Abstract
Structural features of proteins capture underlying information about protein evolution and function, which enhances the analysis of proteomic and transcriptomic data. Here we develop Structural Analysis of Gene and protein Expression Signatures (SAGES), a method that describes expression data using features calculated from sequence-based prediction methods and 3D structural models. We used SAGES, along with machine learning, to characterize tissues from healthy individuals and those with breast cancer. We analyzed gene expression data from 23 breast cancer patients and genetic mutation data from the COSMIC database as well as 17 breast tumor protein expression profiles. We identified prominent expression of intrinsically disordered regions in breast cancer proteins as well as relationships between drug perturbation signatures and breast cancer disease signatures. Our results suggest that SAGES is generally applicable to describe diverse biological phenomena including disease states and drug effects.
Introduction
Traditional transcriptomic and proteomic analyses often rely on gene names or Gene Ontology (GO) terms to interpret biological data. While effective, this "name-based" approach overlooks the physical and structural properties of the encoded proteins. Proteins with different names may share similar structural folds or functional domains, and co-expressed genes in disease states may share structural motifs that explain their collective behavior. To bridge this gap, we developed the Structural Analysis of Gene and protein Expression Signatures (SAGES) framework, which translates expression lists into structural "fingerprints."
The SAGES Framework
The SAGES pipeline converts expression data into a vector of structural features through a multi-step process:
- Feature Generation: Features are derived from sequence-based predictions (e.g., IUPRED2a for intrinsically disordered regions, ANCHOR for binding sites) and 3D structural models (e.g., AlphaFold2 and DSSP for secondary structures like alpha-helices and beta-sheets).
- Normalization: Features are normalized based on input gene expression levels, ensuring that the structural signature reflects the most abundant proteins in the sample.
- Machine Learning: Random Forest models are employed to determine if these structural features can independently identify tissue types or disease states.
Statistical Comparison
To pinpoint significant structural shifts between disease samples (e.g., breast cancer) and healthy backgrounds (e.g., GTEx), the framework utilizes statistical tests to find features with significant log-ratios.
Key Insights
Structural Identity of Tissues
A primary finding of this study is that structural features alone—independent of gene names—can predict tissue type with 97% accuracy. This indicates that the "structural grammar" of the proteome is a fundamental component of biological identity. For instance, the Random Forest model achieved an AUROC of 0.949 for predicting normal breast tissue using only structural features, nearly matching the performance of models using gene names (0.951).
The Role of Intrinsically Disordered Regions (IDRs) in Cancer
Breast cancer signatures were notably characterized by an overexpression of Intrinsically Disordered Regions (IDRs). Unlike rigid, folded proteins, IDRs are flexible segments that often serve as signaling hubs. In both the COSMIC and patient datasets, features like "Disordered Binding" and "Disorder" showed significant overexpression (log-ratios > 2), suggesting that cancer cells favor high-plasticity proteins to rewire cellular pathways.
Transcriptome-Proteome Discrepancies
Interestingly, the structural signature of breast cancer differed between RNA and protein data. While RNA data highlighted IDRs, proteomic data showed an enrichment of transmembrane helices and immunoglobulin domains. This discrepancy may be due to the transient nature and short half-lives of IDR-heavy signaling proteins, which makes them more difficult to detect in steady-state proteomics compared to transcriptomics.
Structure-Based Drug Reversion
SAGES demonstrated that breast cancer drugs induce structural perturbations in cell lines that closely align with the disease's own structural signature. Breast cancer drugs showed fewer structural differences (314 features) from the disease background compared to non-targeted drugs (321 features). This suggests that SAGES can be used for structure-based drug repurposing by identifying molecules that target the specific structural neighborhood of a disease.
Conclusion
SAGES provides a novel lens for analyzing biological data by focusing on the physical properties of proteins. Our findings highlight the importance of protein flexibility and IDRs in breast cancer and demonstrate the predictive power of structural signatures. Future work will focus on applying SAGES to other cancer types and utilizing these signatures to screen for new therapeutic agents that can reverse disease-associated structural states.
Summary updated