Antimicrobial Resistance Prediction in Salmonella enterica Using Frequency Chaos Game Representation and ResNet-18

This study developed and evaluated a deep learning model combining Frequency Chaos Game Representation (FCGR) and ResNet-18 to predict antimicrobial resistance in *Salmonella enterica* and *Staphylococcus aureus*, demonstrating promising results for specific antibiotics like cephalosporins while acknowledging that the approach currently requires further improvement to match the performance of established gene-based tools like ResFinder for clinical application.

The Gist

The Big Picture: A High-Speed "Genomic X-Ray"

Imagine a bacterial infection is like a burglar trying to break into a house (your body). Doctors need to know quickly which lock-picking tools (antibiotics) will work to stop the burglar and which ones the burglar has already learned to bypass (resistance).

Traditionally, doctors have to wait 18 to 24 hours to grow the bacteria in a lab to see what works. This paper tries to build a super-fast AI scanner that looks at the bacteria's genetic blueprint (DNA) and instantly predicts which drugs will fail, without needing to grow the bacteria first.

The Problem: The "Dictionary" Approach vs. The "Pattern" Approach

Current methods for predicting resistance are like using a dictionary.

• How it works: Scientists have a list of known "bad words" (resistance genes). If the bacteria's DNA contains one of these words, the AI says, "Alert! This bacteria is resistant."
• The flaw: If the bacteria invents a new way to resist drugs, or uses a complex combination of words the dictionary doesn't know, the dictionary method fails. It can't read between the lines.

This paper tries a different approach: The "Art Gallery" method.
Instead of reading the words, they turn the entire DNA sequence into a 2D picture (an image). They use a technique called Frequency Chaos Game Representation (FCGR).

• The Analogy: Imagine taking a long, boring book of DNA letters and folding it up into a tiny, intricate mosaic tile. Even though you can't read the words anymore, the pattern of the tiles tells you if the book is a mystery novel or a cookbook.
• The Goal: Train a computer vision AI (like the one used to recognize cats in photos) to look at these DNA mosaics and say, "That specific pattern means the bacteria is resistant to Penicillin."

The Experiment: Two Different Neighborhoods

The researchers tested their "AI Art Scanner" on two very different bacterial neighborhoods:

1. Salmonella: A Gram-negative bacteria (like a house with a double-layered brick wall).
2. Staphylococcus aureus: A Gram-positive bacteria (like a house with a single, thick stone wall).

They wanted to see if their AI could learn the "visual language" of resistance in both types of houses.

The Results: A Mixed Bag

Here is how the "Art Scanner" (the AI) performed compared to the "Dictionary" (the standard tool called ResFinder):

1. The Winners (Cephalosporins):
For a specific family of antibiotics (cephalosporins), the AI was a star performer. It predicted resistance almost perfectly.

• Why? It seems these antibiotics share a common "visual signature" in the DNA mosaic. The AI got really good at spotting that specific pattern.

2. The Strugglers (Tetracycline & Ampicillin):
For other drugs, the AI stumbled. It often missed resistant bacteria or thought sensitive ones were resistant.

• The Reality Check: The standard "Dictionary" tool (ResFinder) was still much better at these. The AI couldn't find the patterns as well as the experts who know exactly which genes to look for.

3. The "Leakage" Trap:
One of the smartest parts of this paper was how they split the data.

• The Problem: If you train an AI on a picture of a specific dog, and then test it on a picture of its twin, the AI looks smart but is actually just cheating (memorizing the family).
• The Fix: The researchers used a special "homology-aware" filter. They made sure the AI never saw the "twins" (closely related bacteria) during training. This proved the AI was actually learning the concept of resistance, not just memorizing specific bacteria.

The Verdict: Promising, But Not Ready for the ER

The Good News:
The study proves that you can turn DNA into pictures and use AI to predict drug resistance. It works across different types of bacteria (Gram-negative and Gram-positive), showing the method is flexible.

