Computational Workflow for Chemical Compound Analysis: From Structure Generation to Molecular Docking

Abstract

Drug discovery is a complex and expensive process in which only a small proportion of candidate molecules reach clinical approval. Computational methods, particularly computer-aided drug design (CADD), have become fundamental to accelerate and optimize early stages of discovery by integrating chemical, biological, and pharmacokinetic information into predictive models. This review outlines a complete computational workflow for chemical compound analysis, covering molecular structure generation, database selection, evaluation of absorption, distribution, metabolism, excretion and toxicity (ADMET), target prediction, and molecular docking. It focuses on freely accessible and web-based tools that enable reproducible, cost-effective, and scalable in silico studies. Key platforms such as PubChem, ChEMBL, RDKit, SwissADME, TargetNet, and SwissDock are highlighted as examples of how different resources can be integrated to support rational compound design and prioritization. The article also discusses essential methodological principles, data curation strategies, and common limitations in virtual screening and docking analyses. Finally, it explores future directions in computational drug discovery, including the incorporation of artificial intelligence, multi-omics integration, and quantum simulations, to enhance predictive accuracy and translational relevance.

1. Introduction

Drug discovery and development are highly complex, costly, and time-consuming, with only a small fraction of candidates reaching the market [1]. Between 2009 and 2018, the U.S. Food and Drug Administration (FDA) authorized 355 new drugs; after accounting for failed trials, the median capitalized investment per drug was estimated at USD 985.3 million, with a mean of USD 1335.9 million [2,3]. To mitigate these challenges, computer-aided drug design (CADD) and artificial intelligence (AI) have become central to modern discovery, complementing structure-based and ligand-based approaches [4].

CADD provides a framework for storing, managing, and modeling chemical entities, supporting hit identification, lead optimization, ADMET profiling, and early safety evaluation [4,5]. A key advantage of computer-aided drug design (CADD) resides in its capacity to efficiently screen large chemical libraries, thereby reducing experimental burden as well as overall time and cost. Its predictive capabilities facilitate the early prioritization of promising lead candidates, minimizing resources allocated to compounds with limited therapeutic potential. Beyond these operational advantages, CADD affords valuable mechanistic insight into drug–receptor interactions, enabling a rational interpretation of binding modes and affinity determinants and underscoring its central role in contemporary drug discovery workflows [6,7].

Between 2021 and 2023, numerous CADD-driven studies reported the successful identification of bioactive compounds across a broad range of therapeutic targets, including inhibitors of promiscuous ABC transporters [8]. In prostate cancer research, VPC-17821 and VPC-17160 were discovered as potential treatments through targeting of the androgen receptor and DNA binding domain dimerization [9], while VPC-70619 was reported to target N-Myc–Max signaling [10]. Additional studies described the discovery of an inhibitor of MyD88–Toll/interleukin-1 receptor domain [11] and compound 64, an inhibitor of Bruton’s tyrosine kinase [12]. Further examples include XST-119, a novel inhibitor of FOXM1 [13], and Zinc-09 targeting serum and glucocorticoid regulated kinase 3 (SGK3) [14], together with the identification of a novel class of D-amino acid oxidase inhibitors for schizophrenia [15]. CAAD-based investigations also enabled the discovery of a cannabinoid antagonist [16], 5-HT2A receptor agonist [17], and ROCK1 kinase inhibitors [18]. Beyond human targets, computational approaches supported the identification of 4,5-pin2bpy (L4) and inhibitor of the Staphylococcus aureus MurG enzyme [19], as well as CDMS-01 targeting Trypanosoma cruzi sirtuin 2 enzyme for Chagas disease [20]. Finally, drug repositioning efforts led to the identification of nifuroxazide (NFZ) as a potent inhibitor of E-26 transformation-specific (ETS)-related gene (ERG) in prostate cancer [21]. These examples highlight the versatility and translational relevance of CADD-based pipelines. Importantly, CADD has also enabled the identification and repositioning of FDA-approved drugs, with deep learning-based approaches predicting antiviral activity against SARS-CoV-2 for agents such as atazanavir, remdesivir, efavirenz, ritonavir, and dolutegravir, underscoring the clinical impact of CADD-assisted drug discovery [22].

