Cross-Species Antimicrobial Resistance Prediction from Genomic Foundation Models

This paper demonstrates that cross-species antimicrobial resistance prediction, which fails with standard k-mer baselines due to out-of-distribution generalization challenges, can be significantly improved by leveraging genomic foundation model embeddings from Evo-1-8k-base and applying MiniRocket to preserve localized resistance signals rather than relying on global pooling.

The Gist

Imagine you are a detective trying to solve a mystery: Why is a specific bacterium resistant to an antibiotic?

For a long time, scientists tried to solve this by looking at the bacteria's DNA like a list of ingredients. They looked for specific "bad words" (k-mers) that usually mean resistance. But this approach had a huge flaw: it was like trying to recognize a person only by their accent. If you learned to identify a criminal by their New York accent, you might fail to catch a criminal from London who committed the exact same crime but speaks with a British accent.

This thesis, written by Huilin Tai, tackles this problem using a new kind of AI called a Genomic Foundation Model. Think of this model as a super-smart student who has read every book in the library of bacterial DNA. It understands the "language" of genes deeply.

However, even this super-student has trouble when asked to predict resistance in a new type of bacteria it has never seen before. Here is the story of how the author fixed this, explained through three simple analogies.

1. The Problem: The "Accent" Trap

The main challenge is Cross-Species Prediction.

• The Old Way: The AI was trained on E. coli and then tested on Salmonella. It failed because it learned to recognize "E. coli-ness" (the accent) rather than the actual "resistance mechanism" (the crime).
• The Analogy: Imagine you are teaching a robot to identify "fire." You show it pictures of campfires in the woods. The robot learns that "fire = wood + smoke." Then, you show it a gas stove fire. The robot says, "No fire here, because there is no wood!" It failed because it focused on the background (wood) instead of the signal (flame).

2. The First Fix: Finding the "Sweet Spot" in the Brain

The AI model (called Evo) has 32 layers of "thinking," like a skyscraper with 32 floors.

• The Mistake: Most people assume the top floor (the final layer) has the best answers. But in this model, the top floor is actually a bit "fried." The numbers get too big, and the signal gets messy (like a radio station with too much static).
• The Discovery: The author built a diagnostic tool to check every floor. They found that Floor 10 is the "Goldilocks Zone."
  • Floors 1–9: Too raw, not enough understanding.
  • Floor 11+: Too compressed, the signal is distorted.
  • Floor 10: Just right. It's stable, clear, and holds the most useful information without the noise.
• The Analogy: It's like listening to a song. The bass (low floors) is too muddy, and the treble (top floors) is too sharp and distorted. The middle range (Floor 10) is where you can clearly hear the melody.

3. The Second Fix: The "Zoom Lens" vs. The "Wide Angle"

Once the AI reads the DNA, it has to summarize the whole genome into a single report. This is where the author introduced two different ways to look at the data.

Method A: The "Wide Angle" Lens (Global Pooling)

This method takes the average of the entire genome.

• How it works: It calculates the "average mood" of the whole bacteria.
• When it works: It's great for resistance that is spread out everywhere, like a slow-burning fever caused by many small changes in the body's system (chromosomal mutations).
• The Flaw: If the resistance is caused by a tiny, specific "cassette" of genes (like a hidden weapon), averaging the whole genome dilutes it. It's like trying to find a single needle in a haystack by measuring the average height of the hay. You miss the needle.

Method B: The "Zoom Lens" (MiniRocket)

This method treats the DNA as a story or a signal that flows in order.

• How it works: Instead of averaging, it uses a technique called MiniRocket to scan the DNA for specific patterns and sequences, looking for those tiny, localized "cassettes" (like plasmids carrying resistance genes).
• The Analogy: Imagine you are looking for a specific phrase in a book.
  • Global Pooling reads the whole book and tells you the average sentiment (e.g., "This book is mostly sad"). It misses the specific sentence that changes the plot.
  • MiniRocket scans the pages looking for that specific sentence, even if it's only on page 42. It preserves the local detail.

The Big Surprise: It Depends on the "Crime"

The most important discovery of this thesis is that neither method is always better. It depends on how the bacteria is resisting the drug.