The Bad News:
Right now, the AI is not better than the existing dictionary tools.

• The "Dictionary" (ResFinder) is still the gold standard because it is built on decades of biological research.
• The AI is like a talented art student who can guess the genre of a book by looking at the cover art, but the librarian (ResFinder) still knows the plot better because they've read the index.

Why This Matters for the Future

Think of this paper as a prototype for a self-driving car.

• It proves the engine works and the car can drive on the road.
• However, it's not safe enough to replace the human driver (the current lab tests) just yet.

The authors conclude that while this "AI Art Scanner" is a fascinating proof-of-concept, we need more data, better training, and more powerful computers before we can trust it to save lives in a hospital. It's a step toward a future where we can diagnose antibiotic resistance in seconds, but we aren't there yet.

Technical Summary

1. Problem Statement

Antimicrobial Resistance (AMR) is a critical global health threat, yet current methods for predicting resistance face significant limitations:

• Traditional Methods: Phenotypic testing takes 18–24 hours, which is too slow for immediate clinical decision-making.
• Current Genomic Tools: Tools like ResFinder and ABRicate rely on matching sequences against known resistance gene databases. While effective, they cannot detect novel resistance mechanisms, miss complex multi-gene interactions, and fail to capture resistance patterns arising from evolutionary pathways not yet cataloged.
• Machine Learning Challenges: Existing deep learning approaches for AMR often suffer from data leakage (training and testing on genetically similar isolates), lack of cross-species generalizability, and insufficient comparison against standard gene-based baselines.

2. Methodology

The authors propose a deep learning pipeline that predicts AMR phenotypes directly from whole-genome assemblies using an alignment-free approach.

A. Data Preparation & Homology-Aware Splitting

• Datasets: Two datasets were used: Salmonella enterica (7 antibiotics) and Staphylococcus aureus (5 antibiotics).
• Leakage Prevention: To prevent genomic data leakage, the authors avoided standard random splitting. Instead, they used sourmash (MinHash sketching) to calculate pairwise genomic distances. Genomes with a distance $d \le 0.05$ (approx. 95% shared 31-mers) were clustered into connected components.
• Splitting Strategy: Entire clusters were assigned to either the training or test set (GroupKFold), ensuring no homologous genomes appeared in both sets. This resulted in an 87.57% / 12.43% split for Salmonella and a 90.55% / 9.45% split for S. aureus.

B. Feature Engineering: Frequency Chaos Game Representation (FCGR)

• Transformation: DNA sequences were converted into 2D images using FCGR, which maps $k$-mer frequencies to a grid.
• Parameters: The authors tested $k=6, 7, 8$ and selected $k=8$ based on empirical performance. This generates $256 \times 256$ matrices (65,536 unique $k$-mers), providing sub-gene-level resolution sufficient to capture conserved motifs in resistance determinants.
• Preprocessing: Matrices were min-max normalized to $[0, 1]$. Labels were binarized where Susceptible (S) = 1 (positive class) and Resistant/Intermediate (R/I) = 0, chosen to improve model convergence stability.

C. Deep Learning Architecture

• Model: A ResNet-18 architecture was employed as the backbone.
• Modifications: The input layer was adapted for single-channel images, and the final layer was replaced with a linear classifier for multi-label antibiotic prediction.
• Training: Models were trained from scratch (no pre-trained weights) using the Adam optimizer with a one-cycle learning rate schedule.
• Regularization: MixUp augmentation ($\alpha=0.1$) was used. Class imbalance was handled via weighted binary cross-entropy loss.
• Evaluation: Metrics included Balanced Accuracy, MCC, F1-score, and ROC AUC.

