Perseus: Lineage-Aware Refinement of Kraken2 Taxonomic Classification for Long Read Metagenomes

Perseus is a lineage-aware framework that utilizes a multi-headed convolutional neural network to refine Kraken2's k-mer-based taxonomic classifications for long-read metagenomes, significantly reducing false positives by modeling spatial and hierarchical consistency to prioritize accurate, lineage-coherent assignments.

The Gist

Imagine you are a detective trying to identify a suspect in a crowded room. You have a massive photo album (a database) of known criminals, and you are looking at a blurry, long surveillance video (a long DNA read) of a person walking by.

The Problem: The "One Bad Clue" Mistake
Currently, the most popular detective tool, called Kraken2, works like a very fast but slightly naive assistant. It scans the video and looks for tiny, specific details—a unique scar, a specific hat, a distinct walk. If it sees a scar that matches a criminal named "John," it immediately shouts, "That's John!"

The problem is that in the microbial world (the room full of suspects), many different people share the same scar or wear the same hat. These are like conserved genes (common body parts like ribosomes) that almost all bacteria have.

If your video is long (a long DNA read), Kraken2 might see a "John" scar in the first 10 seconds, but the rest of the video shows the person is actually wearing a "Mary" hat and walking like "Bob." Because Kraken2 treats every clue as an independent vote, that one "John" scar can trick it into making a confident, but completely wrong, guess. It's like identifying a stranger as your neighbor just because they both own a red bicycle, ignoring the fact that they look nothing else alike.

The Solution: Perseus, the "Context Detective"
Enter Perseus. Think of Perseus not as a new scanner, but as a senior detective who reviews the work of the junior assistant (Kraken2).

Perseus doesn't just look at the clues; it looks at how the clues are arranged along the entire video.

1. It checks the neighborhood: If the video shows a "John" scar, but the next 5 minutes show "Mary" and "Bob" features, Perseus realizes, "Wait, this doesn't make sense. You can't be John if the rest of the evidence says otherwise."
2. It respects the family tree: Perseus understands that even if it can't be 100% sure the person is "John," it might be 100% sure they are a "Smith" (a broader category). Instead of guessing "John" and being wrong, Perseus says, "I'm confident this is a Smith, but I'm not confident enough to say it's John."

How It Works (The Metaphor)
Imagine the DNA sequence is a long train track.

• Kraken2 is a train that stops at every station and asks, "Is this station 'New York'?" If it sees a sign that looks sort of like New York, it stops and says, "We are in New York!" even if the next station is clearly "London."
• Perseus is the conductor who looks at the entire journey. It sees that the train passed through "London," "Paris," and "Berlin" before and after that one confusing sign. It realizes, "This train isn't going to New York; that sign was just a misleading advertisement." So, it corrects the destination to a broader, safer location like "Europe" or "The Western Hemisphere," rather than risking a wrong guess.

Why This Matters
In the world of microbiology (studying tiny bugs in soil, guts, or oceans), we often find bugs that have never been seen before.

• Without Perseus: The computer gets overconfident and says, "This is a new species of E. coli!" when it's actually something completely different. This leads to false discoveries.
• With Perseus: The computer says, "I'm not sure exactly what species this is, but I am 99% sure it belongs to the Enterobacteriaceae family."

The Result
Perseus trades specificity (guessing the exact name) for accuracy (getting the family name right). It's better to be correctly identified as a "Dog" than to be confidently but incorrectly identified as a "Poodle" when you are actually a "Cat."

By using a smart neural network (a type of AI) to look at the pattern of clues rather than just counting them, Perseus cleans up the mess, reduces false alarms, and gives scientists a much more reliable map of the microbial world. It turns a chaotic pile of guesses into a structured, trustworthy report.

Technical Summary

1. Problem Statement

Long-read metagenomic sequencing (e.g., PacBio HiFi, Oxford Nanopore) offers superior assembly contiguity and enables genome-resolved analysis of complex microbial communities. However, accurate taxonomic classification of these long reads remains a significant challenge.

• The Core Issue: Highly scalable k-mer-based classifiers like Kraken2 frequently produce over-specific (false positive) taxonomic assignments when applied to long-read data.
• Root Cause: Kraken2 treats sequences as a "bag of votes" (independent k-mer matches) without modeling the spatial distribution or hierarchical consistency of these matches.
  • In complex microbiomes with taxonomic novelty (where the true organism is missing from the reference database), short, localized conserved genomic regions (e.g., rRNA operons, housekeeping genes, mobile genetic elements) can generate clusters of k-mer matches.
  • These sparse, localized matches can dominate the classification decision, leading to biologically implausible species-level assignments even when the broader sequence context does not support that lineage.
• Limitations of Current Solutions:
  • Simply expanding reference databases is difficult due to the lack of high-quality representatives for many environmental lineages.
  • Kraken2's internal confidence thresholds enforce a blunt precision-recall tradeoff, often reducing the call rate aggressively without distinguishing between true lineage-consistent evidence and spurious localized matches.

2. Methodology: The Perseus Framework

Perseus is a lineage-aware confidence estimation framework that acts as a post-processing layer for Kraken2. It reframes taxonomic classification from a single-rank prediction task to a hierarchical confidence estimation problem.

