Microbiome-Aware Quantum–AI Drug Discovery for Precision Antimicrobial Design

1. Abstract

Antimicrobial discovery pipelines rarely account for the impact of candidate therapeutics on the human microbiome, which contributes to antibiotic‑induced dysbiosis, secondary infections, and long‑term health effects. Existing AI and quantum‑enhanced methods focus primarily on target binding and pathogen kill, with minimal consideration of microbiome preservation or bystander effects.

This project extends an operational Quantum–AI drug discovery platform to incorporate an explicit microbiome‑aware modeling layer for small‑molecule candidates targeting New Delhi metallo‑β‑lactamase 1 (NDM‑1) and clinically relevant viral proteases. The extended pipeline will jointly optimize for (1) predicted efficacy against resistance‑associated targets and (2) minimal disruption of commensal microbiota, supporting the design of microbiome‑sparing, precision antimicrobials.

The team already operates a production‑ready, containerized platform integrating quantum simulation, classical molecular modeling, and AI/ML modules on commercial cloud infrastructure, and requires no new facilities or ethical approvals. All work is fully computational, using public molecular and microbiome datasets plus commercial quantum services (e.g., AWS Braket). Results, models, and evaluation code will be released in open repositories where licensing permits, to maximize reproducibility and community reuse. The project is scoped for completion within six months under a total direct cost of $20,000.

2. Introduction and Hypotheses

2.1 Background and Significance

Antimicrobial resistance (AMR) and antibiotic‑induced microbiome disruption are converging crises. Broad‑spectrum agents that successfully clear pathogens often drive dysbiosis, which is associated with higher risk of C. difficile infection, metabolic dysfunction, and other long‑term sequelae. Recent work on “microbiome‑sparing” antibiotics (e.g., lolamicin) demonstrates that it is feasible to design agents with potent activity against target pathogens while sparing commensal species by exploiting pathogen‑specific vulnerabilities.

In parallel, AI‑driven molecular design and structure‑based virtual screening have become standard tools for accelerating antimicrobial discovery, and have successfully produced novel NDM‑1 inhibitors via multi‑step docking and molecular modeling workflows. Quantum‑inspired and quantum‑enabled methods such as variational quantum eigensolvers (VQEs) have been explored for electronic‑structure calculations and binding‑energy refinement, although they face practical challenges related to noise and barren plateau phenomena on current hardware.

Despite these advances, few pipelines optimize for efficacy and microbiome preservation jointly. Most frameworks design for target engagement alone, implicitly tolerating collateral damage to commensal microbes. This proposal aims to fill that gap by adding a microbiome‑aware layer on top of an existing Quantum–AI drug discovery workflow.

2.2 Specific Aims

Aim 1: Extend an existing Quantum–AI drug discovery pipeline to incorporate a microbiome‑aware scoring function that penalizes candidates predicted to significantly disrupt representative human gut microbiota while preserving on‑target efficacy against NDM‑1 and selected viral proteases.
Aim 2: Quantitatively compare candidate ranking and down‑selection between the baseline (efficacy‑only) pipeline and the microbiome‑aware pipeline.

2.3 Hypotheses

H1 (Primary): Incorporating microbiome‑aware modeling into the screening pipeline will yield candidate sets that maintain predicted efficacy against NDM‑1 and viral proteases while substantially reducing predicted microbiome disruption compared to efficacy‑only screening.
H2 (Secondary): Microbiome‑aware scoring will induce statistically significant shifts in candidate rankings relative to a baseline that optimizes only for binding affinity / target inhibition.

3. Methods

This project is entirely in silico, using public molecular and microbiome resources and commercially available quantum backends.

3.1 Data and Targets

Pathogen targets:
- NDM‑1 metallo‑β‑lactamase (carbapenem resistance) with structures and ligand complexes drawn from PDB and prior inhibitor studies.
- One to two clinically relevant viral proteases (e.g., SARS‑CoV‑2 Mpro or related proteases) with high‑quality structural data.
Ligand/candidate space:
- Small‑molecule libraries from ChEMBL and ZINC filtered for drug‑like properties and physicochemical constraints.
- Optionally, focused libraries enriched in scaffolds related to known NDM‑1 inhibitors and protease inhibitors for benchmarking.
Microbiome panel:
- A curated panel of representative gut commensals and opportunistic pathogens (e.g., Bacteroides, Faecalibacterium, Lactobacillus, Enterococcus, Escherichia, Clostridioides).
- Where specific ligand–microbiome interaction data exists (e.g., growth inhibition profiles or predicted off‑target binding for known antibiotics), this will be used to train or calibrate microbiome response models.

