[Paper Review] Cross-Species Antimicrobial Resistance Prediction from Genomic Foundation Models

Huilin Tai|arXiv (Cornell University)|Mar 11, 2026

Antibiotic Use and Resistance0 citations

TL;DR

The thesis studies cross-species AMR prediction with genomic foundation models, showing that extracting Layer 10 embeddings and using MiniRocket for local-pattern aggregation enables transfer mainly for cassette-mediated resistance, with performance depending on resistance mechanism.

ABSTRACT

Cross-species antimicrobial resistance (AMR) prediction is fundamentally an out-of-distribution (OOD) generalization problem: models trained on one set of bacterial taxa must transfer to phylogenetically distinct genomes that may rely on different resistance mechanisms. Across species, resistance arises from a heterogeneous mixture of localized, horizontally transferred gene cassettes and diffuse species-specific genomic backgrounds, making successful transfer inherently mechanism-dependent. Using a strict species holdout protocol, we first establish an interpretable k-mer baseline with Kover and show that strong within-species performance collapses under true cross-species evaluation. This motivates representation-level approaches that preserve transferable biological signals rather than amplify phylogenetic shortcuts. We investigate genomic foundation model embeddings derived from Evo-1-8k-base and introduce diagnostics for layer selection based on activation scale, isotropy, effective rank, and cross-seed stability under native bfloat16 inference. These analyses identify a stability boundary in deeper layers and reveal that embeddings extracted near this boundary provide more robust representations for downstream prediction. To preserve localized resistance signals, we treat per-window embeddings as an ordered multivariate signal and apply MiniRocket to summarize multi-scale local activation patterns instead of relying on global pooling. Our results show that aggregation strategy plays a central role in cross-species AMR prediction and that preserving local activation patterns substantially improves generalization when resistance mechanisms are localized.

Motivation & Objective

• Address how to achieve cross-species AMR prediction using genomic foundation models under distribution shift.
• Identify which layer and aggregation strategy preserve transferable resistance signals.
• Understand how resistance mechanism type affects cross-species generalization.
• Provide a reproducible, leakage-resistant evaluation framework for AMR models.

Proposed method

• Use Evo genomic foundation model embeddings and extract from Layer 10 based on diagnostics for stability and isotropy.
• Treat per-window embeddings as a multivariate signal and apply MiniRocket for local-pattern aggregation.
• Benchmark against Kover baseline using strict species holdout partitions to assess true cross-species generalization.
• Filter and curate a multi-species dataset with five partitions and three replicates to enable robust evaluation.
• Compare aggregation strategies (MiniRocket vs Global Pooling) across cassette-mediated vs chromosomal resistance mechanisms.

Figure 2.1: Category coverage among the top 30 targets, before (left) and after (right) filtering. The raw data shows extreme imbalance, with many antibiotics lacking sufficient species diversity. Post-filtering retains only targets that preserve diversity across all five partitions.

Experimental results

Research questions

• RQ1 Can genomic foundation-model embeddings transfer to unseen species under strict species holdout?
• RQ2 Which layer and aggregation strategy best preserve transferable resistance signals across species?
• RQ3 How does the dominant resistance mechanism (cassette-mediated vs chromosomal) influence cross-species generalization?
• RQ4 What is the role of model interpretability and neighbor auditing in cross-species AMR predictions?

Key findings

• Kover shows strong within-species performance but substantial cross-species degradation under species holdout across six antibiotics.
• Layer 10 embeddings are the deepest jointly stable extraction point, with Layer 11 showing instability; layer diagnostics identify this stability boundary.
• MiniRocket with k-NN generally yields the highest cross-species MCC (0.753) on the val_outside split for ampicillin, outperforming the k-NN with Global Pooling.
• Across ampicillin, cassette-mediated resistance leads to larger cross-species gains; chromosomal/diffuse resistance is harder to transfer.
• On the independent test_outside set, Global Pooling with linear or tree-based classifiers often matches or surpasses MiniRocket, depending on the resistance mechanism.
• Cross-species predictions improve when resistance signals are localized (cassette-mediated) and preserved during aggregation; same-species performance remains similar between methods.

Figure 2.2: Sample counts among the top 30 targets, before (left) and after (right) filtering. Many antibiotics in the raw data have insufficient samples for stable evaluation. Post-filter sets maintain adequate per-partition counts to yield reliable confidence intervals.