GPAS: an online AI system for rapid and accurate pathogen identification and LLM-based interpretation

The Global Pathogen Analysis System (GPAS) is an accessible online AI platform that combines a novel hybrid machine learning framework for rapid, high-accuracy pathogen identification with a specialized large language model agent to autonomously generate clinically actionable, evidence-based interpretations of complex metagenomic data.

The Gist

Imagine you are a detective trying to solve a crime in a massive, chaotic library. The library contains billions of books (DNA strands) from thousands of different people (microbes). Your job is to find the one specific book that belongs to the criminal (the pathogen causing an illness).

The problem? The library is messy. There are millions of books that look almost identical to the criminal's book (false alarms), and the librarian (current computer software) often gets overwhelmed, shouting out thousands of suspects when there is only one real criminal. Furthermore, even if you find the criminal, the librarian just gives you a list of names without explaining why they are dangerous or what to do about it.

GPAS (Global Pathogen Analysis System) is a new, super-smart detective team designed to solve this exact problem. Here is how it works, broken down into simple parts:

1. The Clean Library (GenoDB)

The Problem: Current libraries are full of duplicate books. If you have 1,000 copies of the same "Criminal Book," the computer gets confused and thinks there are 1,000 different criminals.
The GPAS Solution: The team built a brand-new library called GenoDB. They went through the messy library, threw away all the duplicate copies, and kept only the single, best version of every book. Now, instead of a chaotic mountain of papers, the detective has a neat, organized shelf. This makes the search faster and much less confusing.

2. The Double-Check System (Dynamic Library Alignment)

The Problem: Old detective tools use two different methods. One is very sensitive (catches almost everyone, but includes many innocent people) and the other is very strict (only catches the guilty, but might miss some).
The GPAS Solution: GPAS uses a hybrid team.

• Detective A (Kraken2): "I found 50 suspects! Let's look at all of them!" (High sensitivity).
• Detective B (Sylph): "I only see 2 suspects that are definitely guilty." (High specificity).
• The Chief (The AI Algorithm): The Chief takes the list from both detectives. Using a "cheat sheet" of past mistakes (knowing which books are often confused with each other), the Chief cross-references the lists. If Detective A says "Suspect X" but Detective B says "No," and the cheat sheet says "Suspect X is usually a fake alarm," the Chief removes them.
• The Result: They filter out the noise, leaving only the true criminals with near-perfect accuracy.

3. The "Fingerprint" Check (Genome Coverage Pattern)

The Problem: Sometimes, a computer thinks it found a criminal just because a tiny, random piece of a book matched. It's like finding a single letter "A" and thinking it's the whole word "Apple."
The GPAS Solution: GPAS looks at the whole story.

• If a real pathogen is present, its DNA should be spread out evenly across its entire genome, like a complete book being read from cover to cover.
• If it's a false alarm, the DNA matches are scattered, fragmented, and messy, like finding random words from different books glued together.
• GPAS checks this "reading pattern." If the pattern looks messy, it discards the suspect. If it looks like a complete, coherent story, it confirms the suspect is real.

4. The Expert Translator (The LLM Agent)

The Problem: Even if you find the criminal, the computer just spits out a list of scientific names like Streptococcus pneumoniae. A doctor needs to know: "Is this dangerous? Does it have superpowers (drug resistance)? How does it relate to the patient's fever?"
The GPAS Solution: This is where the AI Agent comes in. Think of it as a brilliant medical translator who speaks both "Microbe" and "Human."

• It connects the list of microbes to a giant Knowledge Graph (a massive web of medical facts, research papers, and drug data).
• It acts like a team of three:
  • The Planner: Figures out what the doctor needs to know.
  • The Researcher: Digs through millions of medical papers to find evidence about the specific microbes found.
  • The Reflector: Double-checks the work to make sure it makes sense.
• The Output: Instead of a raw list, the doctor gets a clear, human-readable report: "We found a bacteria that is likely causing the fever. It is resistant to Penicillin but sensitive to Azithromycin. This matches the patient's history of immune issues."

Real-World Example: The Lupus Patient

The paper tested this on a patient with Systemic Lupus Erythematosus (SLE) who had a fever.

• Old Way: The computer found 2,345 different microbes. The doctor was overwhelmed and couldn't tell which one was the problem.
• GPAS Way: It filtered the list down to just 201 likely microbes. It then used the AI Agent to explain: "The patient's immune system is weak, allowing normal mouth bacteria to overgrow and cause infection. Here is the specific bacteria causing the trouble and how to treat it."

Why This Matters

GPAS is like upgrading from a magnifying glass to a high-tech forensic lab. It takes the confusing, noisy data of modern DNA sequencing and turns it into a clear, actionable story that doctors can use immediately. It lowers the barrier so that any hospital, not just those with super-expensive computer experts, can quickly identify dangerous germs and save lives.

In short: GPAS cleans the library, uses a double-check system to find the real bad guys, checks their fingerprints to make sure they aren't fakes, and then writes a clear report explaining exactly what to do.

Technical Summary

1. Problem Statement

Metagenomic Next-Generation Sequencing (mNGS) is a powerful, hypothesis-free tool for identifying unknown pathogens. However, its routine clinical adoption is hindered by three critical bottlenecks:

• High False Positives & Noise: Current tools struggle to distinguish true pathogens from background noise and highly homologous sequences, leading to excessive false positives.
• Trade-offs in Sensitivity vs. Specificity: Existing classification tools generally follow two paradigms with distinct limitations:
  • k-mer matching (e.g., Kraken2): High sensitivity but prone to misclassification among closely related species due to conserved genomic regions.
  • Sketch-based methods (e.g., Sylph): High specificity but often miss low-abundance taxa (low sensitivity).
• Interpretation Gap: Translating complex microbial lists into clinically actionable insights requires specialized bioinformatics expertise. There is a lack of AI agents capable of integrating multi-source data (host characteristics, clinical literature, microbial ecology) to generate evidence-based diagnostic reports.