Key methodologies of CADD include virtual screening, molecular docking, quantitative structure–activity relationship (QSAR) modeling, and de novo design, which together enable prediction of drug–target interactions and optimization of bioactive compounds (4–6). These strategies are generally divided into structure-based drug design (SBDD), which relies on data from X-ray crystallography, nuclear magnetic resonance, or homology modeling [23].

The exponential growth of genomic data, high-throughput screening results, and curated databases such as ChemBank, DrugBank, and ChemSpider [24] has further strengthened the role of cheminformatics and bioinformatics. This vast data growth facilitates both SBDD (by providing more target protein structures) and ligand-based design (by offering more bioactivity data for machine learning and QSAR models). These fields support property analysis, pharmacokinetic and toxicity prediction, drug–target interaction modeling, and the exploration of biological pathways and genetic variants [25,26]. Together, they provide the foundation for integrated workflows that link chemical structure generation, predictive modeling, and molecular docking.

This review proposes a computational workflow for drug discovery, spanning from linear molecular representations and three-dimensional structure generation to molecular docking. The manuscript is organized to first discuss chemical databases and virtual libraries for compound selection, followed by molecular structure representation and input formats, ADMET property evaluation, target prediction strategies, and molecular docking methodologies. Emphasis is placed on open-access and web-based platforms, with the aim of illustrating how representative tools can be integrated into practical pipelines that support rational compound design and assist researchers during early-stage drug discovery.

2. Chemical Databases and Libraries in Virtual Screening

Molecules can be regarded as the fundamental units of a chemical language that encodes atomic composition, bonding patterns, charges, physicochemical attributes, and biological activities. Mastery of this language is essential for de novo drug design, where the objective is to generate chemically valid molecules with favorable pharmacological properties [27].

Historically, lead identification relied on high-throughput screening of large chemical libraries against biological targets [28]. Although this strategy proved effective, its high cost and limited scalability have prompted the adoption of computational alternatives. Virtual screening has since become a cornerstone of computer-aided drug discovery, broadly categorized into structure-based and ligand-based approaches. Both strategies rely fundamentally on chemical databases and virtual libraries, which provide the structural and physicochemical data required to identify and prioritize bioactive compounds [29].

Initial applications of virtual screening demonstrated its capacity to identify ligands and predict receptor-bound conformations with notable accuracy [28]. Since then, the rapid expansion of compound collections has transformed the field: commercial vendors now provide “make-on-demand” services and virtual catalogs of up to 20 million readily synthesizable compounds [30,31,32]. To support these strategies, chemical libraries are generally classified into three groups: general, natural product-based, and specialized resources (Figure 1). This classification provides a conceptual framework for understanding how different resources contribute to complementary aspects of drug discovery pipelines.

A wide array of open-access repositories is available, covering general, natural, and specialized compound collections. Here we highlight just a few representative examples. Among general databases, PubChem https://pubchem.ncbi.nlm.nih.gov (accessed on 17 September 2025) provides access to over 200 million compounds with detailed structural and physicochemical information [33], while ChEMBL https://www.ebi.ac.uk/chembl/ (accessed on 19 November 2025) adds pharmacological depth with around 2.5 million annotated bioactive molecules [34,35]. The ZINC20/22 platforms https://cartblanche.docking.org (accessed on 19 September 2025), widely used in virtual screening, host billions of purchasable compounds [36,37]. DrugBank, although smaller (~500,000 compounds), provides curated drug–target information of high value for pharmaceutical research [38]. Complementing these, natural product libraries such as COCONUT 2.0 https://coconut.naturalproducts.net (accessed on 21 September 2025) integrate structural and provenance metadata [39,40].

Taken together, chemical databases and virtual libraries provide the essential foundation for computational drug discovery. However, their utility depends on data quality and the need for rigorous curation (e.g., standardization, deduplication, and bioactivity cleaning) prior to processing. The next step in the workflow, therefore, focuses on chemical structure generation and the use of linear input formats, which transform molecular representations into standardized, machine-readable data suitable for computational analysis.

