Optimizing Gene-Based Testing for Antibiotic Resistance Prediction

David Hagerman, Anna Johnning, Roman Naeem, Fredrik Kahl, Erik Kristiansson, Lennart Svensson

TL;DR

The GenoARM framework combines reinforcement learning (RL) with transformer-based models to identify the most informative subset of genes for predicting antibiotic resistance. By optimizing the selection of just a few key genetic markers, the system achieves prediction accuracy comparable to full-genome sequencing while being significantly faster and more cost-effective for clinical use.

The Challenge of Antimicrobial Resistance

Antimicrobial Resistance (AMR) is a global health crisis, projected to cause 10 million deaths annually by 2050. To treat patients effectively and slow the spread of resistance, clinicians require rapid diagnostics to determine which antibiotics a pathogen is resistant to.

Currently, diagnostics face a significant trade-off:

  1. Whole Genome Sequencing (WGS): Highly accurate but slow, expensive, and impractical for rapid bedside decision-making.
  2. Polymerase Chain Reaction (PCR): Fast and affordable, but limited to testing a small subset of genes.

The core challenge is the "selection problem." With over 5,000 known resistance-linked genes and mutations, identifying the 5 to 15 most informative markers for a PCR test is difficult. Simple statistical approaches often fail to capture the complex, redundant patterns of resistance.

Methods: The GenoARM Framework

The authors introduce GenoARM, a dual-architecture framework designed to solve the selection problem through two integrated modules:

  • The AR Prediction Model: A transformer-based architecture that processes a subset of gene test results alongside metadata (e.g., year, country, and sample source). It employs a novel embedding strategy that pairs "gene tokens" with "existence tokens" (present/absent), allowing the model to efficiently process unordered sets of genetic data.
  • The Policy Network: An RL agent trained via Proximal Policy Optimization (PPO). This agent treats gene selection as a sequential decision process, "picking" genes and receiving rewards based on the resulting prediction model's performance.
The GenoARM training loop using Proximal Policy Optimization (PPO) to select informative genes.

Key Innovation: Rather than using a non-smooth metric like raw accuracy for rewards, the authors utilize the unreduced loss from the prediction model. This provides a more stable signal for the RL agent to identify which genes provide the highest information gain.

Datasets and Experimental Setup

The study utilized publicly available data from the NCBI pathogen detection system, focusing on five clinically significant pathogens: Escherichia coli, Klebsiella pneumoniae, Salmonella, Staphylococcus aureus, and Acinetobacter. Data preprocessing involved standardizing gene nomenclature and grouping metadata categories (e.g., consolidating various blood-related sample types into a single "blood" category).

Key Results and Insights

  • Superior Accuracy with Fewer Genes: On E. coli, GenoARM achieved 93.77% accuracy using only 5 genes plus metadata, outperforming genetic algorithm and statistical baselines.
  • The Law of Diminishing Returns: Experimental results show that a subset of just 15 optimized genes performs nearly as well as using the entire set of approximately 200 available genes. This suggests that resistance profiles are highly redundant, and an optimized small-scale test can effectively replace comprehensive sequencing.
Accuracy vs. number of genes tested (K), demonstrating that GenoARM (GenARM) achieves near-peak performance with very few genes.
  • Metadata as a Power Multiplier: Geographic and temporal metadata significantly improve predictions. For instance, knowing a sample originated from a urine source in the USA allows the model to prioritize genes like tet(A) and tet(B), which are highly relevant for those specific conditions.
  • Error Reduction: GenoARM significantly reduced "Very Major Errors" (VMEs)—instances where a resistant sample is incorrectly predicted as susceptible—which is critical for patient safety in clinical settings.

Limitations and Future Work

The model currently relies on a library of known resistance genes; discovery of novel mutations would require retraining. Additionally, geographic bias in the training data (primarily from the US) may affect performance in other regions. Future research aims to expand the dataset to include more diverse global samples and develop "universal" PCR panels that function across multiple bacterial species.

Summary updated

BibTeX
@inproceedings{hagerman2025optimizing,
  author = {David Hagerman and Anna Johnning and Roman Naeem and Fredrik Kahl and Erik Kristiansson and Lennart Svensson},
  title = {{Optimizing Gene-Based Testing for Antibiotic Resistance Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35021},
}
Figure