Machine Learning–Guided and Site-Resolved In Silico Screening for Identification of RSV F-Protein Modulators

Principal Investigator: Juan Fidel Osuna Ramos, PhD

Institutional Affiliation: Laboratory of Virology and Antiviral Research. Autonomous University of Sinaloa. Mexico

Email: osunajuanfidel.fm@uas.edu.mx ORCID: https://orcid.org/0000-0001-8280-9812

1. Overview

At present, respiratory syncytial virus (RSV) is one of the leading causes of severe lower respiratory tract disease globally, particularly in newborns, older adults, and immunocompromised people. The prefusion conformation of the RSV fusion glycoprotein (RSV F) is a biologically central and structurally tractable antiviral target, as structural studies have shown that this metastable trimer contains conserved regions directly involved in membrane fusion and viral entry, together with conformational features that are relevant to ligand binding and functional modulation [1, 2]. More recent work has further strengthened the translational relevance of this target by showing that small molecules identified through drug-repurposing approaches, including lonafarnib, can inhibit RSV through interaction with the fusion protein [5]. Collectively, these findings endorse the use of prefusion RSV F as an appropriate framework for a consistent in silico screening study aimed at methodological comparison and candidate identification.

Classical structure-based virtual screening is a conventional approach for initial hit identification, and AutoDock Vina remains one of the most prevalent docking platforms due to its computational efficiency, accessibility, and extensive utilization in reproducible screening protocols [3, 4, 12]. However, docking-based ranking frequently offers insufficient differentiation among candidate ligands, prompting the proposal of consensus-scoring frameworks as an effective method to enhance enrichment and prioritization efficacy in virtual screening initiatives [7-9]. Simultaneously, machine learning-enabled methodologies have arisen as advantageous supplements to traditional docking for hit prioritization, rescoring, and affinity prediction [6,10,13]. The proposed study will conduct a controlled head-to-head comparison between a machine learning-enabled screening workflow and a classical docking baseline, utilizing the same RSV F target, the same curated repurposing-oriented library, and the same predefined evaluation criteria. The study aims to produce a clear comparison of practical screening performance and a shortlist of structurally viable, chemically diverse RSV F candidate modulators for subsequent validation. The complete workflow will be preregistered and publicly disclosed in accordance with current standards for reproducible computational science [11,14].

Key innovations:

ML as a pre-docking filter to reduce computational load and enrich for biologically relevant compounds.
Site-resolved analysis, enabling identification of ligands targeting functionally critical regions: central cavity, interprotomer interfaces, and fusion-associated domains.

2. Introduction

Recent advancements in the prevention of RSV have significantly enhanced the public health landscape; however, they have not obviated the necessity for further therapeutic discovery. Vaccines are now available for older adults, and infant protection can be achieved through maternal immunization or long-acting monoclonal antibodies. However, these interventions are preventive rather than therapeutic, are recommended for specific populations rather than universally administered, and remain incompletely implemented in real-world practice. Post-licensure evidence indicates that uptake among older adults and maternal vaccine recipients remains well below universal coverage, even as effectiveness data continue to accumulate. Consequently, prevention alone does not resolve the clinical need for treatment of established infection. [15-18,22,23]

At the clinical level, management of RSV infection remains predominantly supportive. Current public-health guidance emphasizes hydration, symptom control, and monitoring of respiratory status and specifically notes that antiviral medication is not routinely recommended for most RSV infections. In parallel, recent antiviral-development reviews conclude that although ribavirin remains the only approved antiviral treatment, multiple direct-acting candidates are still investigational, and none has yet become a broadly adopted standard therapy for RSV disease. This gap is especially important because severe disease continues to occur in infants, older adults, and immunocompromised individuals who may not be fully protected by available preventive strategies. [19-21,24]

