PanTEon: a cross-kingdom framework to guide the design of transposable element classifiers

The paper introduces PanTEon, a cross-kingdom deep learning framework comprising a harmonized database and a modular benchmarking platform that enables reproducible, standardized training and evaluation of transposable element classifiers across diverse eukaryotic lineages.

The Gist

Imagine your genome (your body's instruction manual) is a massive, ancient library. Most of the books in this library are well-written stories called "genes." But scattered throughout the shelves are millions of sticky notes, torn pages, and random scribbles that keep copying themselves and sticking to new pages. These are Transposable Elements (TEs).

For a long time, scientists called these scribbles "junk DNA" and ignored them. But now we know they are actually the library's chaotic architects—they shape how the library grows, evolves, and sometimes even causes problems. The trouble is, because these scribbles copy themselves so fast and mutate so much, they look different in every species. Trying to sort them out is like trying to organize a library where every book is written in a different language, half the pages are missing, and the titles keep changing.

Enter PanTEon: The Ultimate Librarian's Toolkit.

This paper introduces PanTEon, a new, all-in-one system designed to help scientists finally organize this chaos. Think of it as a "Swiss Army Knife" for genome librarians, built with two main parts:

1. The PanTEon Database: The "Master Reference Collection"

Before PanTEon, scientists had to hunt for TE examples in scattered, messy databases. Some were free but unorganized; others were high-quality but locked behind paywalls. It was like trying to learn a language using only a few scattered dictionary pages from different countries.

PanTEon fixes this by gathering 240,000 high-quality TE examples from over 2,700 different species (animals, plants, and fungi).

• The Analogy: Imagine a massive, perfectly organized museum where every single type of "scribble" from every corner of the animal and plant kingdom is displayed, labeled, and cleaned up.
• The Magic: The team didn't just copy-paste these; they used a robot (an automated curation process) to check every single one, ensuring they were complete and correctly labeled. This creates a "Gold Standard" reference that anyone can use to train their own sorting machines.

2. The PanTEon Platform: The "Tournament Arena"

Even with a great library, you need a way to test which sorting method works best. Currently, there are many different AI tools (computer programs) that try to sort these TEs, but they all speak different languages and use different rules. Comparing them was like trying to race a Ferrari, a bicycle, and a boat on the same track without a standard finish line.

PanTEon provides a standardized arena where these AI tools can compete fairly.

• The Arena: It takes a pile of messy TE sequences and asks, "What is this?"
• The Contestants: It runs nine different top-tier AI classifiers (the "Ferraris" of the TE world) against the same test.
• The Results: It tells you exactly who wins, who loses, and under what conditions.

What Did They Discover? (The Plot Twist)

When they ran the tournament, they found some surprising things:

• The "One-Size-Fits-All" Myth: They thought a single AI model could sort TEs for all life forms. It turns out, that's like trying to use a single recipe to cook a steak, a salad, and a cake. The AI performed great on animals and plants but struggled terribly with fungi (mushrooms and molds). Why? Because the training data for fungi was too scarce.
• The Power of Specialization: When they trained the AI specifically on just fungi or just insects, the performance skyrocketed. It's like hiring a specialist who only knows how to sort mushrooms, rather than a generalist who tries to know everything.
• The Teamwork Effect: They tried a "team strategy" (Ensemble learning), where they let all the AI tools vote on the answer. The result? The team was smarter than any single expert. By combining their opinions, they got much more accurate results, especially for the tricky fungal cases.
• Speed vs. Smarts: Some tools were incredibly fast but slightly less accurate. Others were slow but brilliant. PanTEon helps users choose the right tool for their specific job (e.g., "I need speed" vs. "I need maximum accuracy").

Why Does This Matter?

Think of PanTEon as the foundation for the next generation of genome science.

• For Scientists: It stops them from reinventing the wheel. They can stop arguing about which tool is best and start using the PanTEon framework to build better tools.
• For the Future: It allows us to finally understand the "junk" in our DNA. By sorting these elements correctly, we can learn how species evolve, how diseases happen, and how life adapts to its environment.

In a Nutshell:
PanTEon is a massive, clean library of genetic "scribbles" combined with a fair testing ground for AI. It proves that while sorting the chaos of our DNA is hard, having the right tools and the right data makes it possible to turn a messy library into a well-organized masterpiece.

Technical Summary

1. Problem Statement

Transposable elements (TEs) are critical drivers of genome evolution, yet their annotation and classification remain inconsistent, difficult to reproduce, and highly fragmented across species. Key challenges include:

• Data Heterogeneity: Existing TE databases (e.g., Repbase, Dfam) are often incomplete, require payment, lack curation, or are taxonomically biased (heavily skewed toward plants and animals, with poor fungal representation).
• Methodological Inconsistency: TE classification tools use disparate training datasets, nomenclatures, and evaluation metrics, making direct comparison impossible.
• AI Limitations: While Deep Learning (DL) and Machine Learning (ML) have improved TE classification, there is no standardized framework to benchmark these models fairly, nor is there a large, high-quality, automatically curated dataset to train robust, cross-kingdom models.
• Generalization Failure: Current models often fail to generalize across different taxonomic kingdoms (e.g., performing well on plants but poorly on fungi) due to training data bias.

2. Methodology

The authors developed PanTEon, a comprehensive framework consisting of two main components: a curated database and a modular software platform.

A. The PanTEon Database

• Scale & Diversity: Contains 239,946 automatically curated TE sequences from 2,790 species across three kingdoms: Animalia, Plantae, and Fungi.
• Data Sources: Integrated curated sequences from Dfam 3.9, uncurated Dfam sequences, libraries from APTEdb, Ensembl 2025, and the B-GUT database (specifically for fungi).
• Curation Pipeline: Used RepeatModeler2 (RM2) for initial detection and MCHelper for automatic curation. Sequences were filtered to ensure structural completeness (e.g., requiring LTRs for retrotransposons and specific domain counts).
• Standardization: All sequences were renamed and classified into a consistent hierarchical format (Class / Order / Superfamily) following Wicker et al. (2007) taxonomy.

B. The PanTEon Platform

A modular software framework (Python-based, using TensorFlow/PyTorch) designed for:

1. Benchmarking: Parallel execution and comparison of multiple ML/DL classifiers on standardized datasets.
2. Training: Retraining existing models or training new custom architectures on user-defined datasets.
3. Inference: Predicting TE classifications for new sequences.
4. Library Generation: Extracting specific TE subsets based on taxonomy or classification levels.
5. Extensibility: Allows users to integrate custom ML/DL architectures by placing Python scripts in a designated folder.

C. Experimental Design

• Benchmarking: Seven state-of-the-art tools (ClassifyTE, CREATE, DeepTE, NeuralTE, TEClass2, TERL, Terrier) were tested on a balanced dataset of ~38,742 sequences.
• Re-training: Nine architectures (the seven above plus Inpactor2_Class and BERTE) were re-trained from scratch on the PanTEon Database to eliminate training-data bias.
• Ensemble Methods: Three strategies were tested: simple majority voting, weighted voting, and XGBoost stacking.
• Specific Tasks:
  • Cross-kingdom generalization tests.
  • Kingdom/Phylum-specific model training (e.g., models trained only on Fungi or Chordata).
  • Binary classification: Distinguishing TE sequences from non-TE sequences (genes/RNAs) to filter false positives.

3. Key Contributions

1. PanTEon Database: The first large-scale, taxonomically broad, and automatically curated TE repository covering animals, plants, and fungi, addressing the lack of high-quality training data for AI.
2. Standardized Framework: A unified platform that harmonizes training, evaluation, and inference across different ML/DL architectures, enabling fair comparison.
3. Extensibility: A "plug-and-play" architecture allowing the community to easily integrate new classifiers and custom models.
4. Comprehensive Benchmarking: The first independent, large-scale evaluation of TE classifiers across multiple kingdoms and superfamilies, revealing specific biases and performance gaps.

4. Key Results

Performance of Existing Tools

• Kingdom Bias: All tools performed significantly better on Animals and Plants than on Fungi. Fungal classification F1-scores dropped as low as 42% for some tools, highlighting the lack of fungal training data in existing models.
• Top Performers: NeuralTE and Terrier emerged as the best-performing tools.
  • NeuralTE excelled at Class I (retrotransposons) due to its use of structural features and k-mer frequencies.
  • Terrier achieved high speed (0.32 hours) and accuracy by using a simple integer mapping of nucleotides and hierarchical loss functions.
• Ensemble Success: Ensemble approaches, particularly XGBoost stacking, improved F1-scores by 7–13% over the best individual models, especially for under-represented taxa like Fungi.

Re-training and Architecture Analysis

• Taxon-Specific Models: Training models on specific taxonomic groups (e.g., Chordata, Angiosperms) often outperformed general cross-kingdom models. However, for Fungi, specific models still lagged behind general models due to the small size of the fungal dataset.
• Architecture Trends:
  • CNNs generally performed well.
  • Transformers (BERTE, TEClass2) showed poor performance when re-trained on this dataset, likely due to insufficient data volume relative to model complexity (though TEClass2 improved significantly when trained on specific plant data).
  • Feature Engineering: Models using structural features (e.g., terminal repeats) combined with k-mers (NeuralTE) outperformed those relying solely on sequence embeddings.
• Efficiency vs. Accuracy: There was no correlation between model size (parameters) and performance. Small models (e.g., ClassifyTE with 502k parameters) performed well but were computationally expensive due to feature extraction steps.

Binary Classification (TE vs. Non-TE)

• The framework successfully trained models to distinguish TEs from non-TE sequences (genes, RNAs).
• Most models achieved >0.95 F1-score, with NeuralTE reaching 0.99, demonstrating the framework's utility for automating the manual curation step of filtering false positives.

5. Significance and Future Impact

• Standardization: PanTEon establishes a reproducible, quantitative foundation for TE research, moving the field away from inconsistent, manual-heavy workflows.
• Community Resource: By providing a massive, open database and a modular platform, it lowers the barrier for developing new AI tools for TE biology.
• Addressing Bias: The study highlights the critical need for expanding fungal and non-model organism TE datasets to improve cross-kingdom generalization.
• Future Directions: The framework is designed to evolve, supporting future tasks like de novo TE detection and sequence trimming, and facilitating the development of TE-specific foundation models as data scales up.

In conclusion, PanTEon bridges the gap between genomic data availability and AI-driven analysis, providing the necessary infrastructure to transform TE annotation into a mature, standardized, and scalable discipline.