-sheet. In the case of M29-complex, the ligand underwent significant positional and orientational changes at the active site, with pronounced rotation at 50 ns and minimal rotational motion thereafter. From 100 ns onward, the ligand moved slightly upward, eventually returning to its position at 50 ns by 200 ns. The protein backbone showed the movement of the right α-helix (absent before but observed after 50 ns till the end), along with motions in the left β-sheet and central loops. In the M32-complex, the ligand depicted minimal delocalization and rotation until 100 ns, followed by rotation and a slight downward shift in the position along with the noticeable motion of the α-helix and loops (150 ns). In the M34-complex, the ligand remained stable with some rotation for the first 50 ns. After this period, it slightly shifted upward along with the protein backbone, maintaining minimal rotational and translational movement until the end of the simulation. The protein backbone showed notable dynamics, including the disappearance of the left α-helix after 1 ns and upward movement of the top lying α-helix (100 ns). The loops fluctuated throughout the simulation. The M26-complex was excluded from Fig 8 as the ligand showed displacement from the orthosteric site after 150 ns, suggesting weak binding or instability that compromised complex integrity. This indicated that the binding affinity of −9.4 kcal/mol, even though better than that observed for M32 and M28 complexes, was not sufficient to retain the pose and position at the active site. This implies that MD does not necessarily provide information about the stability of the complexes.

Periodic monitoring of adducts’ dynamical behavior provided valuable insights into molecular evolution, which could be linked to specific structural descriptors. Overall, the results demonstrated that the ligand’s pose and the protein backbone’s structural integrity were nearly preserved across the top four complexes (M28, M29, M32, and M34) with minimal structural changes and no major disruptions, suggesting the stability of the complexes. On the other hand, the M26-complex was unstable due to ligand displacement. These findings can be correlated with the RMSD and RPDF curves, which will be discussed next.

RMSD of ligand and protein backbone in the complex: The stability and dynamic behavior of the protein-ligand complexes were analyzed by examining the RMSD of the ligands and protein backbones. The RMSD for both the ligands and protein backbones with respect to the protein backbone was calculated from the MDS trajectories of various complexes and is displayed in Figs 9 and 10, respectively.

Fig 9. RMSD curves of top ligands relative to protein backbone; M26-complex (green), M28-complex (blue), M29-complex (red), M32-complex (maroon), and M34-complex (magenta).

Fig 10. RMSD curves of protein backbone with respect to protein backbone in holo form along with the protein backbone in apo form (black); M26-complex (green), M28-complex (blue), M29-complex (red), M32-complex (maroon), and M34-complex (magenta).

The RMSD of ligands relative to the protein backbone (Fig 9) provides insight into the extent of the conservation of the pose over time. The RMSD profiles of M28 (blue) and M29 (red) in their respective complexes exhibited smooth trajectories, with average RMSD of 0.38 ± 0.11 nm and 0.56 ± 0.08 nm, respectively. A slight increase in fluctuation was observed for M29 after 155 ns, attributed to an orientation shift of the ligand beyond 150 ns, as illustrated in (Fig 8). In the case of M34 (magenta), the RMSD trajectory was relatively flat after 95 ns, whereas M32 (maroon) depicted a moderate RMSD curve throughout the simulation period with some fluctuation after 120 ns. The fluctuation before 95 ns in the M34 complex and after 120 ns in M32 can be corroborated with the ligand’s orientation as seen previously in snapshots (Fig 8). The average RMSD for M34 and M32 was determined to be 0.63 ± 0.16 nm and 0.49 ± 0.16 nm, respectively. Conversely, ligand M26 (green) initially displayed a stable trajectory up to 170 ns, followed by a sharp increase in RMSD, reaching approximately 2.0 nm. This trend suggested an unstable nature of the complex, which aligns with the interpretation from the snapshots.

The backbones (Fig 10) of M28 (blue) and M34 (magenta) followed similar trajectories, with average RMSD of 0.31 ± 0.07 nm and 0.34 ± 0.08 nm, respectively, closely matching that of the apo protein (black = 0.30 ± 0.05 nm). In contrast, the backbones of M29 (red) and M32 (maroon) exhibited slightly lower average RMSD of 0.26 ± 0.04 nm and 0.29 ± 0.03 nm, respectively. The observed spikes in the RMSD plots corresponded to minor backbone adjustments, as depicted in the snapshots (Fig 8). These findings suggest that the protein backbone remained largely stable across the four complexes, indicating that ligand binding had minimal impact on its overall structure compared to the apo form. Since the M26 was unstable, the protein backbone of the M26-complex was not discussed.