From a therapeutic-development perspective, this persistent gap continues to justify focused small-molecule discovery against the viral fusion machinery. The RSV fusion protein remains central to viral entry and membrane fusion and therefore continues to represent the most compelling mechanistic target for antiviral intervention. Recently, the field has moved toward an integrated state of the art in which target-based antiviral discovery, repurposing strategies, and increasingly data-driven prioritization approaches are used in combination rather than isolation. Within this context, the unresolved question is no longer whether RSV F is a tractable target, but which computational strategy is most informative for prioritizing experimentally plausible compounds under realistic budgetary and reproducibility constraints. [21,24,25]

The present study is designed to address that methodological problem while responding to the ongoing therapeutic need. Specifically, it evaluates whether machine learning–enabled prioritization provides practical advantages over a classical docking baseline when both are applied to the same RSV F target, the same curated repurposing-oriented library, and the same predefined downstream criteria. By doing so, the study links a clinically relevant unmet need with a reproducible computational benchmarking problem and aims to generate both comparative methodological evidence and a shortlist of chemically diverse candidate RSV F modulators for future validation. [21-25]

Hypotheses:
H₁: ML-based filtering reduces non-informative ligands entering docking.
H₂: ML-prioritized compounds show improved docking performance.
H₃: Site-resolved analysis identifies biologically relevant and experimentally viable candidates.

The project produces experiment-ready candidate compounds while maintaining reproducibility and efficiency.

3. Methods

Participants/Sample

Target Sample:
The research will employ a curated computational dataset comprising approximately 10,152 small-molecule compounds sourced from the FDA Unique Ingredient Identifier (UNII) database (https://open.fda.gov/data/unii/). This database comprises approved pharmaceuticals, investigational agents in advanced clinical stages, and bioactive compounds with established pharmacokinetic and safety profiles. The dataset is chosen to optimize translational potential for RSV F-protein modulation and drug repurposing initiatives.

Inclusion criteria:

Valid chemical structure: Only compounds with correctly annotated SMILES or InChIKey representations will be included to ensure accurate molecular modeling.
Drug-like properties: Compounds must satisfy at least Lipinski’s rule-of-five criteria (MW ≤ 500, LogP ≤ 5, H-bond donors ≤ 5, H-bond acceptors ≤ 10) and Veber rules for rotatable bonds (≤ 10) to guarantee oral bioavailability.
Successful conformer generation: 3D conformers must be successfully generated for each ligand to ensure suitability for ML-based and docking structural analysis.

Exclusion Criteria:

Structural errors: Compounds with incomplete, inconsistent, or ambiguous structural representations are excluded.
Failed docking preparation: Compounds that cannot be converted into docking-compatible formats (PDBQT) due to chemical complexity or incompatibility with force fields will be removed.
Redundant entries: Duplicate or highly similar compounds (e.g., tautomers or stereoisomers without distinct activity evidence) will be filtered to maintain chemical diversity.

Recruitment Strategy for Computational Dataset:

The chemical dataset will be systematically extracted from the public database FDA UNII.
A script in Jupyter Notebook Python code language will be used to fetch structures, standardize chemical representations, remove duplicates, and compute physicochemical descriptors for filtering.
Candidate molecules meeting inclusion criteria will then undergo 3D conformer generation and energy minimization to ensure compatibility with AutoDock Vina and ML screening pipelines.

Sample Size Justification:

The dataset, comprising approximately 10,152 compounds, offers extensive chemical representation of both approved and clinical-stage drugs, optimizing computational practicality and the potential for hit discovery.
Power considerations for in silico screening are theoretical: presuming a genuine hit rate of 1–5% and a machine learning-guided enrichment factor of 2–3, this sample size is adequate to identify high-performing compounds with statistically significant enhancements in predicted binding scores compared to random selection (Mann–Whitney U test, α = 0.05, power ≈ 0.8).
This dataset guarantees that the hybrid ML and docking workflow is statistically robust for identifying high-confidence candidate modulators for subsequent in vitro assessment while preserving computational efficiency.

