Abstract
Generative biology aims to transform our ability to design novel proteins and pathways, yet its progress is hindered by a lack of data capturing the complex genetic architecture of protein function. In this study, researchers systematically mutagenized all 54 human basic-leucine zipper (bZIP) domains, quantifying their interactions through approximately 2 million measurements using a quantitative deep mutational scanning assay (bindingPCA). The results reveal a "local simplicity, global complexity" architecture: while mutational effects are largely additive within the vicinity of a specific wild-type bZIP, they diverge significantly across the family due to strong context dependency. Despite this complexity, a convolutional neural network (CNN) trained on this family-wide dataset accurately predicted binding scores and enabled the design of highly specific synthetic bZIPs. This work demonstrates that capturing family-wide diversity is essential for building accurate models of protein-protein interactions (PPIs).
Introduction
The ability to design proteins from scratch—generative biology—holds immense potential for medicine and biotechnology. However, a fundamental challenge remains: understanding how the effect of a mutation depends on the rest of the protein sequence, a phenomenon known as epistasis. Most existing studies focus on single proteins, missing the context dependencies that define how entire protein families function.
To address this, the researchers mapped the interaction landscape of the human basic-leucine zipper (bZIP) family. Comprising 54 transcription factors, the bZIP family is an ideal model system due to its structural regularity (coiled-coil) and critical role in cellular regulation. By mapping nearly 2 million mutations, the study reveals the underlying genetic architecture of this family and provides a roadmap for training predictive models in generative biology.
Experimental Mapping of the bZIP Landscape
The researchers employed a high-throughput approach combining experimental biology with quantitative modeling:
- Deep Mutational Scanning (DMS) and bindingPCA: The team systematically mutated all 35 interfacial positions across 54 human bZIP domains to all 20 amino acids. They used bindingPCA (bPCA), a quantitative assay where yeast growth serves as a proxy for protein interaction strength. In this system, interacting bZIPs reconstitute the DHFR enzyme, allowing yeast to survive in the presence of a selection drug.
- Thermodynamic Modeling: To determine if mutational effects were additive, the researchers used a two-state thermodynamic model to convert binding scores into free energy changes (ΔG).
- Data Scale: The resulting dataset includes ~2 million quantitative interaction measurements, capturing the effects of single amino acid substitutions across 53 human bZIPs (excluding ATF7 due to data quality). This includes a baseline map of 2,916 possible wild-type bZIP pairs.
Local Simplicity vs. Global Complexity
A central finding of the study is the distinction between local and global genetic architectures. Within the context of a single bZIP protein, mutations were found to be largely additive, meaning their effects could be predicted by summing the contributions of individual residues. However, when the same mutations were examined across different bZIP homologs, their effects shifted significantly.
This "context dependency" explains why a global additive model—which assumes a mutation has the same effect regardless of the background—explained only 47% of the variance in the data. In contrast, individual models tailored to each bZIP explained 88% of the variance. This suggests that while the rules of interaction are simple at a local level, they are complex and highly specific at the family-wide level.
Predictive Modeling and De Novo Design
The researchers demonstrated that this complexity is not an insurmountable barrier to prediction, provided the training data is sufficiently diverse. A 1D Convolutional Neural Network (CNN) trained on the family-wide dataset achieved an $R^2$ of 0.88 on held-out test data.
To test the practical utility of this model, the team used a genetic algorithm paired with the CNN to design "de novo" synthetic bZIPs. These synthetic proteins were significantly different from any natural human sequence (up to 59 amino acids different) yet bound their intended targets with high specificity. The experimental validation of 80 such designs matched the model's predictions almost perfectly, proving that the landscape is highly predictable when family-wide diversity is captured.
Discussion and Implications
The study reveals that the human bZIP interaction network is highly "rewirable." Most bZIP pairs do not naturally interact but are only one or two mutations away from doing so—a state the authors call the "edge of binding." This proximity allows for rapid evolutionary innovation but also necessitates high specificity to prevent crosstalk.
While the structural regularity of bZIPs may contribute to the observed "local simplicity," the study's primary takeaway is universal: high-accuracy protein design requires massive datasets that cover entire families to reveal context dependencies. Future work may involve integrating these large-scale datasets with Large Language Models (LLMs) like ESM2 to further improve generalization across even more complex, globular protein families. This study marks a significant step toward a future where we can design specific protein interactions with mathematical precision.
Summary updated