TL;DR
This paper introduces a novel "bacteria-centric" pipeline for predicting antimicrobial peptides (AMPs) directly from bacterial genomes. By integrating comparative biology and considering peptide structure at multiple levels, it aims to overcome the bias in current AMP prediction tools that heavily rely on eukaryotic data, leading to better discovery of naturally occurring bacterial AMPs.
What Problem Does This Paper Address?
Antimicrobial peptides (AMPs) are a promising alternative to traditional antibiotics due to their broad-spectrum activity and diverse mechanisms. However, despite bacteria being the vast majority of life on Earth, existing AMP databases and prediction tools are heavily biased towards eukaryotic AMPs. This means many naturally occurring bacterial AMPs are likely being missed. The challenge lies in developing prediction methods that are specifically tailored to the unique characteristics of bacterial genomes and AMPs, moving beyond the current reliance on eukaryotic-centric training data and algorithms.
Methods
The authors developed a multi-stage pipeline that integrates various computational approaches to identify candidate AMPs. It starts by extracting open reading frames (ORFs) from bacterial nucleotide sequences and translating them into peptides. Crucially, it then applies a series of filters and analyses:
- Primary Structure Analysis: This includes filtering based on absent tripeptide combinations found in eukaryotic AMPs, physicochemical properties of amino acids, and specific motifs like the double glycine triplet and low-complexity regions.
- Secondary and Tertiary Structure Analysis: The pipeline incorporates predictions of secondary structures (alpha helices, beta sheets) and uses homology searches against protein databases (like Foldseek) to assess tertiary structure. It also employs frequent itemset mining and binary logistic regression models trained on autocorrelation descriptors.
- Motif Discovery: Multiple sequence alignments of known AMPs are used to discover novel motifs for AMP identification.
The pipeline generates ten batches of AMP candidates, with final candidates being those identified by at least three different methods.
Datasets
- Antimicrobial Peptide Database (APD): Used as a reference set for validated AMPs, particularly for identifying absent tripeptide combinations and for motif discovery.
- Real Bacterial Genomes: A set of 27 bacterial genomes known to produce AMPs were used for evaluation.
- Fake Bacterial Genomes: Control genomes with randomized nucleotides were generated to mimic the size and composition of real genomes, serving as a baseline for comparison.
- ESKAPE Pathogens: Genomes of seven multi-drug resistant nosocomial pathogens were specifically analyzed.
Key Insights and Ideas
- The "Eukaryotic Bias" is a Major Hurdle: The vast majority of known AMPs are from eukaryotes, leading current prediction tools to overlook many bacterial AMPs. This paper directly addresses this by creating a "bacteria-centric" approach.
- Multi-Level Structural Analysis is Key: AMPs are not just defined by their amino acid sequence. Considering primary, secondary, and tertiary structural properties provides a more comprehensive and robust prediction.
- Motifs and Physicochemical Properties Matter: Specific amino acid motifs and physicochemical properties are strong indicators of AMP activity, and leveraging these from bacterial contexts is crucial.
- Comparative Biology as a Design Principle: The pipeline's design is inspired by the sequence properties of bacterial genomes, rather than solely relying on statistical machine learning models trained on existing, biased datasets.
Key Results
- The pipeline successfully identifies known AMPs that are missed by state-of-the-art AMP predictors.
- It yields a significantly higher number of AMP candidates from real bacterial genomes compared to fake genomes.
- AMP detection rates are notably higher in the genomes of seven nosocomial pathogens than in fake genomes.
- When compared to amPEPpy (a highly-rated AMP predictor), the pipeline shows a different prediction profile, suggesting complementarity. For instance, the pipeline identified validated APD AMPs that amPEPpy missed (observed positive rate of 16.3% vs. 91.7% for amPEPpy on the same set).
- The pipeline's output on real genomes is more likely to be predicted as AMPs by AmpGram (AUC=0.532) than its output on fake genomes (AUC=0.416).
Limitations and Future Work
- Bias in Existing Databases: While the pipeline aims to overcome the eukaryotic bias, it still uses the APD (which is biased) for some filtering and motif discovery.
- Computational Intensity: The multi-stage pipeline, involving various software and analyses, is likely computationally intensive.
- Future Work: The authors suggest that their approach can serve as a complementary tool to existing machine learning-based AMP predictors, potentially improving overall AMP discovery by combining statistical and biologically-inspired methods. Further refinement of the pipeline's components and exploration of more advanced machine learning techniques within a bacteria-centric framework are implied.
Summary updated