Abstract: Machine learning has been increasingly used for protein engineering. However, because the general sequence contexts they capture are not specific to the protein being engineered, the accuracy of existing machine learning algorithms is rather limited. Here, we report ECNet (evolutionary context-integrated neural network), a deep-learning algorithm that exploits evolutionary contexts to predict functional fitness for protein engineering. This algorithm integrates local evolutionary context from homologous sequences that explicitly model residue-residue epistasis for the protein of interest with the global evolutionary context that encodes rich semantic and structural features from the enormous protein sequence universe. As such, it enables accurate mapping from sequence to function and provides generalization from low-order mutants to higher-order mutants. We show that ECNet predicts the sequence-function relationship more accurately as compared to existing machine learning algorithms by using ~50 deep mutational scanning and random mutagenesis datasets. Moreover, we used ECNet to guide the engineering of TEM-1 β-lactamase and identified variants with improved ampicillin resistance with high success rates.
Introduction
Protein engineering aims to design variants with improved functions, such as enhanced enzyme activity or thermal stability. Traditionally, this is achieved through directed evolution—an iterative process of random mutation and screening. However, the sequence space is astronomically vast; a protein of just 100 amino acids has more possible variants than there are atoms in the universe, making manual search impossible.
Machine learning (ML) offers a solution by predicting high-fitness variants. Yet, existing models suffer from a "specificity gap." Unsupervised models capture broad evolutionary trends but lack protein-specific fitness data, while supervised models often fail to generalize to higher-order mutants because they do not account for epistasis—the complex interactions between mutations at different sites.
The ECNet Framework
ECNet (Evolutionary Context-integrated neural network) bridges this gap by integrating two distinct types of evolutionary information into a supervised Bidirectional Long Short-Term Memory (BiLSTM) neural network architecture.
As shown in the architecture diagram, the model processes information through two parallel paths:
- Local Evolutionary Context: Utilizing Direct Coupling Analysis (DCA) on Multiple Sequence Alignments (MSAs), the model explicitly captures residue-residue dependencies specific to the protein of interest. This serves as a mathematical proxy for epistasis.
- Global Evolutionary Context: The model incorporates representations from large-scale Protein Language Models (LMs) pre-trained on millions of sequences from the "protein universe," learning the general biological "grammar."
These features are combined using an attention mechanism and fine-tuned on quantitative fitness data to map sequences directly to functional measurements.
Epistasis and Generalization
A core insight of the ECNet framework is that local co-evolutionary signals are highly predictive of mutational epistasis. By explicitly feeding these residue-residue dependencies into the model, ECNet learns why certain combinations of mutations are synergistic while others are deleterious. This allows the model to extrapolate from simple single-mutation training data to accurately predict the fitness of complex, higher-order mutants.
Performance and Experimental Validation
ECNet was validated using approximately 50 Deep Mutational Scanning (DMS) datasets. It consistently outperformed state-of-the-art unsupervised models (like EVmutation) and supervised models (like UniRep) across diverse protein families.
Key results include:
- Higher-Order Prediction: When trained only on single-mutant data, ECNet predicted double-mutant fitness with Spearman correlations ranging from 0.73 to 0.94.
- Wet-Lab Success: The authors applied ECNet to engineer TEM-1 $\beta$-lactamase, identifying novel variants with up to 8-fold improved fitness over the wild type. The "hit rate" for identifying improved variants reached as high as 94% in high-concentration antibiotic screens.
Conclusion and Future Directions
ECNet demonstrates that integrating local and global evolutionary priors significantly enhances the accuracy and generalizability of protein fitness predictions. While the model currently depends on high-quality MSAs and focuses on point mutations, it represents a major step forward in data-efficient protein engineering. Future work will aim to address insertions and deletions (indels) and further reduce the amount of initial experimental data required for fine-tuning.
Summary updated