• Scenario 1: The "Hacker" (Cassette-Mediated Resistance)

  • The bacteria stole a specific "hack" (a gene cassette) from another species.
  • Winner: MiniRocket (Zoom Lens). Because the "hack" is a localized, specific pattern, the Zoom Lens finds it perfectly. The AI can say, "Ah, this bacteria has the same 'hack' as that other species, even though they look different!"
  • Result: The AI becomes incredibly accurate, even for bacteria it has never seen before.

• Scenario 2: The "Slow Evolution" (Chromosomal Resistance)

  • The bacteria changed its own internal machinery slowly over time.
  • Winner: Global Pooling (Wide Angle). Because the change is spread out, the average view captures it better. The Zoom Lens gets confused by too much detail.

The Conclusion: A New Rulebook

The author proves that to predict antibiotic resistance across different species, you cannot use a "one-size-fits-all" approach.

1. Don't look at the top floor of the AI brain; look at Floor 10.
2. Don't just average the data; sometimes you need to scan for specific patterns.
3. Match the tool to the problem: If the bacteria uses a "stolen tool" (cassette), use the Zoom Lens. If it uses "slow evolution" (chromosomal), use the Wide Angle.

Why does this matter?
Antibiotic resistance kills over a million people a year. Doctors currently have to wait days for lab tests to see which drugs work. This research gives us a blueprint to build AI that can look at a bacteria's DNA and instantly predict which drugs will work, even if it's a brand new type of bacteria, by understanding the mechanism of the resistance rather than just memorizing the species.

In short: To catch the criminal, you need to understand the crime, not just the criminal's accent.

Technical Summary

Here is a detailed technical summary of the thesis "Cross-Species Antimicrobial Resistance Prediction from Genomic Foundation Models" by Huilin Tai.

1. Problem Statement

The central challenge addressed is Cross-Species Antimicrobial Resistance (AMR) Prediction, which is fundamentally an Out-of-Distribution (OOD) generalization problem.

• The Core Difficulty: Models trained on one set of bacterial taxa often fail when applied to phylogenetically distinct species. This is because resistance mechanisms are heterogeneous:
  • Localized/Modular: Horizontally transferred gene cassettes (e.g., plasmid-borne $\beta$-lactamases) that are conserved across species boundaries.
  • Diffuse/Chromosomal: Species-specific mutations in regulatory genes or membrane permeability that rely on the specific genomic background of a species.
• The Failure of Standard Methods: Traditional $k$-mer based methods (like Kover) and standard foundation model pipelines often rely on "phylogenetic shortcuts" (learning species-specific background signals like GC content or codon bias) rather than causal resistance mechanisms. Consequently, they achieve high accuracy within-species but collapse under strict cross-species evaluation.
• The Scale Problem: Genomic foundation models (e.g., Evo-1-8k-base) produce massive, high-dimensional embeddings (4,096 dimensions per token). A typical bacterial genome requires thousands of windows, resulting in millions of raw features, making naive downstream modeling computationally impractical and prone to signal dilution.

2. Methodology

The thesis proposes a diagnostic-driven framework with two primary innovations to address the problem:

A. Diagnostic-Driven Layer Selection

Instead of defaulting to the final layer of the foundation model (Evo-1-8k-base), the author developed a diagnostic framework to identify the optimal extraction layer under native bfloat16 (bf16) inference.

• Diagnostics: The study analyzed activation scales, isotropy (angular diversity), effective rank, and cross-seed stability across all 32 layers.
• Findings: A sharp stability boundary was identified at Layer 11 (L11). Beyond this layer, the model exhibits:
  • Anisotropy: Collapse of angular diversity.
  • Compression: Reduction in effective rank (singular spectrum compression).
  • Numerical Instability: Massive residuals and "attention sinks" where a few tokens dominate the activation space, exacerbated by bf16 precision limits.
• Solution: Layer 10 (L10) was identified as the deepest jointly stable layer, offering the best balance of transferability, geometric richness, and numerical stability.

B. Local Pattern-Preserving Aggregation

The thesis challenges the standard practice of Global Pooling (mean, std, min, max, etc.), which treats the genome as a bag of words and dilutes localized signals.

• Hypothesis: Resistance mechanisms like plasmid-borne cassettes are spatially localized. Global averaging obscures these sparse but critical signals.
• Solution: The author treats the sequence of L10 token embeddings as an ordered multivariate signal and applies MiniRocket (a time-series classification method).
  • Mechanism: MiniRocket uses random binary convolutions to detect local patterns and summarizes them using the Proportion of Positive Values (PPV).
  • Outcome: This preserves cassette-scale patterns (e.g., a 3kb $\beta$-lactamase cassette) while down-weighting diffuse species-specific background noise.