From the analysis of the RMSD profiles, it was found that all ligands except M26 were bound at the protein’s active site till the end, narrowing the selection of top candidates from five to four. Among the four, ligands M28 and M29 resulted in the most stable complexes, as reflected by their low and consistent RMSD. M34-complex exhibited good stability, whereas M32-complex demonstrated moderate stability. In contrast, the complex with M26 showed significant instability, with a sharp increase in RMSD after 170 ns, indicating a loss of stable binding, and therefore, protein backbone analysis was omitted. The protein maintained the sturdy geometry across the top four complexes, capable of holding the ligand at its catalytic site. Hence, four ligands, except M26, could potentially inhibit the functioning of the FGFR1 protein.

Radial pair distribution function (RPDF): The radial pair distribution function (RPDF) describes how the distance between two entities varies over time. In its reduced form [g(r)], it represents the probability of finding the ligand’s center of mass at a distance r from the protein’s center of mass [71]. Fig 11 represents the RPDF plots for different protein-ligand complexes, which have been derived from the MDS trajectories.

Fig 11. Radial pair distribution function between the center of mass of ligands ((a) M26 = green, M28 = blue, M32 = maroon; (b) M29 = red, M34 = magenta) and protein’s center of mass in various complexes retrieved from the MDS trajectories; a single sharp peak indicates the ligand’s localization and a double peak signifies occupation at two distinct positions during the MDS period.

The RPDF plot revealed distinct binding behavior of the ligands relative to the protein’s center of mass. M32 (maroon), M28 (blue), and M34 (magenta) displayed two peaks at different distances, whereas M29 (red) exhibited a single peak. For M32, two peaks with a taller one at ca. 0.9 nm and a shorter one at ca. 1.2 nm were observed. Similarly, a tall peak at ca. 1.1 nm and a short peak at ca. 1.4 nm were observed for M28. The two different peaks indicated occupancy at two distinct locations, with the longer peak suggesting a preference for a shorter ligand-protein distance. On the other hand, M34 displayed two peaks of comparable height at ca. 1.0 nm and 1.2 nm, implying that it occupied two distinct locations within the protein’s active site for most of the simulation period. The presence of these peaks supported the minor variations in ligand position and orientation within the complex, as observed in Fig 8. In contrast, the M29 (red) complex exhibited a single peak at ca. 1.4 nm, indicating the localization of the ligand’s center of mass relative to the protein’s center of mass throughout the simulation. The occurrence of RPDF maxima at ca. 1.0 nm for the top four ligands indicates that the orthosteric site remained occupied throughout most of the simulation period. These results indicate that after binding to the orthosteric pocket of the receptor protein, the ligands remained largely localized within the site, possibly inhibiting the protein’s regular function. Thus, the RPDF analysis effectively evaluated ligand stability over time, reinforcing earlier conclusions drawn from structural snapshots.

For M26 (green), a single peak was observed at ca. 1.0 nm, but the presence of further smaller broad peaks afterward indicated the delocalization, which supported the instability of the ligand, as previously noted in snapshots and RMSD analysis. Since ligand M26 exhibited instability, further geometrical parameter analysis for this ligand was not conducted.

Fluctuation of α-carbon in the protein backbone of the complex: The root mean square fluctuation (RMSF) of the α-carbon atoms was calculated from the MDS trajectory to identify the flexible and rigid regions of the FGFR1 protein after ligand binding. The RMSF curves for the five ligand-protein complexes, along with that of the apo form (Fig 12), displayed a similar nature of the plot.

Fig 12. Root mean square fluctuation of α-carbon atoms in protein backbone of five complexes (in holo form) along with that in the apo form (black); M28-complex (blue), M29-complex (red), M32-complex (maroon), and M34-complex (magenta).