3.2 Pipeline Overview

The extended workflow has four main components:

Classical structure‑based modeling
- Standard docking and molecular mechanics (MM/GBSA) calculations to estimate ligand–target binding affinity for NDM‑1 and viral proteases.
- Generation of pose ensembles for downstream quantum refinement.
Quantum refinement (VQE layer)
- For a selected subset of small ligand–binding‑site subsystems (e.g., fragments in the NDM‑1 active site), we will apply VQE‑style electronic structure calculations using a hardware‑efficient ansatz on small qubit registers.
- We explicitly constrain quantum workloads to small systems where circuit depth and noise permit meaningful gradients, acknowledging known issues such as noise‑induced barren plateaus and gradient suppression.
- Quantum‑refined energies will be used as a re‑ranking signal on top of classical scores, not as a standalone decision metric, to remain robust under NISQ limitations.
AI/ML modeling
- Graph neural networks (GNNs) / message passing neural networks for ligand–target affinity prediction, trained or fine‑tuned on relevant antimicrobial and NDM‑1 datasets where available.
- Optional sequence/structure transformers for proteases to capture variant‑specific effects.
- Multi‑task learning setup where outputs include:
  - Predicted activity against NDM‑1 and proteases (regression or classification).
  - Predicted activity / off‑target impact against members of the microbiome panel, either via direct training (where data exists) or via transfer learning from related antibiotic–microbiome datasets.
Microbiome‑aware scoring layer
- For each candidate, compute:
  - Efficacy score EE: aggregated from classical docking/MM, quantum refinement (where available), and ML activity predictions on pathogen targets.
  - Microbiome disruption score MM: derived from predicted inhibitory activity against commensal species, weighted by relative abundance and ecological importance (e.g., higher penalties for disrupting keystone taxa).
- Define a combined objective such as
  S=E−λMS=E−λM
  where λλ is a tunable hyperparameter controlling the trade‑off between pathogen kill and microbiome preservation (to be calibrated in Aim 2 experiments).

The existing platform already orchestrates containerized jobs for classical modeling, quantum backends, and AI inference; this project primarily adds microbiome‑focused models, data, and scoring, plus evaluation experiments.

3.3 Planned Analyses

We will evaluate H1 and H2 using the following analyses:

Baseline vs microbiome‑aware rank comparison
- Generate ranked lists of candidates using:
  - Baseline pipeline (efficacy‑only scores, no microbiome term).
  - Extended pipeline (combined S=E−λMS=E−λM).
- Quantify rank shifts via Kendall’s tau and Spearman correlation between the two rankings.
- Identify candidates that move substantially in rank when microbiome penalties are added (e.g., Δrank > N positions), and analyze structural / mechanistic patterns behind those shifts.
Top‑k enrichment and trade‑off curves
- For known active vs inactive compounds (e.g., from published NDM‑1 inhibitor studies), estimate enrichment factors and ROC‑AUC for both pipelines.
- Construct Pareto‑like curves showing predicted efficacy vs microbiome disruption across candidate sets for different λ values, to visualize trade‑off frontiers.
Microbiome disruption simulation
- Using predicted activity profiles against the microbiome panel, simulate simple community perturbation scenarios (e.g., reduction of specific commensals) and compare baseline vs microbiome‑aware‑selected candidates.
- Summarize results in terms of predicted preservation of diversity / key taxa relative to agents selected by efficacy‑only ranking.
Sensitivity analyses
- Vary λ and microbiome weighting schemes to assess robustness of conclusions (e.g., whether qualitative rank shifts persist).
- Evaluate how sensitive outcomes are to inclusion/exclusion of the quantum refinement signal, given NISQ limitations.

We will treat H1 as supported if the microbiome‑aware pipeline yields candidate sets with comparable predicted efficacy but materially lower MM (disruption) metrics than the baseline. H2 is supported if rank correlations are significantly below 1.0 and structured shifts can be explained by microbiome liabilities.

3.4 Ethics and Data Management

No human or animal subjects are involved.
All data comes from open or properly licensed resources (PDB, ChEMBL, PDBBind/ZINC, curated microbiome datasets).
All code for the microbiome‑aware layer, model training scripts, and analysis notebooks will be released under an open‑source license, with configuration files that allow reproduction of experiments on public cloud and quantum backends where users have accounts.

4. Pilot Data and Feasibility

The team operates an existing Quantum–AI discovery platform with the following demonstrated capabilities (details in the linked internal documentation and prior preprints):

Automated ingestion of target structures from PDB and ligand libraries from ChEMBL/ZINC.
Containerized workflows for docking, MM/GBSA, and ML‑based affinity prediction.
Integration with managed quantum services (e.g., AWS Braket) to run small‑scale VQE experiments on toy systems, used today as a re‑ranking signal rather than as a primary engine, in line with current NISQ constraints.

To directly address the reviewer’s concern: for this proposal, we will not claim unsubstantiated performance “wins” for the existing platform. Instead, we will explicitly treat the current pipeline as:

Baseline: validated mainly by its ability to reproduce trends and rankings consistent with published NDM‑1 inhibitor studies and protease inhibitors (e.g., reproducing enrichment of known actives from decoy sets).
Starting point: a stable, orchestrated stack that can be extended with microbiome data and scoring without major engineering risk.