Materials and Procedures

Dataset and Ligand Preparation

A curated dataset comprising approximately 10,152 small-molecule compounds, sourced from public databases such as FDA UNII. Only compounds possessing valid chemical structures, drug-like characteristics, and successful conformer generation will be considered. Compounds demonstrating structural inaccuracies, unsuccessful docking preparation, or numerous Lipinski/Veber violations will be excluded.

Ligand preprocessing will be performed utilizing RDKit v2023.09 in Python 3.11. SMILES strings will be standardized, salts eliminated, tautomers resolved, and three-dimensional conformers generated utilizing the ETKDG algorithm. The protonation states will be modified to align with physiological pH (7.4). The physicochemical properties, such as molecular weight, logP, hydrogen-bond donors and acceptors, and rotatable bonds, will be calculated, and the Lipinski and Veber rules will be implemented. Energy minimization of three-dimensional conformers will be conducted utilizing the MMFF94 force field, and ligands will be transformed to PDBQT format using OpenBabel v3.1.

RSV F-Protein Structure Preparation

Prefusion RSV F-protein structures (PDB: 5EA8) will be retrieved from the RCSB Protein Data Bank. Non-essential atoms, including ligands and ions, will be removed, while polar hydrogens and protonation states will be assigned using PDB2PQR v3.2 with PROPKA 3.1. Missing loops and residues will be modeled as needed. The protein structure will undergo brief energy minimization using GROMACS 2024.1 with the CHARMM36m force field.

Potential binding sites will be identified at the central cavity, interprotomer interfaces, and HRA/HRB regions near the fusion peptide. Candidate cavities will be defined using DoGSiteScorer and cross-validated with known antibody epitopes to define docking grids.

Machine Learning-Based Ligand Prioritization

A graph neural network (GNN) model, Boltz-2, will be used to predict binding probability for each ligand. Molecular graphs of ligands and protein structural features will serve as inputs. The network will comprise three message-passing layers, global pooling, and dense layers that output predicted binding probabilities. The top 10–20% of ligands ranked by predicted binding probability will be selected for subsequent docking.

Classical Docking with AutoDock Vina

All chosen ligands will be docked into each designated site utilizing AutoDock Vina v1.2 on a high-performance computing cluster. The docking parameters will specify an exhaustiveness of 8 for the initial screening, utilizing grid boxes measuring approximately 20–30 Å per side surrounding each cavity, as illustrated in Figure 1 and detailed below. Ligands will exhibit flexibility, while the receptor will maintain rigidity. For each ligand, the nine highest-ranking poses per site will be retained. PDBQT files will be transformed into PDB format for visualization with PyMOL v2.5. All docking runs designated as failed will be omitted.All docking runs designated as failed will be omitted.

Figure 1. Defined Docking Boxes for Classical Site Numbers 0, II, III, IV, and V on the RSV Prefusion F-Protein (PDB: 5EA8).

Classical Docking Box Definitions for Target Sites

During the docking stage, all ligands will be evaluated against five predefined target sites on the prefusion RSV F-protein, numbered according to established structural nomenclature. Each site will have a clearly defined grid box encompassing key residues and structural regions to ensure accurate pose sampling and site-specific interaction analysis.

Site 0
- Location: Trimer axis of the prefusion F-protein, critical for membrane fusion.
- Key residues: LYS201, ASP202, GLU204, and adjacent central core residues.
- Box dimensions: ~25 × 25 × 25 Å, centered at the geometric centroid of these residues.
- Rationale: This site is the primary fusion core and the most conserved structural feature; ligands here may stabilize the preF conformation.
Site II – Interprotomer Interface
- Location: Interface between adjacent protomers, contributing to trimer stability.
- Key residues: MET259, SER262, ILE265, GLN267.
- Box dimensions: ~22 × 22 × 22 Å.
- Rationale: Targeting this site may interfere with conformational changes required for viral fusion.
Site III – HRA/HRB Groove
- Location: Heptad repeat regions forming the six-helix bundle.
- Key residues: Residues 170–177 and 175–180.
- Box dimensions: ~20 × 20 × 20 Å.
- Rationale: Ligand binding may disrupt the coiled-coil assembly, stabilizing the prefusion state.
Site IV – Apical Surface Pocket
- Location: Shallow surface-exposed pocket distal to the central cavity.
- Key residues: PRO385, ASN386, THR387.
- Box dimensions: ~18 × 18 × 18 Å.
- Rationale: Targeting this region may modulate preF accessibility or dynamics without affecting the trimer core.
Site V – Secondary Surface Pocket
- Location: Secondary surface-exposed region near HRA/HRB termini.
- Key residues: THR423, LYS425, ILE428.
- Box dimensions: ~20 × 20 × 20 Å.
- Rationale: May allow identification of allosteric modulators influencing preF stability indirectly.