3. Chemical Structure Generation and Linear Input Formats

Before virtual screening can be performed, chemical structures must be expressed in machine-readable formats. As shown in Figure 2, these include two-dimensional depictions, three-dimensional encodings, linear notations, and connection table-based files. Two-dimensional and three-dimensional formats capture connectivity and spatial coordinates, providing the basis for computational analyses.

Linear notations, among which SMILES is the most widely adopted, encode molecular graphs as alphanumeric strings that are compact, efficient, and suitable for cheminformatics, ADMET prediction, and machine learning [41,42,43]. SMILES provides a concise representation of molecular structures using ASCII characters [44,45,46,47]. Because atom ordering during graph traversal may vary, a single compound can be represented by multiple valid strings. These variants, known as enumerated or randomized SMILES, are generated by selecting different starting atoms, expanding the representation space available for predictive modeling and machine learning [41,48,49,50].

SMILES encodes chemical structures according to fundamental rules that ensure accuracy and prevent misinterpretation [51,52]. To support their use, several platforms provide tools for generating SMILES strings from molecular design. Marvin JS Editor 17.21.0 by ChemAxon (Budapest, Hungary) offers comprehensive capabilities under a commercial license, while free web-based alternatives like the Chemical Identifier Resolver by NCI/CADD Group [53] provide practical options. Furthermore, the open-source toolkit RDKit is foundational, enabling high-throughput programmatic processing of structures (e.g., cleaning, canonicalization, and representation generation) essential in machine learning pipelines. Collectively, these tools integrate structure representation into computational workflows, thereby bridging chemical data into in silico analyses.

4. ADMET Property Evaluation

Many drugs development failures stem from poor pharmacokinetics or toxicity, often detected only in costly late-stage trials. To reduce these setbacks, evaluation of absorption, distribution, metabolism, excretion and toxicity (ADMET) has therefore become crucial to rational drug design [54]. Computational methods provide a practical alternative for integrating ADMET evaluation into early discovery workflows, enabling rapid prediction of pharmacokinetic and toxicity profiles, complementing docking and QSAR analyses [55,56].

Beyond estimating ADMET profiling, models also provide insights into bioavailability and potential safety liabilities [57]. Importantly, protein-based simulations allow the study of compound interactions with critical ADMET-related proteins such as cytochrome P450 isoenzymes, the hERG potassium channel, and P glycoprotein [58,59,60]. By combining predictive and structural data, these methods enhance rational compound selection and optimization.

Current ADMET prediction strategies can be broadly divided into two complementary approaches [54]: structure-based methods using docking/dynamics [61,62], and ligand-based methods using QSAR models derived from chemical and biological datasets [63,64].

Building on these approaches, numerous open-access platforms have been developed to predict ADMET properties and assist in compound prioritization. Among the most frequently used are SwissADME https://www.swissadme.ch (accessed on 25 September 2025), which provides predictions of physicochemical parameters, pharmacokinetics, and drug-likeness, making it a valuable first step in compound prioritization [65]; pkCSM https://biosig.lab.uq.edu.au/pkcsm/ (accessed on 25 September 2025), which uses graph-based signatures to encode molecular structure to train predictive models [66]; and admetSAR https://lmmd.ecust.edu.cn/admetsar2 (accessed on 27 September 2025), one of the largest curated resources for large-scale virtual screening [67,68]. Additional resources include ADMETlab (versions 2.0 and 3.0) https://admet-mesh.scbdd.com (accessed on 27 September 2025) [69,70] and the machine learning-based platform ADMET-AI https://admet.ai.greenstonebio.com (accessed on 28 September 2025) [71]. Furthermore, specialized platforms have been designed for specific ADMET parameters, such as MetaPred for cytochrome p450 isoform prediction (https://webs.iiitd.edu.in/oscadd/metapred/submit.php, accessed on 28 September 2025) [72] and PASS for biological activity prediction https://way2drug.com/dr/ (accessed on 28 September 2025) [73].

In addition to pharmacokinetic and toxicity prediction, many of the platforms discussed also incorporate drug likeness evaluation, providing an early filter for prioritizing viable candidates [74]. Among the most widely applied filters are Lipinski’s Rule of Five [75], Ghose [76], Veber [77], Muegge [78,79], and Egan [80,81].