The RMSF was below ca. 0.8 nm for all top four complexes, whereas it was ca. 1.0 nm for the apo form, indicating the stability of protein geometry [72]. Higher RMSF were observed at the terminal ends and within three specific loop regions around residues 45, 130, and 200. The increased flexibility in these regions can be attributed to the absence of α-helix or β-sheet structures, which typically restrict molecular motion and reduce degrees of freedom [55]. Since these flexible regions do not play a significant role in ligand interactions or disruptions, their high RMSF does not indicate structural instability. Therefore, for clarity in the plot, only residues ranging from 12 to 300 were considered, excluding the terminal ends. The RMSF profiles indicated that the α-carbon atom fluctuations exert minimal influence on the ligand’s binding affinity within the active site. Consequently, the stability of the adduct remained unaffected, which potentially may lead to the inhibition of protein activity.

Gyradius (R_g) and solvent-accessible surface area (SASA): The gyradius (R_g), derived from the MDS trajectory, was used to evaluate the protein’s compactness and backbone conformational changes. It represents the average distance of the macromolecule’s components from its central axis, which is a crucial indicator of the stability of the protein-ligand complex [73]. The R_g plot for the protein backbone of the four ligand-protein complexes (Fig 13) revealed similar stability patterns across all systems, with R_g values ranging from ca. 2.00 to 2.15 nm, indicating no significant structural expansion or contraction during the simulation period. In contrast, the M29 (red = 2.05 ± 0.01 nm) and M32 (maroon = 2.06 ± 0.02 nm) complexes displayed some fluctuations, particularly between 60 ns and 145 ns, similar to that of the apo form (black = 2.06 ± 0.02 nm) and M34 (magenta = 2.07 ± 0.02 nm). A pronounced fluctuation in the case of the M34 complex can be correlated with minor changes in ligand orientation and α-helix positioning at 100 ns (Fig 8). Overall, the R_g of the top four complexes closely matched the apo form, suggesting no significant receptor expansion or shrinkage upon ligand binding. These findings indicate that the protein maintained structural integrity even after ligand binding, suggesting that the ligands may contribute to target protein inhibition.

Fig 13. Radius of gyration curves for protein backbone in top five ligand complexes along with the apo form (black): M28-complex (blue), M29-complex (red), M32-complex (maroon), and M34-complex (magenta); extracted from MDS trajectories.

Solvent-accessible surface area (SASA) is a crucial parameter for evaluating protein-solvent interactions, as it measures the exposure of protein residues to water molecules [74]. Changes in SASA can influence the protein’s structure, dynamics, and function [75]. SASA analysis was conducted to evaluate the effect of ligand binding on the conformational behavior of the FGFR1 protein over 200 ns MD simulations (Fig 14). The SASA ranged from 155 to 185 nm² for most complexes, showing similar trends to that of the apo form (black), with an exception for the M34 complex (magenta = 172.94 ± 4.48 nm²) which exhibited slightly higher SASA due to minor surface adjustments after 80 ns. The M28 (blue = 166.00 ± 3.32 nm²), M29 (red = 166.83 ± 2.59 nm²), and M32 (maroon = 168.59 ± 3.75 nm²) complexes displayed comparable average SASA to the apo form (165.47 ± 3.13 nm²), supporting the stable surface geometry. The minimal variations observed (below 5 nm²) suggest that ligand binding did not significantly alter the protein’s hydrophobic regions or shape, ensuring consistent solvent accessibility and reinforcing the structural stability of the complexes for most of the cases, as discussed previously.

Fig 14. Solvent-accessible surface area plots for protein backbone in top five ligand complexes along with the apo form (black): M28-complex (blue), M29-complex (red), M32-complex (maroon), and M34-complex (magenta); extracted from MDS trajectories.

3.3.2. Thermodynamic stability assessment of the protein-ligand complexes.

The spontaneity and feasibility of complex formation reactions were evaluated by analyzing the binding free energy changes in the equilibrated segment of the MDS trajectory (20 ns, 200 frames) for the top four (MDS stable) ligand adducts, as outlined in Table 5. The table reflects the degree of spontaneity in the complex formation reactions from the discrete protein and ligand. The negative ΔG_BFE (ΔG_BFE < 0) signifies the spontaneity of the complex formation reaction, and a smaller value corresponds to higher stability [76].

Table 5. Binding free energy changes (kcal/mol) and their components for protein-ligand complexes extracted from the equilibrated segment of the 20 ns MDS trajectories.