For all sites, the grid boxes will be centered on the geometric centroid of key residues, with dimensions chosen to encompass each pocket fully while avoiding unnecessary space. This setup ensures consistent docking conditions across all ligands and enables site-resolved analysis for consensus scoring and candidate selection. The top nine poses per ligand per site will be retained for interaction analysis and downstream consensus scoring.

Post-Processing and Site-Resolved Binding Analysis

Docking results and ML predictions will be merged into a single dataset including ligand ID, docking score, ML score, 3D pose, Lipinski compliance, and optional ADMET predictions. Ligands will be assigned to binding sites (central cavity, interprotomer interface, fusion-related regions, and surface pockets) based on spatial clustering and residue interactions using PLIP v2.2. Structural clustering will ensure chemical diversity in the top candidates.

Candidate Selection and Consensus Scoring

A multi-criteria consensus score will be computed for each ligand as follows:

Where the top 15–20 candidates will be selected for further refinement based on consensus scores, scaffold diversity, and site relevance.

Molecular Dynamics (MD) Refinement

MD simulations will be conducted for each of the top ligand–protein complexes using GROMACS 2024.1 or AMBER20. Protein parameters will use CHARMM36m, while ligand parameters will be assigned using CGenFF or GAFF. Systems will be solvated in TIP3P water with 10 Å padding and 0.15 M NaCl. Each system will undergo energy minimization, followed by equilibration (1 ns NVT + 1 ns NPT). Production runs of 100 ns will be performed. Stability will be assessed using RMSD and RMSF of protein and ligand, hydrogen bond occupancy, and MM/PB(GB)SA binding free energy calculations. Ligands that dissociate or show unstable interactions will be excluded from the final selection.

Reproducibility and Data Management

All preprocessing, docking, ML, and MD steps will be fully scripted and documented in Python and Jupyter notebooks. Code and data will be stored in GitHub and OSF repositories, with Docker or Conda environments for reproducibility.

Planned Analyses

The analysis pipeline will begin by merging outputs from both classical docking (AutoDock Vina) and machine learning (Boltz-2) for all curated ligands. Preprocessing will involve normalizing docking scores and ML-predicted binding probabilities, verifying structural integrity of ligand poses, and ensuring compliance with Lipinski and Veber criteria. Each ligand will be assigned to a specific binding site based on residue-level interactions and spatial clustering, distinguishing the central cavity, interprotomer interface, fusion-associated regions, and surface pockets.

The following schematic provides a formal, visually appealing, and manuscript-ready representation of the planned analyses, clearly linking preprocessing, ML prioritization, docking, site-specific analysis, consensus scoring, MD refinement, and candidate selection:

Statistical analyses will include calculation of descriptive metrics for docking affinities, ML scores, and interaction densities for each site. Comparisons between ML-prioritized ligands and randomly selected compounds will be conducted using Mann–Whitney U tests to assess enrichment of high-affinity hits (H1). Correlation analysis (Pearson or Spearman) will evaluate agreement between ML-predicted scores and docking energies, testing the hypothesis that ML-prioritized compounds yield improved docking performance (H2). Site-resolved analysis will be used to determine whether top candidates cluster within functionally critical regions, testing H3 regarding biologically relevant site targeting.