However, it is crucial to recognize that these rules are primarily count-based heuristic filters that fail to capture molecular topology or specific target interactivity. Their use in early filtering must be complemented with QSAR analysis and structural evaluations, particularly when screening unconventional candidates like natural products or molecules designed for specialized barriers (e.g., the blood–brain barrier).

Complementing ADMET evaluations, specialized web platforms focus on toxicity prediction. Among freely accessible tools, ProTox 3.0 https://tox.charite.de/protox3/index.php (accessed on 29 September 2025) integrates molecular similarity, fragment analysis, and machine learning [82], while Toxtree applies decision tree algorithms https://apps.ideaconsult.net/data/ui/toxtree (accessed on 1 October 2025) [83]. Beyond open-access solutions, commercial platforms such as eTox-Drug Toxicity Prediction on through the Neurosnap platform (Neurosnap Inc. Computational Biology Platform for Research. Wilmington, DE, USA, 2022) https://neurosnap.ai/ (accessed on 1 October 2025) and MultiCASE https://multicase.com (accessed on 1 October 2025) [84] offer in silico tools to support comprehensive evaluation of compound toxicity and facilitate risk assessment.

Taken together, these platforms, both free and commercial, provide complementary resources that substantially enhance the predictive assessment of compound toxicity.

5. Target Prediction and Receptor Selection

An essential step in compound development is the identification of molecular targets with which a compound may interact [85]. In silico strategies addressing this challenge include comparative genomics [86,87], network-based approaches [87,88,89], and target fishing [90]. Target fishing can be executed through target-centric methods, such as machine learning and QSAR modeling [90,91], or ligand-centric methods, based on molecular similarity assessment [90,92].

Building on these strategies, comparative genomics integrates large-scale genomic data with computational tools to identify essential genes whose inhibition may compromise cellular viability, providing insights for both infectious diseases and complex conditions like cancer [86,87]. Meanwhile, network-based approaches enable the systematic exploration of molecular interactions within biological systems, supporting biomarker discovery, disease diagnosis, and target identification [87,88,89]. Complementarily, target fishing frameworks combine both paradigms—target- and ligand-centric—often enhanced by machine learning algorithms such as Random Forest or Naive Bayes, trained on extensive bioactivity datasets to estimate compound–target interaction probabilities [90,91].

Target prediction is made broadly accessible through web platforms that typically process SMILES input structures, which are converted into molecular fingerprints for prediction [93]. Fingerprints are central to cheminformatics, enabling virtual screening and exploration of chemical space, with substructure-based descriptors often providing the most accurate results. A widely used example is the MACCS fingerprint, where each bit denotes the presence or absence of predefined substructural features [94].

These types of fingerprints are widely employed across web-based prediction platforms. For instance, TargetNet http://targetnet.scbdd.com (accessed on 2 October 2025) integrates seven fingerprinting schemes to evaluate binding potential by generating high-quality QSAR models for each target [95]. SuperPred https://prediction.charite.de/subpages/target_prediction.php (accessed on 2 October 2025) combines target prediction with the ATC classification system [96]. Ligand-based strategies also underpin widely used tools such as SwissTargetPredictions http://www.swisstargetprediction.ch (accessed on 2 October 2025), which predicts the most probable protein targets of bioactive molecules in humans [97], and the Similarity Ensemble Approach (SEA) https://sea.bkslab.org (accessed on 2 October 2025), which infers protein function from chemical similarity ligands [98].

6. Molecular Docking: Evaluating Ligand–Protein Binding Affinity

Molecular docking has emerged as a central tool in structure-guided drug design [99]. Molecular docking is applied to predict ligand conformation and compatibility within active sites. The accuracy of these predictions depends on the docking algorithm, which incorporates spatial orientation, ligand flexibility, and scoring functions [100]. According to Priya et al. (2014), molecular docking can be grouped into two main categories: protein–protein, which predicts the interface between interacting macromolecules, often without requiring prior experimental information, and protein–ligand, the most widely used approach that models how a small molecule fits into an active site of a protein, where it may function as either activator or inhibitor for receptor activity [101].

