Abstract
Predicting how mutations alter RNA function is critical for understanding viral evolution and public health. This study utilizes a Long Short-Term Memory (LSTM) network trained on high-throughput experimental data from the CPEB3 ribozyme. The model demonstrates strong generalization, accurately predicting the functional impact of triple mutations found in nature after being trained only on single and double mutations. This approach offers a safe, computational alternative to high-risk viral experimentation for identifying "gain of function" mutations.
Introduction
RNA viruses, including SARS-CoV-2 and influenza, evolve rapidly through mutation. Some mutations lead to "gain of function," increasing infectivity or replication efficiency. Predicting these outcomes is difficult because RNA function depends on complex 3D folding patterns derived from its sequence. Traditional experimental methods involve live viruses, posing significant biosafety risks. This research proposes a machine learning framework to learn sequence-to-function relationships using safe, non-infectious ribozymes as a model system.
Methods
The researchers employed the CPEB3 ribozyme, a self-cleaving RNA molecule, to generate a large-scale functional dataset.
- Data Generation: A library of approximately 20,000 sequences was synthesized, covering all possible single and double mutations of the CPEB3 ribozyme.
- Functional Assay: High-throughput RNA sequencing (RNA-seq) measured the "fraction cleaved" for each variant, serving as a quantitative measure of RNA function.
- Model Architecture: An LSTM recurrent neural network was trained to perform regression, mapping nucleotide sequences to their corresponding cleavage fractions.
- Validation: The model was tested on an internal set (1-2 mutations) and an external validation set of 518 triple-mutant sequences found in mammalian genomes to assess its ability to generalize to higher-order mutations.
Results
The LSTM model achieved high predictive accuracy across multiple experimental replicates. In Experiment 1, which focused on single and double mutations, the model reached a Pearson Correlation Coefficient (PCC) of 0.921 and an $R^2$ of 0.848. Crucially, in Experiment 2, the model successfully generalized to triple mutations—an out-of-distribution task—maintaining a PCC of up to 0.924.
While performance varied slightly between replicates (e.g., Replicate 1 in Experiment 2 showed an $R^2$ of 0.648), the overall consistency in Replicates 2 and 3 suggests that the model effectively captures the underlying rules of RNA function.
Discussion
The results highlight the potential of using ribozymes as a safe proxy for studying RNA mutation effects. Unlike models based on structural homology, the LSTM learns local sequence patterns that enable generalization to unseen, higher-order mutations. However, the study also identified limitations related to data sparsity; variants with low experimental copy numbers resulted in noisier ground truth data, which negatively impacted model accuracy.
Conclusion
This research demonstrates that machine learning models trained on high-throughput ribozyme data can accurately predict the functional consequences of RNA mutations, including those found in natural evolution. Future work will focus on applying this framework to viral RNA structures, potentially aiding in vaccine development and the early detection of dangerous viral variants.
Summary updated