Thermodynamic Components	Binding Free Energy Change (kcal/mol)
	M28-complex	M29-complex	M32-complex	M34-complex
∆Evdw	−44.43 ± 2.48	−41.17 ± 3.67	−29.49 ± 2.62	−51.58 ± 3.56
∆Eel	−18.89 ± 5.27	−23.75 ± 8.74	−9.07 ± 4.11	−21.72 ± 5.82
∆Epb	50.05 ± 5.34	52.86 ± 11.42	29.28 ± 4.70	56.89 ± 5.27
∆Enpolar	−5.00 ± 0.16	−4.75 ± 0.35	−3.47 ± 0.19	−5.46 ± 0.22
∆G BFE	−18.28 ± 4.30	−16.81 ± 7.20	−12.76 ± 2.83	−21.87 ± 3.98

Note: data are expressed as average±SD

where, ∆G_BFE = Change in Gibb’s binding free energy, ∆E_npolar = Non-polar energy change, ∆E_pb = Change in Poisson-Boltzmann solvation energy, ∆E_el = Electrostatic energy change, ∆E_vdw = van der Waals energy change

All top-ranked protein-ligand complexes exhibited negative ∆G_BFE (ranging from −21.87 to −12.76 kcal/mol), affirming the spontaneity and feasibility of the complex formation reactions. Among them, the M34-complex demonstrated the highest thermodynamic stability, with the lowest ∆G_BFE of −21.87 ± 3.98 kcal/mol. Analysis of the thermodynamic components revealed that the solvent contribution from the Poisson-Boltzmann model (pb) posed a significant destabilizing influence across all the complexes. Nevertheless, this adverse effect was effectively mitigated by substantial positive contributions from electrostatic (el), van der Waals (vdW), and non-polar (np) interactions. These findings suggest that the top four ligands exhibit a natural propensity to associate with the FGFR1 receptor, forming energetically favorable and stable complexes throughout the simulation period. Notably, M34 stood out as the most stable adduct based on its overall energy profile.

To further dissect the energetic contributions at the residue level, decomposition analysis was performed. M29 exhibited the most favorable binding energy (−13.57 ± 2.00 kcal/mol), primarily stabilized by hydrophobic residues such as Phe32, Gly33, and Val35. Minor destabilizing effects were noted from residues Asp184 and Lys57. Similarly, M28 showed strong binding affinity (−10.12 ± 1.98 kcal/mol), with stabilizing contributions from Asn111, Val35, and Leu173, while residues Lys57, Glu74, and Glu105 depicted unfavorable effects. The M34-complex displayed moderate binding energy (−9.87 ± 2.00 kcal/mol), supported by residues Val35, Val104, Asn171, Leu173, and Ala183, and opposed by interactions with Lys57 and Asp184. In contrast, M32 demonstrated the weakest binding (−6.57 ± 1.27 kcal/mol), with stabilizing hydrophobic contacts at residues Leu27, Val35, and Leu173, but significant destabilization from Asp184 and Lys57. Detailed residue-wise free energy contributions are presented in Supplementary Information S1 Table in S1 File. Corresponding bar plots and heatmaps highlighting the interactions between active site residues of FGFR and the respective ligands are shown in Supplementary Information S1-S5 Figs in S1 File, respectively. In these heatmaps, blue shades denote favorable (negative value) contributions, while red shades represent unfavorable (positive value) ones [62].

Overall, the results indicate that hydrophobic and polar residues, particularly Val35 and Leu173, played dominant roles in stabilizing the ligand-FGFR complexes. M29 emerged as the most promising ligand candidate based on per-residue energetic contributions, while M34 demonstrated the greatest thermodynamic stability based on ∆G_BFE. Collectively, these findings suggest that all top ligands formed energetically stable complexes with FGFR1, with M34 and M29 showing particularly strong and favorable interactions. These results highlight their potential as promising FGFR1-targeted inhibitors, although further experimental validation is required to confirm their therapeutic efficacy.

3.4. Protein-ligand interaction heatmaps and histograms

To understand the residue-level binding dynamics, amino acid interaction heatmaps and histograms were generated for the 200 ns (20,000 frames) simulation period and provided in Supplementary Information S6-S13 Figs in S1 File. The analyses revealed that all four ligands complexed with the FGFR1 protein predominantly exhibited stable van der Waals (vdW) interactions, with sporadic hydrogen bonding observed in some cases.