C. Evaluation Protocol

• Strict Species Holdout: The study employs a rigorous Leave-One-Species-Out (LOSO) protocol where training and test sets share zero phylogenetic overlap.
• Dataset: 3,388 genomes from 126 species, focusing on Ampicillin resistance (with analysis of 6 antibiotics total).
• Baselines: Compared against Kover (rule-based $k$-mer learner) and standard Global Pooling pipelines.

3. Key Results

Performance Discrepancies and Mechanism Dependence

The results revealed that no single aggregation strategy universally dominates; performance is mechanism-dependent:

• Cassette-Mediated Resistance: For species where resistance is driven by horizontally transferred elements (e.g., Acinetobacter baumannii, Pseudomonas aeruginosa), MiniRocket significantly outperforms Global Pooling.
  • Example: On the val_outside split, MiniRocket with k-NN achieved an MCC of 0.753, whereas Global Pooling k-NN scored 0.148.
• Chromosomal/Diffuse Resistance: For species relying on chromosomal mutations or diffuse mechanisms (e.g., Enterobacter hoffmannii), Global Pooling often performs better or comparably.
  • Example: On the test_outside split (dominated by chromosomal mechanisms), Global Pooling with LightGBM achieved an MCC of 0.932, outperforming MiniRocket.
• Comparison to Kover: Both foundation model approaches vastly outperformed the Kover baseline under cross-species evaluation, where Kover's F1 scores dropped to near-zero in many splits due to reliance on lineage-specific $k$-mers.

Geometric Reorganization and k-NN

• Feature Space Shift: MiniRocket reorganizes the feature space such that genomes cluster by shared resistance modules rather than phylogenetic distance.
• k-NN Phenomenon: After MiniRocket transformation, simple k-Nearest Neighbors (k-NN) became the top-performing classifier (MCC 0.753), whereas it performed poorly with Global Pooling. This indicates that the local pattern preservation creates a geometry where "nearest neighbors" are mechanistically similar, not just phylogenetically similar.
• Neighbor Auditing: Analysis showed that MiniRocket reduces "phylogenetic hubness" (where test samples incorrectly cluster around a dominant training species) and redirects neighbors toward species with shared resistance cassettes (AMR hubs).

4. Key Contributions

1. Species Holdout Protocol: Established a rigorous, leakage-resistant benchmarking infrastructure for AMR prediction that enforces zero phylogenetic overlap between training and test sets.
2. Layer Selection Diagnostics: Developed a framework to identify Layer 10 as the optimal extraction point for Evo-1-8k-base under bf16, preventing the degradation caused by attention sinks and numerical instability in deeper layers.
3. Local Pattern Aggregation: Introduced MiniRocket for genomic embeddings, demonstrating that treating embeddings as ordered signals preserves critical localized resistance mechanisms that global pooling dilutes.
4. Mechanism-Mix Hypothesis: Provided empirical evidence that cross-species generalization is not a monolithic problem but depends on the dominant resistance mechanism (cassette vs. chromosomal) of the target species.

5. Significance and Implications

• Biological Insight: The work demonstrates that successful cross-species prediction requires matching the aggregation strategy to the biological mechanism. Local pattern preservation is essential for modular, horizontally transferred resistance, while global pooling suffices for diffuse chromosomal resistance.
• Interpretability: By enabling simple k-NN classifiers to perform well, the method allows for neighbor auditing, providing a transparent way to understand why a prediction was made (i.e., "this genome is similar to E. coli because they share a specific plasmid").
• Clinical Relevance: The findings suggest that for pathogens known to carry plasmid-borne resistance (a major clinical concern), local pattern preservation is critical. The framework offers a reproducible path to deploying genomic foundation models in real-world clinical settings where novel pathogens are encountered.
• Computational Efficiency: Identifying Layer 10 allows for streaming extraction without storing full layer stacks, reducing GPU memory requirements by an order of magnitude.

In conclusion, the thesis argues that effective use of genomic foundation models in biology requires a deep understanding of both the model's computational properties (numerical stability, layer depth) and the biological structure of the task (modularity of resistance mechanisms).