De novo design of site-specific protein interactions with learned surface fingerprints

Pablo Gaínza, Sarah Wehrle, Alexandra Van Hall‐Beauvais, Anthony Marchand, Andreas Scheck, Zander Harteveld, Stephen Buckley, Dongchun Ni, Shuguang Tan, Freyr Sverrisson, Casper A. Goverde, Priscilla Turelli, Charlène Raclot, Alexandra Teslenko, Martin Pačesa, Stéphane Rosset, Sandrine Georgeon, Jane Marsden, Aaron S. Petruzzella, Kefang Liu, Zepeng Xu, Yan Chai, Pu Han, George F. Gao, Elisa Oricchio, Beat Fierz, Didier Trono, Henning Stahlberg, Michael M. Bronstein, Bruno E. CorreiabioRxiv 2022 · DOI

TL;DR

This paper introduces MaSIF-seed, a novel geometric deep learning framework for de novo protein-protein interaction (PPI) design. By learning "fingerprints" of protein surfaces, MaSIF-seed can efficiently identify complementary binding sites and design novel protein binders with high affinity, demonstrating success against challenging targets like SARS-CoV-2 spike, PD-1, PD-L1, and CTLA-4.

What Problem Does This Paper Address?

Designing new protein-protein interactions (PPIs) from scratch (de novo) is a fundamental challenge in biology and a critical goal for applications like developing therapeutics. While we have vast amounts of genomic and proteomic data, understanding the precise molecular features that drive these interactions remains difficult. This knowledge gap hinders our ability to fully comprehend cellular PPI networks and to engineer custom protein binders. Existing computational methods often struggle with flat interfaces or rely on identifying existing interaction motifs, limiting their ability to design novel interactions for "undruggable" targets. This paper aims to bridge this gap by developing a computational approach that can predict and design novel, specific protein interactions based on learned surface features.

Methods

The core of this work is the MaSIF-seed framework, built upon a geometric deep learning approach. Here's a breakdown of the key components:

  • MaSIF (Molecular Surface Interaction Fingerprinting): This is the foundational deep learning framework that represents protein surfaces as geometric and chemical "fingerprints." It learns patterns that predict the propensity of two molecules to interact.
  • MaSIF-site: A tool derived from MaSIF that takes a protein as input and outputs a score for each surface point, indicating its propensity to become a buried site in a PPI. This helps identify potential interaction regions.
  • MaSIF-search: Another MaSIF-derived tool designed as a Siamese neural network. It learns to produce similar fingerprints for complementary binding partners and dissimilar fingerprints for non-interacting pairs, enabling the search for suitable binding "seeds."
  • MaSIF-seed Pipeline: This is the novel three-stage approach for de novo PPI design:
  • Prediction of target buried interface sites: MaSIF-site is used to identify regions on the target protein with a high propensity for interaction.
  • Surface fingerprint-based search for complementary motifs (binding seeds): MaSIF-seed searches a large database of pre-computed surface fingerprints to find "seeds" that are complementary to the identified target site.
  • Binding seed transplantation to protein scaffolds: The identified binding seeds are grafted onto stable protein scaffolds, and the interface is further optimized using computational tools like Rosetta to confer stability and additional contacts.

This approach is novel in its reliance on learned surface fingerprints and its ability to search for complementary motifs without being limited to known interaction types.

Datasets

  • Training Data:
    • PRISM database
    • PDBBind database
    • ZDOCK benchmark
    • SabDab database
    • A dataset of 6001 protein-protein interactions for MaSIF-search training.
  • A dataset of 12002 subunits for MaSIF-site training.
  • A database of ~250K helical motifs and ~390K beta-strand motifs for binding seed generation.
  • Benchmark Datasets:
    • 114 dimeric complexes for benchmarking MaSIF-seed's ability to identify true binders from decoys.
  • 31 helix:receptor complexes and 83 non-helix:receptor complexes for computational benchmark of binding motif recovery.

The paper does not explicitly mention the creation of a new benchmark dataset, but rather uses existing datasets for training and evaluating their method against established tools.