Contingent analyses will be implemented as IF–THEN rules:

IF the ML-based prioritization produces no enrichment over random, THEN the workflow will revert to a broader docking-based ranking to ensure candidate capture.
IF docking scores and ML predictions show discordance for a given ligand, THEN candidate selection will consider both scores with a weighted consensus metric to balance the two methods.
IF site-resolved clustering indicates a lack of occupancy in functional regions, THEN additional secondary sites will be included for targeted re-screening.

Final candidate ranking will integrate normalized docking scores, ML probabilities, interaction quality from PLIP, and predicted MD-derived stability into a weighted consensus score, guiding selection of the top 15–20 ligands for subsequent in vitro validation. All steps, thresholds, and parameters will be fully documented in the reproducible workflow to allow independent replication and verification.

Ethics and Data Management

This study will not involve human participants or animal subjects, and therefore it is exempt from formal ethical approval requirements. All computational analyses will utilize publicly available chemical libraries and protein structures, ensuring that no personal or sensitive data are processed. Confidentiality measures will include adherence to licensing terms for all datasets and software used, and any proprietary input data will be anonymized or excluded from public dissemination.

All generated data, including curated ligand libraries, protein structures, docking outputs, ML-predicted scores, and MD simulation results, will be deposited in open-access repositories such as OSF, Zenodo, or GitHub under a CC-BY license to ensure transparency and reproducibility. Metadata, software versions, parameters, and scripts will be fully documented to allow independent replication. Version-controlled workflows will be maintained using GitHub, and containerized environments (Docker/Conda) will be provided for reproducible execution of the in silico pipeline.

Preregistration of the study protocol, including hypotheses, methods, and analysis plan, will be completed on ResearchHub as a Stage 1 Registered Report. Interim progress will be shared via ResearchHub’s updates platform, and final outputs will be made publicly available, including visualization files, consensus scoring tables, and MD stability metrics, to facilitate community use and further translational research.

4. Pilot Data

Before full-scale in silico screening, the computational pipeline will undergo preliminary development and testing to ensure methodological feasibility and readiness for high-performance execution. The code implementing ligand preprocessing, machine learning prioritization, classical docking, and consensus scoring will be constructed and systematically debugged on a local computational server. Compilation and execution tests will be performed to identify and correct potential errors in script logic, input/output handling, and environment dependencies. These tests will provide proof-of-error correction and validate the workflow structure, ensuring that all modules can execute successfully in a controlled environment.

This pilot testing will not generate finalized candidate predictions and is clearly distinct from the main study; it is intended solely to demonstrate the operational viability of the pipeline, verify software compatibility, and establish protocols for data management. The local server environment will simulate aspects of cloud and HPC execution, including parallelized docking and ML scoring, to anticipate performance and resource requirements. Once the microgrant budget is allocated, the pipeline will be fully deployed on cloud servers and HPC nodes, enabling large-scale ligand evaluation and integration with consensus scoring and molecular dynamics refinement.

These preparatory activities confirm that the in silico workflow is methodologically feasible, reproducible, and scalable, providing confidence that subsequent candidate selection will proceed efficiently and reliably when connected to high-performance computational resources.

5. Budget

The proposed study will utilize the confirmed funding of $5,000 USD to cover computational, personnel, software, data storage, and contingency costs. All expenses are flexible and will be adjusted based on insights from pilot data, such as observed docking throughput or ML processing times.

High-Performance Computing (HPC) Resources – $3,000 USD: Funding will support the execution of AutoDock Vina docking across multiple sites for ~10,000 ligands, as well as molecular dynamics simulations for the top 15–20 candidates. HPC credits will cover parallelized docking and short MD runs (100 ns) in GROMACS.
Software Licenses – $500 USD: Costs will include any required licenses for specialized software, such as the Boltz-2 ML model, and auxiliary packages for preprocessing and visualization. Open-source tools (RDKit, OpenBabel, PyMOL) will be used wherever possible to optimize budget efficiency.
Data Storage – $300 USD: Budget will support secure archiving of all raw and processed data, including ligand libraries, docking outputs, ML predictions, and MD trajectory summaries, in open-access repositories (OSF, Zenodo).
Personnel – $1,500 USD: This allocation will fund a graduate research assistant to execute ligand curation, docking preparation, ML workflow execution, data integration, and documentation of the pipeline for reproducibility.
Contingency – $200 USD: Reserved for unexpected costs such as additional computational time, minor software updates, or data storage expansion.