Conceptually, docking involves two steps: first, the algorithm generates alternative ligand poses; second, a scoring function evaluates and ranks these poses according to noncovalent interactions, providing an estimate of binding affinity [102,103]. Once these principles are established, the general workflow of molecular docking becomes clearer (Figure 3). It involves (i) ligand and receptor preparation, (ii) active site identification, (iii) docking simulation to generate multiple poses, and (iv) ranking and selection of the most favorable binding conformation [104,105].

The reliability of docking results is intrinsically limited by the scoring functions used. These functions are typically empirical and simplify key thermodynamic terms, making them generally more effective at predicting the correct pose (orientation) than at predicting the absolute affinity (Ki or ΔG values).

Major limitations include their poor handling of solvation effects (the energetic cost of displacing water molecules) and the conformational entropy of the system. Therefore, docking results must be complemented by more rigorous computational methods, such as molecular dynamics (MD) and MM/GBSA or MM/PBSA calculations, for a more accurate estimation of binding affinity and complex stability.

Today, researchers can choose from a wide array of docking tools, each tailored to specific applications and presenting distinct strengths and limitations. Although many powerful options are available for structure-based approaches, no single program is universally optimal for all molecular systems. For this reason, numerous molecular docking programs have been developed, ranging from commercial packages to freely accessible tools; the most representative docking software that exemplifies methodological diversity and relevant for different computational applications are shown in

Table 1. Examples of molecular docking tools mostly used in drug design studies.

Software	License Type	Main Features	Recommended Use	Cite
Autodock	Open-source	Binding orientation and affinity of small molecules to 3D receptors	Virtual screening, structure-based design, recommended for academic projects	[106]
Discovery Studio	Commercial and Academic	Docking with ligand conformational search (Monte Carlo) and LigandFit preparation	Comprehensive modeling, docking workflows, industry-level projects	[107]
DOCK	Commercial and Academic	Geometric matching to place ligands/fragments in binding sites; includes solvent effects	Academic research, fragment docking, solvent-inclusive studies	[108]
Dockey	Open-source	Graphical interface integrating preparation, parallel docking, interaction detection, and visualization	Comprehensive and user-friendly docking workflows	[109]
DOT	Open-source	Docking of macromolecule interactions; predicts binding via electrostatic and van der Waals energies	Protein–protein and large complex docking; biologically relevant models	[110]
FitDock	Academic	Improves protein–ligand docking by using similar co-crystal structures; enhances sampling and scoring	Structure-based drug design with accuracy improvement	[111]
FlexX	Commercial and Academic	Uses incremental construction: docks ligand fragments	Fast docking of fragment-based ligands in diverse binding pockets	[112]
Glide	Commercial	Ligand–receptor docking, supports virtual screening and binding mode prediction	High-precision docking, virtual screening in pharma research	[113]
GOLD	Commercial and Academic	Genetic algorithm for ligand binding predictions; flexible across diverse protein targets	Reliable docking in drug discovery, protein–ligand interaction studies	[114]
GRAMM	Commercial	Explores intermolecular energy landscape; predicts stable and transient protein–protein docking poses	Protein–protein interaction modeling and complex prediction	[115]
iGEMDOCK	Open-source	Identifies pharmacological interactions by virtual screening	Ligand screening and pharmacological interaction prediction	[116]
LeDock	Academic	Fast and accurate flexible docking of small molecules;	High-throughput virtual screening and pose prediction	[117]
LigandFit	Commercial	Shape-based docking using cavity detection, Monte Carlo conformational search, and grid-based scoring	Protein–ligand docking, pose prediction, and high-throughput virtual screening	[118]
MetalDock	Open-source	Specialized in metal–organic docking; supports multiple metal types and automates workflow	Protein, DNA, and biomolecule docking with metal complexes	[119]
MOE	Commercial	Integrated modeling platform: docking, QSAR, pharmacophore design, homology modeling	Comprehensive drug discovery workflows, method development, academic evaluation	[120]
Molegro Virtual Docker	Commercial	Docking platform with novel optimization algorithm and user-friendly interface	Protein–ligand docking, virtual screening with high usability	[121]
MSU SLIDE	Commercial and Academic	Manages large binding-site templates with multi-stage indexing; ranks ligands by steric complementarity	Efficient virtual screening of large libraries with binding-site template matching	[122]
MzDOCK	Open-source	GUI-based docking tool; simplifies workflows and improves reproducibility	User-friendly option for beginners and teaching	[123]
Qsite	Commercial	QM/MM multi-scale tool combining quantum and molecular mechanics to predict configurations, energetics, and electronic structures	Accurate modeling of reactive systems, catalytic sites, and mechanistic studies	[124]
rDOCK	Open-source	Docking of small molecules to proteins and nucleic acids	Virtual screening, binding mode prediction, protein and nucleic acid targets	[125]
SuperStar	Commercial	Generates protein interaction maps from crystallographic data; predicts “hot-spots” for favorable interactions	Binding site analysis, hot-spot prediction, and molecular design support	[126]