3. Key Contributions

1. Alignment-Free Deep Learning Pipeline: Demonstrated the feasibility of using FCGR images combined with ResNet-18 to predict AMR without relying on gene databases.
2. Rigorous Data Splitting: Implemented a homology-aware clustering strategy using MinHash to eliminate genomic leakage, addressing a common flaw in previous ML-based AMR studies.
3. Cross-Species Generalization: Validated the pipeline on both Gram-negative (S. enterica) and Gram-positive (S. aureus) bacteria, proving the method is not restricted to a single lineage.
4. Comprehensive Benchmarking: Directly compared the deep learning model against the gold-standard gene-based tool ResFinder (via ABRicate) and other ML baselines (Random Forest, simple CNN).

4. Results

A. Performance on Salmonella enterica

• Overall: Achieved a Balanced Accuracy (BA) of 0.86 and MCC of 0.73.
• Drug-Specific:
  • High Performance: Cephalosporins (cefoxitin, ceftiofur, ceftriaxone) showed near-perfect prediction (BA $\ge$ 0.94, MCC $\ge$ 0.84). The authors attribute this to correlated resistance labels driven by shared $\beta$-lactamase mechanisms.
  • Lower Performance: Tetracycline (BA 0.79), Ampicillin (BA 0.71), and Amoxicillin-clavulanic acid (BA 0.74) showed lower accuracy.
• Comparison with ResFinder:
  • ResFinder significantly outperformed the CNN on Tetracycline (BA 0.98 vs. 0.79) and Ampicillin (BA 0.96 vs. 0.71).
  • The CNN was competitive with ResFinder on Cephalosporins (e.g., Cefoxitin: CNN BA 0.94 vs. ResFinder BA 0.97).

B. Performance on Staphylococcus aureus

• Overall: Achieved a BA of 0.74 and MCC of 0.44.
• Drug-Specific: Strongest results were for Methicillin (BA 0.85), consistent with the well-defined mecA mechanism. Performance was weaker for Erythromycin and Clindamycin due to heterogeneous resistance mechanisms.
• Comparison: ResFinder consistently outperformed the CNN across all S. aureus drugs (e.g., Erythromycin: ResFinder BA 0.90 vs. CNN 0.68).

C. Baseline Comparisons

• Random Forest: Performed poorly (BA 0.56–0.65), indicating that simple gene presence/absence features were insufficient without the spatial context provided by FCGR.
• Simple CNN: A 3-block CNN achieved lower performance (BA 0.69) than ResNet-18, highlighting the benefit of deeper architectures.
• EfficientNet: A larger architecture (EfficientNet) achieved BA 0.85, approaching but not surpassing ResFinder.

D. Interpretability

• Saliency maps revealed that the model's predictions for cephalosporin resistance were not driven solely by the specific bla$_{CMY-2}$ gene (only 3.6% of top $k$-mers fell within this gene). Instead, the model relied on broader genomic patterns, likely reflecting clonal lineage effects and population structure rather than direct causal mechanisms.

5. Significance and Limitations

Significance:

• The study proves that alignment-free, database-independent deep learning models can achieve meaningful AMR prediction, particularly for drugs with correlated resistance mechanisms (like cephalosporins).
• It establishes a robust framework for preventing data leakage in genomic ML studies, which is critical for clinical applicability.
• It demonstrates that the FCGR + ResNet pipeline is adaptable across different bacterial phyla (Gram-negative and Gram-positive).

Limitations:

• Performance Gap: The model generally underperforms compared to established gene-based tools (ResFinder) for most antibiotics, particularly those with diverse resistance mechanisms.
• Computational Cost: Generating FCGR matrices is computationally intensive.
• Information Loss: FCGR discards positional context (e.g., plasmid vs. chromosomal location), which is crucial for understanding resistance spread.
• Clinical Readiness: The model requires prospective clinical validation and regulatory approval (IVD) before deployment. The current performance gap with ResFinder is a barrier to clinical adoption.

Conclusion:
While the FCGR-based deep learning approach does not yet surpass gene-based methods in accuracy, it offers a promising alternative for detecting resistance patterns that may be missed by database matching, provided future work addresses computational efficiency, incorporates positional context, and validates performance on larger, multi-species datasets.