Where available, we will reference specific benchmark reproductions (e.g., top‑k enrichment against public NDM‑1 datasets) in an appendix or technical supplement rather than relying solely on internal blog posts or non‑reviewed writeups.

The existing team is already running and maintaining this platform in production settings, which substantially de‑risks feasibility.

5. Team and Roles

This project will be executed by an existing interdisciplinary team:

Principal Investigator (PI): Leads overall project, coordinates aims, ensures scientific rigor and reporting.
Quantum / Computational Physics Lead: Designs and runs VQE experiments, chooses ansatz/cost functions, and ensures quantum components remain within realistic NISQ limits.
AI/ML Lead: Develops and fine‑tunes GNN and transformer models for pathogen and microbiome predictions, oversees training and validation.
Cheminformatics / Microbiome Specialist: Curates ligand sets, microbiome panels, and interprets predicted microbiome disruption in the context of microbiome health and microbiome‑sparing antibiotic strategies.
DevOps / Platform Engineer (part‑time allocation): Maintains containerized infrastructure, cloud and quantum integrations, and reproducibility pipelines.

No new hires are strictly required for the 6‑month scope; the requested budget mainly covers compute, quantum service credits, and partial support for existing team members’ time.

6. Budget (Total: $20,000)

All costs are in USD and reflect a 6‑month project timeline.

Compute ($6,000)
- Cloud GPU/CPU time for model training and docking/MM simulations.
- Managed quantum service credits (e.g., AWS Braket) for VQE experiments on selected subsystems.
Personnel ($8,000)
- Partial support (fractional FTE) for PI, ML lead, and microbiome specialist to design experiments, interpret results, and prepare manuscripts.
- Does not cover full‑time salaries; this is a top‑up against existing funded time.
Data and Tooling ($2,000)
- Access to specialized cheminformatics or microbiome datasets if needed (where not freely available).
- Licenses for any required commercial docking/ADMET tools not already held by the team.
Validation and Benchmarking ($2,000)
- Internal validation experiments against public NDM‑1 and protease inhibitor benchmarks (compute + analysis time).
- Sensitivity analyses and ablation studies (with and without quantum layer and microbiome term).
Reporting and Dissemination ($2,000)
- Preparation of preprints and, if applicable, open‑access fees or conference registrations.
- Documentation and packaging of code and datasets for public release.

This breakdown directly answers prior feedback regarding vague items such as “Validation” and clarifies who performs which work under each line.

Rezaei, A. R., Ates, F., Sulik, A., & Toczyłowski, K. (2025). “Smart”, microbiome‑sparing antibacterial therapy with a focus on the novel Lolamicin: An overview. Infection, 53(6), 2267–2276. https://doi.org/10.1007/s15010-025-02538-4
Varkevisser, M. (2025, May 8th). Reprogramming Drug Discovery: How Vibration Robotics is Merging Quantum and AI to Accelerate Therapeutics. LinkedIn. https://www.linkedin.com/pulse/reprogramming-drug-discovery-how-vibration-robotics-xxjuc
Muñoz, K. A., et al. (2024). A Gram‑negative‑selective antibiotic that spares the gut microbiome. Nature, 630(8016), 429–436. https://doi.org/10.1038/s41586-024-07502-0
Sanapalli, B. K. R., Palit, S., Deshpande, A., Tokala, R., Sigalapalli, D. K., & Sanapalli, V. (2026). Artificial Intelligence and the discovery of antibiotics: Reinventing with opportunities, challenges, and clinical translation. Antibiotics, 15(2), 233. https://doi.org/10.3390/antibiotics15020233
Wang, X., Yang, Y., Gao, Y., & Niu, X. (2020). Discovery of the novel inhibitor against New Delhi metallo‑β‑lactamase based on virtual screening and molecular modelling. International Journal of Molecular Sciences, 21(10), 3567. https://doi.org/10.3390/ijms21103567
Yang, Y., Guo, Y., Zhou, Y., Gao, Y., Wang, X., Wang, J., & Niu, X. (2020). Discovery of a novel natural allosteric inhibitor that targets NDM‑1 against Escherichia coli. Frontiers in Pharmacology, 11, 581001. https://doi.org/10.3389/fphar.2020.581001
(For the VQE limitations) Noise‑induced barren plateaus in variational quantum algorithms. Nature Communications. (2021). https://www.nature.com/articles/s41467-021-27045-6

Microbiome-Aware Quantum–AI Drug Discovery for Precision Antimicrobial Design

Microbiome-Aware Quantum–AI Drug Discovery for Precision Antimicrobial Design

Funding Progress

Nonprofit Recipient

Funders

Topics

DOI