Deep learning-based joint sequence–structure de novo membrane protein design

Lucas S. P. Rudden, Remo Bättig, Virginia Andrews, Julie V. Nguyen, Martin Stoll, Lorenzo Scutteri, M. Winnicki, Melissa Call, Matthew E. Call, Damien Thévenin, Patrick BarthbioRxiv 2025 · DOI

TL;DR

This paper introduces TMDiffusion (TMDF), a novel deep learning model that designs transmembrane (TM) proteins from scratch by jointly considering both their amino acid sequence and 3D structure. TMDF achieves high experimental success rates in designing diverse TM protein architectures, including inhibitors and signaling receptors, paving the way for new therapeutics and a deeper understanding of membrane protein function.

What Problem Does This Paper Address?

Designing transmembrane (TM) proteins de novo (from scratch) is a significant challenge in protein engineering. Unlike soluble proteins, TM proteins are embedded in cell membranes, making their structural characterization difficult due to scarce experimental data. Their function is dictated by complex interactions within the hydrophobic membrane environment and intricate conformational dynamics. Existing design methods often rely on approximations or struggle to capture the full complexity of these interactions. This limits our ability to engineer novel TM proteins with specific functions, which are crucial for many biological processes and therapeutic applications.

Methods

The core innovation is TMDiffusion (TMDF), a joint all-heavy-atom sequence-structure diffusion model. Unlike previous methods that often treat sequence and structure design separately or sequentially, TMDF models them simultaneously. It's trained on a large dataset of natural TM protein structures to learn the intricate relationships between sequence, structure, and biophysical interactions within the membrane.

Key aspects of the method include:

  • Joint Sequence-Structure Diffusion: The model learns to generate both the amino acid sequence and the 3D atomic coordinates of TM proteins in a single step.
  • All-Atom Representation: TMDF operates on a full all-atom description, capturing fine-grained details of protein structure and interactions.
  • Membrane-Specific Interactions: The model is specifically trained to understand and recapitulate TM-specific interactions like hydrogen bonds, polar networks, and hydrophobic packing.
  • Diverse Architectures: TMDF can design various TM protein architectures, including inhibitors, associating domains, and conformational switches.
  • Experimental Validation: The designed proteins are experimentally validated, demonstrating their functionality and accuracy.

Datasets

The training data for TMDF was curated from a large reservoir of 5379 multipass membrane proteins, in addition to existing homo- and heterodimeric datasets. Dimeric binding motifs were extracted from these proteins, adhering to strict criteria such as resolution (< 3 Å), membrane burial, and helical structure. This extensive dataset allows TMDF to learn a comprehensive representation of TM protein interactions.

Key Insights and Ideas

  • Simultaneous Sequence-Structure Design is Crucial for TM Proteins: The paper strongly suggests that treating sequence and structure design jointly is essential for accurately modeling the complex biophysical interactions that govern TM protein function.
  • Deep Learning Can Capture Subtle Membrane-Specific Interactions: TMDF demonstrates that deep learning models can learn and leverage the nuanced, often weak, interactions (like CA-H…O hydrogen bonds) that are critical for TM protein assembly and function.
  • Designing Conformational Switches is Feasible: The ability to design TM proteins that can toggle between different conformations opens doors for creating sophisticated biosensors and regulators.
  • Atomic Accuracy in Design is Achievable: The experimental validation, including a crystal structure matching predictions with atomic accuracy, highlights the precision of TMDF.

Key Results

  • High Experimental Success Rate: TMDF achieved a success rate of >70% in experimental validation for designed TM architectures.
  • Accurate Structure Prediction: A crystal structure of a designed protein matched the TMDF prediction with sub-angstrom accuracy.
  • Diverse Design Capabilities: TMDF successfully designed TM inhibitors, associating domains, and conformational switches.
  • Synthetic Receptors with Programmable Function: The model enabled the creation of synthetic single-pass receptors whose TM domains could be precisely controlled to toggle between conformations, influencing signaling outputs.
  • Improved Binding Affinity: In some cases, designed proteins exhibited higher affinity oligomerization than wild-type counterparts (e.g., GpA12 exceeding WT by more than twofold).

Limitations and Future Work

  • Stereochemical Accuracy and Protein Length Control: The authors note that the approach could be improved by enhancing stereochemical accuracy and better controlling protein length. Steric clashes in generated structures can arise from modeling contacts via residue centers of mass.
  • Complex Multi-pass Membrane Protein Design: While TMDF excels at designing dimers, extending its capabilities to assemble multiple, partially overlapping dimer interfaces for complex multi-pass membrane proteins is a promising future direction.
  • Exploration of Diverse TM Architectures: The current work focused on TM α-helical dimers. Future work could explore other TM protein folds and architectures.
  • Broader Application of DL for Membrane Proteins: TMDF establishes a framework, but further development is needed to fully explore the vast and largely uncharted territory of membrane protein design using deep learning.

Summary updated

Abstract

Abstract Deep learning has revolutionized soluble protein design, yet de novo transmembrane (TM) protein engineering remains hindered by scarce structural data, complex membrane-specific interactions and conformational dynamics. We developed TMDiffusion (TMDF), a joint all-heavy-atom sequence–structure diffusion model trained to capture the full interaction diversity of natural TM proteins, including weak and polar contact networks. TMDF designs diverse TM architectures—associating domains, inhibitors, and conformational switches—in a single step, achieving &gt;70% experimental success. A crystal structure of designed proteins matches predictions with atomic accuracy. Leveraging TMDF, we built synthetic single-pass receptors whose de novo TM domains toggle between conformations, enabling precise control of signalling outputs consistent with predicted equilibria. These results show that membrane-adapted DL models can accurately encode and program TM association energetics and conformations. TMDF establishes a general framework for bottom-up design of TM proteins with programmable functions, advancing both mechanistic studies of membrane proteins and development of next-generation therapeutics.
BibTeX
@misc{rudden2025deep,
  author = {Lucas S. P. Rudden and Remo Bättig and Virginia Andrews and Julie V. Nguyen and Martin Stoll and Lorenzo Scutteri and M. Winnicki and Melissa Call and Matthew E. Call and Damien Thévenin and Patrick Barth},
  title = {{Deep learning-based joint sequence–structure <i>de novo</i> membrane protein design}},
  year = {2025},
  doi = {10.1101/2025.08.15.670493},
  howpublished = {bioRxiv},
}
Figure