and Table 2. Despite methodological variations, the overall goal remains the same: to generate plausible ligand conformations and rank them according to predicted affinity.

To support the appropriate selection of docking software, performance should be evaluated using clear and reproducible criteria, particularly docking accuracy, computational efficiency, and screening enrichment. Docking accuracy is commonly assessed based on the rank-one solution, which reflects whether the top-ranked pose accurately reproduces the experimental binding mode. Using this criterion, comparative benchmarks have reported acceptable docking accuracies of approximately 0.47 for AutoDock, 0.31 for DOCK, 0.35 for FlexX, and 0.52 for GOLD, indicating that widely used tools can differ substantially in pose prediction performance. In terms of computational efficiency, FlexX has been consistently reported as one of the fastest approaches, whereas AutoDock typically exhibits longer execution times, a factor that becomes particularly relevant in large-scale virtual screening campaigns. Moreover, screening studies using DOCK and FlexX have demonstrated notable differences in ligand enrichment, emphasizing that the recovery of potentially active compounds within the top-ranked fraction of a library depends strongly on the selected platform [127].

Additional comparative studies using diverse protein–ligand datasets further highlight variability in robustness and pose prediction accuracy across docking platforms. A comparative evaluation of docking programs highlights substantial differences in robustness and pose prediction accuracy across platforms. In a study involving 195 diverse protein–ligand complexes, Glide (v4.5), GOLD (v3.2), and LigandFit (v2.3) were assessed for their ability to reproduce crystallographic binding orientations. GOLD successfully processed the complete dataset, whereas Glide and LigandFit failed to process 25 and 8 complexes, respectively, indicating differences in robustness when handling structurally diverse systems. Regarding docking accuracy, approximately 40% of the docking solutions generated by these programs achieved an RMSD below 1.0 Å, while Glide and GOLD exhibited higher success rates of approximately 60% on this highly diverse test set [128].

Recent comparative evaluations of multiple docking programs reported success rates of approximately 40–60% for top-scored poses and 60–80% for best poses, with RMSD values generally below 2 Å relative to native conformations. Among academic tools, AutoDock Vina (49.0%), AutoDock (PSO) (47.3%), UCSF DOCK (44.0%), and AutoDock (LGA) (37.4%) showed variable performance in predicting top-ranked poses. In parallel, commercial platforms such as GOLD (59.8%) and Glide, operating in Extra Precision (57.8%) and Standard Precision (53.8%) modes, demonstrated comparable or slightly higher success rates, while LigandFit (46.1%) showed moderate performance. Overall, the averaged success rates of commercial and academic docking programs were similar (54.0% vs. 47.4% for top-scored poses and 67.8% vs. 68.4% for best poses), indicating that both categories of algorithms are capable of adequately sampling conformational space and generating reliable docking solutions across diverse protein–ligand complexes [129].

In addition to traditional docking programs, a growing number of free web-based servers now provide accessible platforms for molecular docking (Table 2). These resources integrate streamlined workflows with user-friendly interfaces, enabling efficient docking simulations without the need for local installations or advanced computational infrastructure. By lowering technical barriers, web servers have become valuable tools for both specialized research and educational purposes, facilitating rapid prototyping, virtual screening, and exploratory studies in structure-based drug discovery.

