Orchestrating open-access bioinformatics for seamless drug discovery

Abstract#

Computational antibiotic discovery requires combining chemical generation, structural modelling, physical validation, biological interpretation and synthetic feasibility into a reproducible workflow. This study presents the PROTO-NOOS pipeline, an in silico pipeline for prioritising small-molecule candidates in an E. coli DHFR context. The workflow used REINVENT4 for de novo molecule generation, RDKit-based physicochemical and Gram-negative entry filtering, Boltz-2 for protein-ligand complex and affinity prediction, GROMACS for molecular dynamics stability checks, BioTransformer3 and COBRApy/iML1515 for metabolite and target-perturbation analysis, cellTE for ODE-based intracellular target engagement, and AiZynthFinder for retrosynthetic accessibility.

The pipeline connects chemical, structural, MD, metabolic, target-level and retrosynthetic evidence in one screening process. Its outputs should be interpreted as computational ranking features for hypothesis generation, not as confirmed antibacterial activity: synthesis, antibacterial assays, toxicity testing and wet-lab validation were outside the scope of this work.

Main findings#

• Multi-signal ranking narrows the search space. Of 100 generated compounds, 27 retained a non-zero score; the rest were treated as low-priority under the current evidence model.
• cellTE adds context beyond K_D. Rescoring after structure prediction combines predicted affinity with entry and efflux dynamics, so compounds with similar K_D separate by intracellular exposure.
• Short MD remains a low-confidence gate. The 100 ps high-throughput mode is useful for flagging unstable systems, but its kinetic proxies are explicitly weak labels.
• Retrosynthesis changes lead priority. The top-ranked molecule was not solved by AiZynthFinder in the reported batch, while several lower-ranked candidates had practical routes.

Figures#

Scope and limits#

PROTO-NOOS is an in silico prioritisation framework, not a confirmed antibiotic discovery. This study includes no compound synthesis, MIC testing, target-engagement assays, cytotoxicity assays or wet-lab antibacterial experiments. The practical conclusion is limited to workflow feasibility, hypothesis generation and candidate triage before deeper simulation or experimental follow-up.

Selected references#

1. Stokes, J. M., Yang, K., Swanson, K., et al. A deep learning approach to antibiotic discovery. Cell, 2020.
2. Abraham, M. J., Murtola, T., Schulz, R., et al. GROMACS: high performance molecular simulations through multi-level parallelism. SoftwareX, 2015.
3. Passaro, S., Corso, G., Wohlwend, J., et al. Boltz-2: towards accurate and efficient binding affinity prediction. bioRxiv, 2025.
4. Djoumbou-Feunang, Y., Fiamoncini, J., Gil-de-la-Fuente, A., et al. BioTransformer: a comprehensive computational tool for small molecule metabolism prediction. Journal of Cheminformatics, 2019.
5. Loeffler, H. H., He, J., Tibo, A., et al. REINVENT 4: modern AI-driven generative molecule design. Journal of Cheminformatics, 2024.
6. Genheden, S., Thakkar, A., Chadimova, V., et al. AiZynthFinder: a fast, robust and flexible open-source software for retrosynthetic planning. Journal of Cheminformatics, 2020.