Abstract
In further study, extended vision (or k-mer) is shown to be essential for EVMP. We also found that EVMP can alleviate the over-smoothing phenomenon, which may contributes to its effectiveness. Our work suggests that EVMP can highlight the mutation information of synthetic promoters and significantly improve the prediction accuracy of strength. The source code is publicly available on GitHub: https://github.com/Tiny-Snow/EVMP.
Introduction
In the field of synthetic biology, promoters act as the "volume knobs" of gene expression, determining the rate at which genetic information is transcribed. To optimize metabolic pathways, researchers often construct massive libraries of synthetic promoters. However, characterizing the strength of thousands of variants through traditional "wet lab" experiments is both time-consuming and expensive.
While Machine Learning (ML) offers a promising alternative for predicting promoter strength, existing models struggle with a phenomenon known as excessive proximity. Synthetic libraries are typically generated by introducing minor mutations into a common "base" sequence. Consequently, two promoters may be 99% identical yet exhibit significantly different expression levels. Standard ML models often fail to distinguish these subtle differences, leading to "over-smoothing," where the model predicts nearly identical outputs for highly similar sequences.
The EVMP Framework
To address these challenges, we developed the Extended Vision Mutant Priority (EVMP) framework. EVMP is a universal architecture designed to enhance the performance of various ML models—ranging from traditional algorithms like Random Forests to deep learning architectures like Transformers—by prioritizing mutation information.
The framework operates through several key stages:
- Decomposition: Instead of processing the entire sequence as a single unit, EVMP decomposes it into a Base Promoter and its specific $k$-mer mutations. This "extended vision" ($k$-mer) approach includes the mutated base and its immediate neighbors, providing necessary structural context.
- Dual Encoders: The architecture utilizes two parallel tracks: a BaseEncoder for the background sequence and a VarEncoder that focuses exclusively on the mutations.
- Mutation Representation: Mutations are represented using either Vars+PE (adding positional encoding, suitable for LSTMs) or Masking (hiding non-mutated bases, suitable for XGBoost).
- Data Augmentation: By treating different promoters within a library as potential "bases" for one another, EVMP can generate multiple synthetic views of the same data, significantly increasing the training set's robustness.
Key Insights and Results
Our evaluation, primarily using the Trc synthetic promoter library (3,665 variants), revealed several critical insights:
- Curing Over-smoothing: By analyzing average embedding distances, we demonstrated that EVMP forces models to create distinct mathematical representations for similar sequences, effectively "zooming in" on functional differences.
- Data Efficiency: EVMP-enhanced models are remarkably efficient. A model trained on only 56% of the available data using EVMP can match the performance of a state-of-the-art (SOTA) model trained on the full dataset.
- Performance Breakthroughs: The framework provided substantial gains across all tested models. Notably, the LSTM's Mean Absolute Error (MAE) improved by 61.30%. The EVMP-Transformer established a new SOTA record, outperforming the previous best model by 15.25% in MAE and improving the $R^2$ correlation by 4.03%.
Conclusion
The EVMP framework represents a significant step forward in the computational design of synthetic promoters. By shifting the focus to local mutation neighborhoods, it overcomes the limitations of traditional ML models in handling highly similar biological sequences. Future work will explore applying this "mutant priority" logic to protein engineering and incorporating 3D structural data to further refine mutation representations.
Source Code: https://github.com/Tiny-Snow/EVMP
Summary updated