In the M28-complex, the ligand maintained strong and consistent vdW interactions with residues Leu27, Glu105, Ala107, and Ser108. Among these, Ser108 and Glu105 accounted for the highest number of contacts, followed by Ala107and Leu27, suggesting their central role in ligand stabilization. M29 showed stable vdW contacts with residues Gly33, Gln34, and Met58, while transient interactions were observed with Gly28 and Glu29 residues. A hydrogen bond with residue Asn20 emerged after frame 13,600. Residues Gly33 and Gln34 were the most frequently engaged residues in the M29-complex, with moderate contacts involving residues Met58, Gly29, Gly28, and Asn202. In the M32-complex, the ligand demonstrated stable vdW contacts with residues Ala107 and Ser108 post frame 11300, along with moderate transient interactions involving residues Leu27 and Glu29. Residue Leu27 also formed intermittent hydrogen bonds, further contributing to ligand retention. The residue Ser108 was the most frequently interacting residue in this complex, followed by Ala107 and Leu27. For the M34-complex, stable vdW interactions were primarily observed with residues Glu29, Leu27, Gly18, and Arg170. Occasional hydrogen bonding was also noted with residue Leu27. The corresponding bar plot/histogram indicated residue Glu29 as the most engaged residue, followed by Leu27, with moderate contributions from Arg170, Glu105, and Gly28 residues.

Collectively, these findings underscore the importance of stable vdW contacts, particularly with residues Glu29, Gly33, Glu105, and Ser108 in securing ligand binding and maintaining complex stability. Their frequent engagement across multiple complexes highlights their potential relevance as key anchoring sites in FGFR1-ligand interactions.

3.5. Similarity measure analysis

The chemical similarity of four top candidate compounds (as suggested by MDS) against known FGFR1 drugs was evaluated. Fig 15 shows how structural modifications influence similarities, providing insights into the potential effectiveness of new compounds using a two-color scheme to highlight conserved and divergent features [77]. Green regions represent structural elements shared with reference drugs, suggesting retention of critical pharmacophores needed for FGFR1 inhibition, while pink areas indicate novel modifications or distinct scaffolds [66].

Fig 15. Similarity analysis between known and potential inhibitors of FGFR1.

The analysis revealed that M32 exhibits strong green overlap with infirgatinib, implying it likely maintains a similar binding mode and could serve as a high-priority candidate for lead optimization. Compounds M28 and M29 displayed a mix of green and pink regions, indicating partial structural conservation with opportunities for hybrid scaffold development to fine-tune selectivity or potency. In contrast, M34 showed prominent pink regions, suggesting unique structural features that may either introduce novel binding mechanisms or require further validation to confirm target specificity. The overall results suggest that M32 is the most promising candidate due to its strong structural similarity to known FGFR1 inhibitors, while M28-M34 offer opportunities for optimization or novel scaffold exploration.

The dice similarity index calculated between the known and potential inhibitors of FGFR1 is shown in Table 6.

Table 6. Dice similarity index between the known and potential FGFR1 inhibitors.

Candidate Compounds	Erdafitinib	Pemigatinib	Infirgatinib	Futibatinib
M28	0.200000	0.260870	0.313043	0.263158
M29	0.251656	0.294872	0.282051	0.232258
M32	0.257426	0.283019	0.452830	0.266667
M34	0.243243	0.235294	0.209150	0.263158

The dice similarity index revealed that M32 has the strongest structural resemblance to infirgatinib (0.4532830), making it the most promising candidate. M28 showed moderate similarity across all inhibitors, with the highest value for infirgatinib (0.313043). M29 exhibited comparable similarity to erdafitinib (0.251656) and pemigatinib (0.294872), indicating shared structural features. Meanwhile, M34 had the lowest overall similarity, peaking with futibatinib (0.263158). Overall, these results suggest that all hit compounds are viable leads, with M32 standing out as the most promising, warranting further optimization to enhance potency.

3.6. Prediction of pIC₅₀ values

In addition to classification models, a regression model was used to estimate the pIC₅₀ values of the candidate compounds. Table 7 summarizes the performance metrics of the top 20 regression models.

