Benchmarking Machine Learning Methods for Antimicrobial Peptide Prediction and Discovery in Ancient Microbiomes

Antimicrobial peptides (AMPs) offer a promising alternative to conventional antibiotics amid the growing global threat of antimicrobial resistance. Expanding AMP discovery requires both effective computational tools to accurately identify candidate peptides and access to underexplored biological resources that may harbor novel AMPs. However, the comparative performance of existing computational methods remains poorly understood, while ancient microbiomes have received limited attention as potential reservoirs of AMPs. Here, we systematically benchmarked six machine learning methods across with twelve feature configurations spanning sequence, structural, and protein language model features under a unified framework using eight curated AMP datasets. The best-performing methods were subsequently applied to three ancient metagenomic datasets representing the human gut (archaeological fecal sediments), human oral cavity (dental calculus), and mammoth-associated communities (teeth, skulls, and skin). In parallel, we employed reinforcement learning to select candidate peptides directly from ancient sequence pools without relying on modern AMP labels. Our benchmarking shows that protein language model embeddings provide the most informative features, and deep learning models, including convolutional, transformer-based, and recurrent architectures, outperform classical methods. Applying these models to ancient microbiomes identified candidates occupied the physicochemical space characteristic of AMPs but differed from those of candidates predicted from modern microbiome catalogs. We identified 111, 76, and 50 high-confidence candidate AMPs from the Neanderthal gut, ancient oral, and mammoth-associated communities, respectively. Reinforcement learning selected additional candidate AMPs with distinct properties compared with those predicted by models trained on modern AMP databases. Together, these findings establish a systematic benchmark for AMP prediction and highlight ancient microbiomes as an untapped source of antimicrobial peptide candidates. By integrating supervised machine learning with reinforcement learning, this study expands AMP discovery beyond modern microbiome catalogs and provides new opportunities to identify potentially novel AMPs.

Authors

Institutions

Publication Details

Journal
bioRxiv (Cold Spring Harbor Laboratory)
Published
2026-11-18
DOI
https://doi.org/10.64898/2026.11.17.767431
Primary Topic
Antimicrobial Peptides and Activities
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Benchmarking Machine Learning Methods for Antimicrobial Peptide Prediction and Discovery in Ancient Microbiomes

Shanlin Ke, Can Chen, Qin Ma, Jie Luo et al.
bioRxiv (Cold Spring Harbor Laboratory)
Antimicrobial Peptides and Activities
preprint

Benchmarking Machine Learning Methods for Antimicrobial Peptide Prediction and Discovery in Ancient Microbiomes

Shanlin Ke, Can Chen, Qin Ma, Jie Luo, Quan Zhao
preprint en

Abstract

Antimicrobial peptides (AMPs) offer a promising alternative to conventional antibiotics amid the growing global threat of antimicrobial resistance. Expanding AMP discovery requires both effective computational tools to accurately identify candidate peptides and access to underexplored biological resources that may harbor novel AMPs. However, the comparative performance of existing computational methods remains poorly understood, while ancient microbiomes have received limited attention as potential reservoirs of AMPs. Here, we systematically benchmarked six machine learning methods across with twelve feature configurations spanning sequence, structural, and protein language model features under a unified framework using eight curated AMP datasets. The best-performing methods were subsequently applied to three ancient metagenomic datasets representing the human gut (archaeological fecal sediments), human oral cavity (dental calculus), and mammoth-associated communities (teeth, skulls, and skin). In parallel, we employed reinforcement learning to select candidate peptides directly from ancient sequence pools without relying on modern AMP labels. Our benchmarking shows that protein language model embeddings provide the most informative features, and deep learning models, including convolutional, transformer-based, and recurrent architectures, outperform classical methods. Applying these models to ancient microbiomes identified candidates occupied the physicochemical space characteristic of AMPs but differed from those of candidates predicted from modern microbiome catalogs. We identified 111, 76, and 50 high-confidence candidate AMPs from the Neanderthal gut, ancient oral, and mammoth-associated communities, respectively. Reinforcement learning selected additional candidate AMPs with distinct properties compared with those predicted by models trained on modern AMP databases. Together, these findings establish a systematic benchmark for AMP prediction and highlight ancient microbiomes as an untapped source of antimicrobial peptide candidates. By integrating supervised machine learning with reinforcement learning, this study expands AMP discovery beyond modern microbiome catalogs and provides new opportunities to identify potentially novel AMPs.

bioRxiv (Cold Spring Harbor Laboratory)
University of North Carolina at Chapel Hill (US), Purdue University West Lafayette (US), The Ohio State University (US)
Antimicrobial Peptides and Activities
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon