F5C-finder: An Explainable and Ensemble Biological Language Model for Predicting 5-Formylcytidine Modifications on mRNA

Guohao Wang, Ting Liu, Hongqiang Lyu, Ze Liu

TL;DR

f5C-finder is the first computational model specifically designed to predict 5-formylcytidine (f5C) modifications on mRNA. By combining five different feature extraction methods with an ensemble of LSTM and multi-head attention-based "biological language models," it achieves state-of-the-art performance (AUC of 0.827) while providing interpretable insights into the genetic motifs that drive these modifications.

What Problem Does This Paper Address?

RNA is more than just a simple messenger; it is subject to over 170 different chemical modifications that regulate biological processes—a field known as epitranscriptomics. One such modification, 5-formylcytidine (f5C), is found across various species and is critical for mitochondrial function and embryonic development.

However, mapping these modifications in the lab is a nightmare. Traditional experimental methods like "f5C-seq" are laborious, expensive, and time-consuming. While computational tools exist for other modifications (like m6A or m5C), there has been a total lack of specialized recognition models for f5C. Researchers needed a cost-effective, high-throughput way to scan transcriptomes and identify potential f5C sites before heading to the wet lab.

Methods

The authors treat RNA sequences as a form of "biological language," where nucleotides (A, C, G, U) are like letters and functional motifs are like words. To capture the complexity of this language, they developed f5C-finder, an ensemble model built on two pillars:

  1. Diverse Feature Extraction: Instead of relying on one way to represent RNA, they used five:
  2. One-hot & Binary Encoding: Simple positional mappings.
  3. Hash Decimal Conversion: Captures relationships between adjacent bases.
  4. Codon-based features: Focuses on the frequency of specific trinucleotide patterns.
  5. Position Encoding for Bio-sequence (PEB): Maps nucleotides to numerical values while maintaining their relative order.
  6. Hybrid Architecture: They didn't just use one type of neural network. They built five sub-models:
  7. LSTMs (Long Short-Term Memory): Used for OnehotNet and BinaryNet to capture long-range sequential dependencies.
  8. Multi-Head Attention: Used for PEBNet, HashNet, and FusionCodonNet. This mechanism allows the model to "attend" to specific parts of the sequence that are most relevant to the modification, regardless of distance.

The final f5C-finder is a linear ensemble that averages the predictions of these five sub-models to ensure robustness and higher accuracy.

Datasets

The study utilized a high-resolution f5C dataset from Saccharomyces cerevisiae (yeast).

  • Source: 1,892 f5C sites identified via f5C-seq.
  • Preprocessing: Sequences were trimmed to 101bp (centered on the f5C site). To prevent the model from "cheating" by memorizing similar sequences, they used CD-HIT software with a strict 0.8 threshold to remove redundant data.
  • Final Count: 1,805 positive and 1,805 negative samples. The data was split into a training set (1,444 samples per class) and an independent testing set (361 samples per class).

Key Insights and Ideas

  • The "Neighborhood" Matters: The model's interpretability analysis (using attention heatmaps) revealed that f5C modification isn't random. There is a "conserved grammar" around these sites. Specifically, the motif "TTATTT" is enriched upstream, and "TAAGA" is enriched downstream.
  • Semantic vs. Sequential: By using both LSTMs and Attention mechanisms, the authors show that biological sequences have both "order" (sequential) and "meaning" (functional semantics). The attention mechanism, in particular, was able to highlight that the region immediately surrounding position 51 (the target cytosine) is the most information-rich.
  • Ensemble Superiority: The paper reinforces a common machine learning truth: a "committee" of models looking at the data through different lenses (different feature extractions) consistently outperforms any single "expert" model.

Key Results

f5C-finder set a new benchmark for this specific modification:

  • Performance: In 10-fold cross-validation, it achieved an AUC of 0.807. On the independent test set, this rose to 0.827.
  • Comparison with SOTA: Since no f5C model existed, the authors adapted state-of-the-art (SOTA) models designed for other modifications (like BERT-m7G and i6mA-Vote). f5C-finder outperformed these adapted models across almost all metrics, including Accuracy (ACC), Matthews Correlation Coefficient (MCC), and F1-score.
  • Robustness: The model showed significantly "smoother" ROC and Precision-Recall curves compared to traditional classifiers like SVM or Random Forest, indicating better generalization to unseen data.

Limitations and Future Work

While f5C-finder is a major step forward, the authors acknowledge two main hurdles:

  1. Local vs. Global Context: The model currently only looks at a local 101bp window. In reality, RNA folding and distant structural elements might influence modifications.
  2. Data Scarcity: The model was trained primarily on yeast data. As more experimental f5C data becomes available for humans and other mammals, the model will need to be retrained or fine-tuned to handle species-specific variations.

The authors have made the tool accessible via a webserver and provided the source code on GitHub, inviting the community to build upon this foundation.

Summary updated

Abstract

As a prevalent and dynamically regulated epigenetic modification, 5-formylcytidine (f5C) is crucial in various biological processes. However, traditional experimental methods for f5C detection are often laborious and time-consuming, limiting their ability to map f5C sites across the transcriptome comprehensively. While computational approaches offer a cost-effective and high-throughput alternative, no recognition model for f5C has been developed to date. Drawing inspiration from language models in natural language processing, this study presents f5C-finder, an ensemble neural network-based model utilizing multi-head attention for the identification of f5C. Five distinct feature extraction methods were employed to construct five individual artificial neural networks, and these networks were subsequently integrated through ensemble learning to create f5C-finder. 10-fold cross-validation and independent tests demonstrate that f5C-finder achieves state-of-the-art (SOTA) performance with AUC of 0.807 and 0.827, respectively. The result highlights the effectiveness of biological language model in capturing both the order (sequential) and functional meaning (semantics) within genomes. Furthermore, the built-in interpretability allows us to understand what the model is learning, creating a bridge between identifying key sequential elements and a deeper exploration of their biological functions.
BibTeX
@misc{wang2024f,
  author = {Guohao Wang and Ting Liu and Hongqiang Lyu and Ze Liu},
  title = {{F5C-finder: An Explainable and Ensemble Biological Language Model for Predicting 5-Formylcytidine Modifications on mRNA}},
  year = {2024},
  eprint = {2404.13265},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.13265},
}
Figure