Synergizing sequence and structure representations to predict protein variants

Tong Chen, Pranam ChatterjeeCell 2024 · DOI · OpenAlex

Introduction

Predicting the functional consequences of single amino acid substitutions is a cornerstone of modern bioengineering and precision medicine. While traditional methods have relied heavily on evolutionary history, a new research highlight in Cell Research introduces ProMEP (Protein Multimodal Ensemble Protocol). ProMEP is a multimodal representation model that synergizes protein sequence and structural information to achieve state-of-the-art accuracy in predicting mutational effects. By operating "zero-shot"—without the need for Multiple Sequence Alignments (MSAs)—ProMEP offers a transformative approach for engineering proteins, particularly those designed de novo or those lacking close evolutionary relatives.

The Challenge of Mutation Prediction

Historically, two paradigms have dominated the field of protein variant prediction:

  1. Alignment-based methods: Tools like AlphaMissense rely on evolutionary conservation. If a specific amino acid has remained unchanged across millions of years, a mutation at that site is typically predicted to be deleterious. However, these methods require MSAs, which are computationally intensive and inapplicable to synthetic or de novo proteins that have no natural evolutionary history.
  2. Protein Language Models (pLMs): Models such as ESM-2 are trained on vast sequence databases to learn the "grammar" of proteins. While powerful, they often lack an explicit understanding of the 3D spatial arrangement of atoms, which is the physical basis of protein function.

The goal of ProMEP is to bridge this gap by creating a fast, MSA-independent model that understands both the linguistic patterns of sequences and the geometric constraints of 3D structures.

The ProMEP Architecture

ProMEP introduces several technical innovations to handle the complexity of protein data. Unlike models that represent structures as rigid graphs or simplified 2D contact maps, ProMEP utilizes a point cloud representation. This allows the model to capture atomic-resolution details of the protein's 3D structure.

The ProMEP workflow integrates sequence and structural data to predict the functional impact of mutations.

To process this 3D data effectively, the model employs an SE(3)-Transformer. This architecture ensures "3D equivariance," meaning the model's internal representation of a protein remains consistent regardless of its orientation or position in space. The model was pre-trained on a massive dataset of approximately 160 million proteins using a masked training paradigm, learning to reconstruct missing sequence and structural components based on their context.

Results and Real-World Applications

The effectiveness of ProMEP was validated across 15 different benchmark datasets, where it consistently achieved top-tier performance in zero-shot mutation effect prediction.

Bypassing Evolutionary Constraints

A key advantage of ProMEP is its ability to generalize to proteins with low homology. Because it learns the fundamental physical relationship between structure and function, it does not require "relative" sequences to make accurate predictions. This makes it uniquely suited for de novo protein design, where traditional alignment-based tools often fail.

Engineering Better Biotools

The researchers demonstrated ProMEP's utility in practical engineering tasks, specifically targeting the RNA-guided nuclease TnpB and the CRISPR adenine base editor ABE8e. For ABE8e, ProMEP successfully prioritized mutations that not only maintained high editing efficiency but also reduced "off-target" effects in mammalian cells. This highlights the model's potential to accelerate the development of safer and more effective gene-editing therapeutics.

Conclusion and Future Directions

ProMEP represents a significant step forward in protein engineering, offering a faster and more versatile alternative to MSA-dependent models. However, some challenges remain. The model currently relies on static structural snapshots, which may limit its performance on intrinsically disordered proteins or those that undergo significant conformational changes during function. Future work may involve integrating dynamic structural data or leveraging generative models to better capture the flexibility inherent in biological systems. Despite these limitations, ProMEP's ability to integrate multimodal data sets a new standard for the computational design of functional proteins.

Summary updated

BibTeX
@inproceedings{chen2024synergizing,
  author = {Tong Chen and Pranam Chatterjee},
  title = {{Synergizing sequence and structure representations to predict protein variants}},
  booktitle = {Cell},
  year = {2024},
  doi = {10.1038/s41422-024-01010-6},
}
Figure