A. Feature Encoding

Perseus processes the raw output of Kraken2 (k-mer assignments along a sequence) to create a structured, multi-channel representation:

1. Binning: The sequence is divided into non-overlapping bins (e.g., 1 kb).
2. 22-Channel Tensor: For each sequence-taxon pair, Perseus constructs a feature tensor capturing the spatial distribution of k-mer evidence across 22 channels. These channels categorize evidence based on taxonomic relationships to the candidate taxon ($c$):
   • Raw Evidence: K-mers assigned exactly to $c$.
   • In-Lineage Evidence: K-mers assigned to ancestors of $c$ (supporting the same lineage at higher ranks).
   • Out-of-Lineage Evidence: K-mers assigned to taxa unrelated to $c$ (conflicting signals).
   • Descendant Evidence: K-mers assigned to strict descendants of $c$.
3. Goal: This encoding allows the model to distinguish between a coherent lineage signal (consistent across the sequence) and spurious signals (localized clusters of conserved regions).

B. Model Architecture

• Network Type: A multi-headed 1D Convolutional Neural Network (CNN).
• Process:
  1. Convolutional Layers: Five 1D convolutional blocks extract local spatial patterns of taxonomic consistency across adjacent bins.
  2. Pooling: Masked average pooling generates a fixed-length embedding from variable-length inputs.
  3. Multi-Head Output: The network branches into seven independent output heads, corresponding to the canonical taxonomic ranks (Kingdom, Phylum, Class, Order, Family, Genus, Species).
  4. Training: Uses Focal Loss to handle class imbalance and emphasize difficult examples.
  5. Calibration: Post-hoc isotonic regression is applied to convert raw CNN scores into calibrated probability estimates for correctness at each rank.

C. Decision Logic

Based on the calibrated confidence scores, Perseus makes one of three decisions for a sequence:

1. Confirm: If confidence is high at a fine rank (e.g., species), the assignment is kept.
2. Back-off: If evidence is insufficient for a fine rank but strong for a higher rank (e.g., genus), the assignment is moved up the hierarchy.
3. Abstain: If evidence is too sparse or conflicting, the sequence is marked as unclassified to avoid false positives.

3. Key Contributions

1. Lineage-Aware Modeling: Perseus is the first framework to explicitly model the spatial distribution and hierarchical consistency of k-mer evidence, moving beyond the "bag of votes" approach.
2. Hierarchical Confidence Estimation: Instead of predicting a single fixed rank, it estimates the deepest taxonomic rank supported by coherent evidence, allowing for dynamic "back-off" to higher ranks.
3. Integration: It integrates seamlessly with existing Kraken2 workflows without requiring changes to reference databases or the classifier itself.
4. Mechanistic Insight: The authors provide a detailed analysis showing that misclassifications often stem from short (50–200 bp) conserved regions (rRNA, housekeeping genes) that create false k-mer clusters, validating the need for spatial modeling.

4. Results and Evaluation

The authors evaluated Perseus on simulated datasets (with varying degrees of taxonomic novelty), CAMI II benchmarks (marine and plant-associated), and real-world data (ZymoBIOMICS and German soil metagenomes).

• Reduction in False Assignments: Perseus consistently and substantially reduces the False Assignment Rate (FAR). In family-exclusion scenarios (simulating high novelty), Kraken2's FAR approached 50%, while Perseus significantly lowered this.
• Improved Precision: The Lineage-Correct Positive Predictive Value (LC-PPV) increased dramatically, often approaching 0.99.
• Handling Novelty: In scenarios where the true species was absent from the database, Perseus successfully "backed off" to the correct higher taxonomic ranks (e.g., genus or family) rather than forcing an incorrect species-level call.
• Long-Read Advantage: Improvements were most pronounced for long reads and assembled contigs, where the longer spatial context allows the CNN to better discriminate between consistent lineage signals and localized conserved matches.
• Real-World Validation:
  • On the ZymoBIOMICS mock community, Perseus filtered out spurious species-level errors by backing off to the correct genus level.
  • On the Schönbuch Forest soil dataset, Perseus reduced species-level assignments by >2 orders of magnitude (from ~1.1M to ~6k), reassigning the majority to Order or Class levels with much stronger k-mer support (median k-mer support increased from 5 to 187).
• Comparison: Perseus-filtered Kraken2 outperformed other classifiers (Centrifuger, Metabuli) in LC-PPV, FAR, and Lineage-Correct F1-score across nearly all experiments.

5. Significance

• Paradigm Shift: Perseus demonstrates that separating evidence accumulation (Kraken2's strength) from confidence estimation (Perseus's role) significantly improves robustness in metagenomics.
• Biological Realism: It addresses the reality that many environmental microbiomes contain novel taxa. In such cases, a confident, correct assignment at the genus or family level is more scientifically valuable than a high-confidence but incorrect species-level assignment.
• Scalability: The tool is computationally efficient, operates on standard Kraken2 outputs, and is open-source, making it immediately applicable to large-scale metagenomic studies.
• Future Direction: The work suggests that future taxonomic classifiers should move away from single-rank predictions toward hierarchical, confidence-aware frameworks that can explicitly handle taxonomic novelty and reference database gaps.