Table 2. Online examples of web servers for molecular docking tools.

Web Server	Tools/Features	Suggested Use Cases	URL	Cite
PatchDock	Geometry-based docking; detects shape complementarity with minimal clashes and large interface areas	Protein–protein, protein–ligand, and protein–DNA docking	https://bioinfo3d.cs.tau.ac.il/PatchDock/ (accessed on 5 October 2025)	[130]
ZDOCK	Protein–protein docking using 3D FFT; statistical potentials; improved speed (>8×) and reduced memory	Large scale protein–protein docking, flexible molecule docking	https://zdock.wenglab.org (accessed on 5 October 2025)	[131]
CB-DOCK2	Blind protein–ligand docking; cavity detection + AutoDock Vina + template guidance (FitDock)	Binding site prediction and docking for homologous proteins	https://cadd.labshare.cn/cb-dock2/index.php (accessed on 5 October 2025)	[111,132]
SwissDock	Docking with AutoDock Vina (fast) and Attracting Cavities (accurate); flexible input formats; web access	Small-molecule docking, virtual screening, quick tests, covalent docking	https://www.swissdock.ch (accessed on 7 October 2025)	[133]
HADDOCK	Data-driven docking using experimental or biophysical restraints (AIRs)	Protein–protein docking guided by NMR or mutagenesis data	https://rascar.science.uu.nl/haddock2.4/ (accessed on 7 October 2025)	[134]
Webina 1	In-browser AutoDock Vina via WebAssembly; includes PDBQT Convert	Quick ligand–receptor docking; teaching and rapid tests	https://durrantlab.pitt.edu/webina/ (accessed on 7 October 2025)	[135]
ProteinsPlus	Tools for structure check (EDIA), hydrogen placement (Protoss), conformations (SIENA), interaction diagrams (PoseView), interface classification (HyPPI), pocket detection and druggability (DoGSiteScorer)	Preprocessing, binding site analysis, early-stage modeling	https://proteins.plus (accessed on 7 October 2025)	[136]
HPEPDOCK 2.0	Blind protein–peptide docking; hierarchical algorithm with MODPEP ensembles; global and local docking	Protein–peptide interaction modeling; global and local docking	http://huanglab.phys.hust.edu.cn/hpepdock/ (accessed on 11 October 2025)	[137]
HawkDock	Deep learning flexible docking (GeoDock); binding affinity (VD-MM/GBSA); mutation analysis	Protein–protein docking, affinity prediction, mutation impact studies	https://cadd.zju.edu.cn/hawkdock/ (accessed on 11 October 2025)	[138,139]
EDock	Blind docking with replica exchange Monte Carlo; integrates I-TASSER and COACH	Docking on low-resolution protein models; binding site prediction	https://zhanggroup.org/EDock/ (accessed on 11 October 2025)	[140]

Table 2. Online examples of web servers for molecular docking tools.