The total projected budget is $5,000 USD, providing sufficient resources to complete the in silico screening and refinement workflow while maintaining flexibility to accommodate any adjustments indicated by pilot results.

Timeline and Workplan

The project will be executed over a six-month period, aligning computational tasks with budget allocation to ensure efficient use of HPC resources, software, and personnel. Key milestones include ligand and protein preparation, ML-based prioritization, docking, site-resolved analysis, consensus scoring, MD refinement, and data archiving. Each stage is scheduled sequentially, with partial overlap where feasible to optimize compute utilization.

Month	Task / Milestone	Budget Emphasis
1	Curate ligand library (~10,152 compounds), prepare protein structures (prefusion RSV F), define docking sites.	Personnel, initial software setup, and HPC acquisition.
2–3	Execute ML scoring (Boltz-2) to select the top ~10–20% ligands; perform AutoDock Vina docking for selected compounds at multiple binding sites.	HPC compute cloud and server credits , software licenses purchases
4	Integrate ML and docking outputs; perform site-resolved analysis; compute consensus scores; shortlist ~200–300 preliminary hits.	Personnel time for analysis, minor HPC usage
5	Conduct MD simulations for top ~15–20 ligand–protein complexes; evaluate stability (RMSD, RMSF, H-bonds, MM/PB(GB)SA binding energies).	HPC compute, software, storage
6	Finalize candidate list (~15–20 ligands); document workflows, deposit data in OSF/GitHub/Zenodo; prepare Stage 1 RRR submission.	Personnel, storage, dissemination