2. Methodology

The authors developed GPAS (Global Pathogen Analysis System), an integrated online platform comprising three core technical innovations:

A. GenoDB: A Non-Redundant Microbial Genome Database

• Construction: Built from NCBI RefSeq data (Archaea, Bacteria, Fungi, Viruses, and Hosts).
• Redundancy Reduction: Utilized Average Nucleotide Identity (ANI) clustering (95% threshold) to group intra-species genomes. A single representative genome (highest N50/longest length) was selected per cluster.
• Result: Reduced database size by 90% while maintaining comprehensive species coverage, significantly improving computational efficiency and reducing false positives caused by sequence redundancy.

B. Dynamic Library Alignment (DLA) Algorithm

This is the core engine for species identification, designed to synergize the sensitivity of k-mer methods with the specificity of alignment methods.

1. Preliminary Profiling: Jointly uses Kraken2 (high sensitivity) and Sylph (high specificity) to generate a candidate species list.
2. Hybrid Statistical Inference:
   • Uses pre-computed inter-species misclassification probability matrices derived from 40,000 simulated datasets.
   • Employs a mixed statistical AI model (combining Elastic Neural Networks and Bayesian inference) to calibrate initial assignments.
   • Anchor Species Selection: Identifies high-confidence "anchor" species based on Sylph results and prior reliability categories.
   • Deconvolution: Uses the anchor species to model expected read distributions, filtering out false positives and recalling potential false negatives.
3. Dynamic Alignment: Dynamically extracts reference genomes from GenoDB for the inferred species list and performs precise alignment using Minimap2.

C. Genome Coverage Pattern Recognition

• Hypothesis: True pathogens exhibit non-fragmented, consistent genome coverage patterns across samples, whereas false positives show biased, fragmented distributions.
• Implementation:
  • Constructed a reference library of coverage patterns from 24,164 real-world metagenomic samples.
  • For a target species, the system compares its coverage profile against the reference distribution using statistical tests (Jaccard similarity and Pearson correlation).
  • Outcome: Assigns a statistical confidence score (p-value) to each identification, effectively filtering false positives without sacrificing sensitivity.

D. GPAS-LLM: Pathogen Intelligence Agent

• Knowledge Graph: Integrated 1,242 pathogen species, 10,493 curated articles, and 38.82 million relational triplets into a unified microbial knowledge graph.
• Multi-Agent Architecture:
  • Planner: Parses user goals and decomposes tasks.
  • Researcher: Retrieves evidence from the knowledge graph and bioinformatics tools.
  • Reflector: Performs error checking and feedback optimization.
• Training: Fine-tuned on the curated knowledge graph using Reinforcement Learning to autonomously generate evidence-based, human-readable clinical reports.

3. Key Results

• Superior Accuracy:
  • False Positives: On simulated data (10× depth), GPAS achieved zero false positives for 99.8% of species, compared to 1–5 false positives for other tools (Kraken2, Centrifuger, Ganon2) in >50% of cases.
  • Specific Family Performance: In the Enterobacteriaceae family (highly similar sequences), GPAS reduced false positives to zero in 76.6% of assemblies, whereas Kraken2 averaged ~91 false positives.
  • F1-Score: Achieved an F1-score of 0.925 on the CAMI II marine metagenome dataset, outperforming all benchmarks.
• Confidence Assessment: The genome coverage pattern model removed 96.8% of false positives while retaining 91.2% of true positives in validation datasets.
• Clinical Interpretation (LLM Agent):
  • In a cohort of 100 febrile patients, GPAS-LLM correctly interpreted fever-related symptoms in 91.0% of cases, significantly outperforming DeepSeek V3.2 (61.0%).
  • In 82 confirmed clinical cases, GPAS-LLM identified causative pathogens with 75.6% accuracy vs. 53.7% for DeepSeek V3.2.
• Real-World Application (SLE Case Study):
  • Applied to a throat swab from a Systemic Lupus Erythematosus (SLE) patient.
  • Reduction: Reduced reported species from 2,345 (Kraken2) to 201 (GPAS).
  • Insight: The system identified microbial dysbiosis and pathobiont overgrowth, linking these findings to SLE-associated immune dysregulation and providing a mechanistic explanation for the patient's susceptibility to respiratory infection.

4. Significance and Impact

• Democratization of Diagnostics: By automating the complex bioinformatics pipeline and interpretation, GPAS lowers the technical barrier, making high-precision mNGS accessible to clinicians without specialized bioinformatics training.
• Paradigm Shift: Moves metagenomics from descriptive profiling (listing species) to mechanistic clinical insight (explaining why a pathogen is present and its clinical implications).
• Public Health Utility: The system is freely available online (https://gpas.nh.ac.cn), offering a scalable solution for outbreak response, public health surveillance, and routine clinical diagnostics.
• Future-Proofing: The integration of a dynamic knowledge graph and LLM agents allows the system to evolve with new pathogen discoveries and clinical literature, addressing the limitations of static databases.

In conclusion, GPAS represents a comprehensive computational ecosystem that successfully bridges the gap between raw sequencing data and actionable clinical decisions through algorithmic innovation (DLA, Coverage Patterns) and semantic intelligence (Multi-agent LLM).