Web Server	Tools/Features	Suggested Use Cases	URL	Cite
PatchDock	Geometry-based docking; detects shape complementarity with minimal clashes and large interface areas	Protein–protein, protein–ligand, and protein–DNA docking	https://bioinfo3d.cs.tau.ac.il/PatchDock/ (accessed on 5 October 2025)	[130]
ZDOCK	Protein–protein docking using 3D FFT; statistical potentials; improved speed (>8×) and reduced memory	Large scale protein–protein docking, flexible molecule docking	https://zdock.wenglab.org (accessed on 5 October 2025)	[131]
CB-DOCK2	Blind protein–ligand docking; cavity detection + AutoDock Vina + template guidance (FitDock)	Binding site prediction and docking for homologous proteins	https://cadd.labshare.cn/cb-dock2/index.php (accessed on 5 October 2025)	[111,132]
SwissDock	Docking with AutoDock Vina (fast) and Attracting Cavities (accurate); flexible input formats; web access	Small-molecule docking, virtual screening, quick tests, covalent docking	https://www.swissdock.ch (accessed on 7 October 2025)	[133]
HADDOCK	Data-driven docking using experimental or biophysical restraints (AIRs)	Protein–protein docking guided by NMR or mutagenesis data	https://rascar.science.uu.nl/haddock2.4/ (accessed on 7 October 2025)	[134]
Webina 1	In-browser AutoDock Vina via WebAssembly; includes PDBQT Convert	Quick ligand–receptor docking; teaching and rapid tests	https://durrantlab.pitt.edu/webina/ (accessed on 7 October 2025)	[135]
ProteinsPlus	Tools for structure check (EDIA), hydrogen placement (Protoss), conformations (SIENA), interaction diagrams (PoseView), interface classification (HyPPI), pocket detection and druggability (DoGSiteScorer)	Preprocessing, binding site analysis, early-stage modeling	https://proteins.plus (accessed on 7 October 2025)	[136]
HPEPDOCK 2.0	Blind protein–peptide docking; hierarchical algorithm with MODPEP ensembles; global and local docking	Protein–peptide interaction modeling; global and local docking	http://huanglab.phys.hust.edu.cn/hpepdock/ (accessed on 11 October 2025)	[137]
HawkDock	Deep learning flexible docking (GeoDock); binding affinity (VD-MM/GBSA); mutation analysis	Protein–protein docking, affinity prediction, mutation impact studies	https://cadd.zju.edu.cn/hawkdock/ (accessed on 11 October 2025)	[138,139]
EDock	Blind docking with replica exchange Monte Carlo; integrates I-TASSER and COACH	Docking on low-resolution protein models; binding site prediction	https://zhanggroup.org/EDock/ (accessed on 11 October 2025)	[140]

In summary, molecular docking has consolidated its role as a central component of modern drug discovery workflows. Its successful application in both urgent contexts, such as COVID-19 [141,142], and long-term research fields like oncology highlights its versatility [143,144]. Nevertheless, docking outcomes are inherently dependent on scoring functions and structural quality, and must be complemented with experimental validation to ensure biological accuracy.

Despite the central role of molecular docking and other in silico approaches in modern drug discovery, multiple factors continue to limit their translational success. In addition to methodological constraints inherent to docking, such as simplified scoring functions and sensitivity to structural quality, broader challenges related to data sourcing, integration, and representation substantially contribute to the high attrition of CADD-derived candidates. Redundant and inconsistent datasets, data sparsity, class imbalance, and the limited availability of well-annotated negative samples undermine predictive reliability, while feature representation strategies often struggle to capture the biological complexity of drug–target interactions [145]. Furthermore, the limited interpretability of advanced machine learning models and the incomplete integration of pharmacokinetic, toxicological, and clinical data exacerbate the gap between in silico predictions and experimental or clinical outcomes. Collectively, these limitations underscore the need for integrative, high-quality datasets, improved predictive models, and complementary experimental validation to enhance the success of CADD-driven drug discovery pipelines.

7. Integrative Workflow Summary

In this review we suggest a computational drug discovery workflow, based on the search for compound properties and target interactions. The proposed computational workflow for chemical compound analysis integrates the selection of candidate molecules, conversion of structures into linear and three-dimensional formats, ADMET evaluation, and chemical database mining, through to virtual library generation and molecular docking studies (Figure 4). This scheme synthesizes the key methodological steps discussed throughout the review, providing an integrative overview of the in silico pipeline for bioactive compound discovery and evaluation.

8. Conclusions and Future Perspectives

Computational methods have reshaped drug discovery by enabling coherent workflows that connect molecular representation, chemical databases, ADMET evaluation, target prediction, and molecular docking. These approaches streamline early-stage research, reduce costs and timelines, and minimize reliance on animal testing by filtering unsuitable candidates before experimental validation. The combination of open-access platforms, curated libraries, and predictive algorithms has made these strategies broadly accessible, offering reproducible pipelines that enhance efficiency and selectivity in drug design. Looking forward, the continued integration of emerging technologies will expand these capabilities. Generative artificial intelligence holds promise for the de novo design of novel compounds; multi-omics data can strengthen the biological context of predictions; and advances in quantum simulations may enable unprecedented accuracy in modeling molecular interactions. Together, these developments point to a future where computational pipelines not only complement but also guide experimental research, accelerating the translation of in silico predictions into clinically relevant therapies.