6. References

McLellan JS, Chen M, Leung S, Graepel KW, Du X, Yang Y, et al. Structure of RSV fusion glycoprotein trimer bound to a prefusion-specific neutralizing antibody. Science. 2013;340(6136):1113-7. doi:10.1126/science.1234914.
Gilman MSA, Furmanova-Hollenstein P, Pascual G, van 't Wout AB, Langedijk JPM, McLellan JS. Transient opening of trimeric prefusion RSV F proteins. Nat Commun. 2019;10:2105. doi:10.1038/s41467-019-09807-5.
Trott O, Olson AJ. AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J Comput Chem. 2010;31(2):455-61. doi:10.1002/jcc.21334.
Lionta E, Spyrou G, Vassilatis DK, Cournia Z. Structure-based virtual screening for drug discovery: principles, applications and recent advances. Curr Top Med Chem. 2014;14(16):1923-38. doi:10.2174/1568026614666140929124445.
Gaudillière B, Wang H, Rajakulendran MI, et al. Drug repurposing screen identifies lonafarnib as respiratory syncytial virus fusion protein inhibitor. Nat Commun. 2024;15:3145. doi:10.1038/s41467-024-45241-y.
Boltz-2: towards accurate and efficient binding affinity prediction. bioRxiv [Preprint]. 2025. doi:10.1101/2025.06.14.659707.
Feher M, Kleywegt GJ. Consensus scoring for protein-ligand interactions. Curr Opin Struct Biol. 2006;16(6):694-9. doi:10.1016/j.sbi.2006.08.002.
Charifson PS, Corkery JJ, Murcko MA, Walters WP. Consensus scoring: a method for obtaining improved hit rates from docking databases of three-dimensional structures into proteins. J Med Chem. 1999;42(25):5100-9. doi:10.1021/jm990352k.
Shen C, Sali A. Statistical potential-based consensus scoring in virtual screening. Proteins. 2006;64:553-66. doi:10.1002/prot.20935.
Wang Z, Liu X. Integration of machine learning and docking for hit prioritization in structure-based drug design. Brief Bioinform. 2019;20(6):2050-66. doi:10.1093/bib/bbz026.
Munafò MR, Nosek BA, Bishop DVM, Button KS, Chambers CD, Percie du Sert N, et al. A manifesto for reproducible science. Nat Hum Behav. 2017;1:0021. doi:10.1038/s41562-016-0021.
Eberhardt J, Santos-Martins D, Tillack AF, Forli S. AutoDock Vina 1.2.0: new docking methods, expanded force field, and Python bindings. J Chem Inf Model. 2021;61(8):3891-8. doi:10.1021/acs.jcim.1c00203.
McNutt AT, Francoeur P, Aggarwal R, Masuda T, Meli R, Ragoza M, et al. GNINA 1.0: molecular docking with deep learning. J Cheminform. 2021;13(1):43. doi:10.1186/s13321-021-00522-2.
Wilkinson MD, Dumontier M, Aalbersberg IJJ, Appleton G, Axton M, Baak A, et al. The FAIR Guiding Principles for scientific data management and stewardship. Sci Data. 2016;3:160018. doi:10.1038/sdata.2016.18.
Kampmann B, Madhi SA, Munjal I, Simões EAF, Pahud BA, Llapur C, et al. Bivalent prefusion F vaccine in pregnancy to prevent RSV illness in infants. N Engl J Med. 2023;388(16):1451-64. doi:10.1056/NEJMoa2216480.
Hammitt LL, Dagan R, Yuan Y, Baca Cots M, Bosheva M, Madhi SA, et al. Nirsevimab for prevention of RSV in healthy late-preterm and term infants. N Engl J Med. 2022;386(9):837-46. doi:10.1056/NEJMoa2110275.
Centers for Disease Control and Prevention. Vaccines for adults | RSV [Internet]. 2026 [cited 2026 Apr 2].
Centers for Disease Control and Prevention. RSV immunization guidance for infants and young children [Internet]. 2025 [cited 2026 Apr 2].
Centers for Disease Control and Prevention. Symptoms and care of RSV [Internet]. 2026 [cited 2026 Apr 2].
U.S. Food and Drug Administration. Respiratory syncytial virus (RSV) [Internet]. 2024 [cited 2026 Apr 2].
Bonneux B, Jacoby E, Ceconi M, Stobbelaar K, Delputte P, Herschke F. Direct-acting antivirals for RSV treatment, a review. Antiviral Res. 2024;229:105948. doi:10.1016/j.antiviral.2024.105948.
Trusinska D, Lee B, Ferdous S, Kwok HHY, Gordon B, Gao J, et al. Real-world uptake of nirsevimab, RSV maternal vaccine, and RSV vaccines for older adults: a systematic review and meta-analysis. EClinicalMedicine. 2025;84:103281. doi:10.1016/j.eclinm.2025.103281.
Trusinska D, Lee B, Ferdous S, Lansbury L, Burden C, Anand A, et al. Real-world evidence on RSV vaccine uptake, effectiveness, and safety in older adults: a systematic review and meta-analysis. Lancet Reg Health Eur. 2026;64:101623. doi:10.1016/j.lanepe.2026.101623.
Brüssow H. Respiratory syncytial virus: health burden, disease prevention, and treatment-recent progress and lessons learned. microLife. 2025;6:uqaf003. doi:10.1093/femsml/uqaf003.
Battles MB, McLellan JS. Respiratory syncytial virus entry and how to block it. Nat Rev Microbiol. 2019;17(4):233-45. doi:10.1038/s41579-019-0149-x.

Machine Learning–Guided and Site-Resolved In Silico Screening for Identification of RSV F-Protein Modulators

Machine Learning–Guided and Site-Resolved In Silico Screening for Identification of RSV F-Protein Modulators

Peer Reviews

DOI