Table 7. General performance of 20 different regression models.

Model	R-Squared	RMSE
LGBMRegressor	0.71	0.68
HistGradientBoostingRegressor	0.71	0.69
GammaRegressor	0.68	0.72
TweedieRegressor	0.68	0.72
BayesianRidge	0.67	0.73
RandomForestRegressor	0.67	0.73
SVR	0.66	0.75
ElasticNetCV	0.65	0.75
PoissonRegressor	0.65	0.75
NuSVR	0.65	0.75
LassoLarsCV	0.65	0.75
LassoCV	0.65	0.75
BaggingRegressor	0.64	0.76
GradientBoostingRegressor	0.63	0.78
KNeighborsRegressor	0.55	0.85
PassiveAggressiveRegressor	0.55	0.86
HuberRegressor	0.55	0.86
RidgeCV	0.52	0.88
OrthogonalMatchingPursuitCV	0.49	0.91
ExtraTreesRegressor	0.46	0.94

Light Gradient Boosting Machine (LGBM), Hist Gradient Boosting (HGB), and Gamma Regressor (GR) were combined into a voting regressor for superior performance. For LGBM, the parameters were set to n_estimators = ‘600’ and learning_rate = ‘0.05’; for HGB, max_iter = ‘400’ and learning_rate =’0.05’; for GR, max_iter = ‘500’ and alpha = ‘0.02’; all other parameters were set to their default values.

The five-fold cross-validation of individual regression models and the voting regressor is shown in Table 8.

Table 8. Five-fold cross-validation of individual regressors and voting regressor.

Model	R-Squared	MAE	RMSE
LGBM	0.76 ± 0.03	0.47 ± 0.02	0.65 ± 0.03
HGB	0.76 ± 0.03	0.47 ± 0.02	0.65 ± 0.03
GR	0.74 ± 0.04	0.50 ± 0.02	0.66 ± 0.04
VR	0.77 ± 0.03	0.47 ± 0.02	0.63 ± 0.04

Experimental versus predicted pIC₅₀ values are shown in Fig 16, while Table 9 compares these values for known and potential inhibitors.

Fig 16. Experimental versus predicted pIC₅₀ for training and testing set.

Table 9. Predicted and experimental pIC₅₀ values for known inhibitors and potential inhibitors of FGFR1.

Compound	Experimental pIC50	Predicted pIC50
Erdafitinib	8.92	8.84
Pemigatinib	9.40	9.16
Infigratinib	9.04	8.16
Futibatinib	8.41	8.21
M28	–	7.07
M29	–	7.47
M32	–	7.29
M34	–	7.17

The predicted pIC₅₀ values of the hit candidate compounds (7.07–7.47), while lower than those of FDA-approved selective FGFR1 inhibitors, demonstrate promising lead-like potency. These values suggest that the compounds possess potential as FGFR1 inhibitors and could benefit from further structural optimization. Collectively, the findings support their candidacy for continued validation and development in FGFR1-targeted drug discovery.

Overall, the study identified four hit drug candidates as FGFR1 inhibitors by screening a large dataset. The combination of machine learning-based virtual screening with an in silico method has been found to accelerate the preliminary drug discovery process, enabling efficient analysis of extensive datasets within a short period despite the resource constraints. This integration has not only improved accuracy but also ensured the reliable identification of high-potential drug candidates for FGFR1 inhibition, streamlining the selection for further studies.

4. Conclusion

The study demonstrates the effectiveness of combining AI-guided screening with molecular docking and dynamics simulations in identifying structurally stable and energetically favorable FGFR1 inhibitors. Four hit molecules were identified from a pool of 10 million molecules using this approach. The added insights from per-residue energy decomposition and long-term interaction histogram profiling offered valuable mechanistic insights into ligand-residue interactions, reinforcing the reliability of the identified candidates. These computational results offer a cost effective and quick foundation for further investigation. Future work will focus on pharmacophore modelling, structural optimization of hit compounds and their biological evaluation through in vitro and in vivo studies to support preclinical development.

Supporting information

S1 File. Supplementary data.

This file contains additional data supporting the findings of this study.

(DOCX)

Data Availability

All relevant data are within the paper and its Supporting Information files. Further raw data can be made available on request.

Funding Statement

The author(s) received no specific funding for this work.