Key Insights and Ideas

  • Surface Fingerprints as a Universal Language: The core idea is that geometric and chemical features of protein surfaces, when represented as learned "fingerprints," can capture the essence of molecular recognition. This allows for a more general approach to PPI design, moving beyond specific interaction motifs.
  • Efficiently Navigating Vast Search Spaces: The sheer number of possible protein surface configurations is immense. MaSIF-seed's fingerprint-based search dramatically prunes this space, making de novo design computationally feasible.
  • Targeting "Undruggable" Flat Interfaces: The method demonstrates success in designing binders for targets with flat interfaces (like PD-1/PD-L1), which are notoriously difficult to target with small molecules. This opens new avenues for therapeutic development.
  • Synergy of Deep Learning and Traditional Design: The approach effectively combines deep learning for initial site identification and seed discovery with established computational design tools (like Rosetta) for refinement and optimization.
  • De Novo Design Beyond Native Motifs: The success in designing binders with alpha-helical motifs for targets that natively use beta-hairpins (e.g., PD-1/PD-L1) highlights the method's ability to explore novel structural solutions.

Key Results

  • Successful Design Against Multiple Targets: The method successfully designed binders for SARS-CoV-2 spike, PD-1, PD-L1, and CTLA-4.
  • Nanomolar Affinities Achieved: Several designed binders reached nanomolar affinities, with some requiring experimental optimization. For example, DBR3_03 achieved a KD of 80 nM against SARS-CoV-2 RBD, and DBL1_04 achieved a KD of 256 nM against PD-L1.
  • High Accuracy in Predictions: Structural characterization (cryo-EM, crystallography) confirmed that the designed binders engaged the target sites as predicted, with high accuracy in interface predictions.
  • Speed and Efficiency: MaSIF-seed demonstrated significant speedups (20-200 fold) compared to other docking methods in benchmark tests, while maintaining or exceeding their performance in identifying correct binders.
  • Specificity Demonstrated: Control experiments and competition assays confirmed the specificity of the designed interactions.
  • "In Silico" Designs with Native-like Affinities: The study achieved binders with native-like affinities using purely computational design, a significant step forward.

Limitations and Future Work

  • Absence of Conformational Flexibility: The current approach does not fully account for protein backbone flexibility and adaptation to mutations, which can influence binding.
  • Difficulty in Designing Polar Interactions: Balancing hydrophobic and polar interactions for optimal affinity and specificity remains a challenge, as observed in the PD-L1 binder designs where polar interactions were initially suboptimal.
  • Future Methodological Developments: The authors suggest optimizing neural network architectures to better capture conformational flexibility. They also highlight the potential of generative algorithms for backbone construction, which could further enhance the conceptual framework based on surfaces.
  • Experimental Validation is Still Crucial: While computational design has advanced significantly, experimental optimization and validation remain essential steps to achieve high-affinity binders.

Summary updated

Abstract

Abstract Physical interactions between proteins are essential for most biological processes governing life. However, the molecular determinants of such interactions have been challenging to understand, even as genomic, proteomic, and structural data grows. This knowledge gap has been a major obstacle for the comprehensive understanding of cellular protein-protein interaction (PPI) networks and for the de novo design of protein binders that are crucial for synthetic biology and translational applications. We exploit a geometric deep learning framework operating on protein surfaces that generates fingerprints to describe geometric and chemical features critical to drive PPIs. We hypothesized these fingerprints capture the key aspects of molecular recognition that represent a new paradigm in the computational design of novel protein interactions. As a proof-of-principle, we computationally designed several de novo protein binders to engage four protein targets: SARS-CoV-2 spike, PD-1, PD-L1, and CTLA-4. Several designs were experimentally optimized while others were purely generated in silico , reaching nanomolar affinity with structural and mutational characterization showing highly accurate predictions. Overall, our surface-centric approach captures the physical and chemical determinants of molecular recognition, enabling a novel approach for the de novo design of protein interactions and, more broadly, of artificial proteins with function.
BibTeX
@misc{gainza2022i,
  author = {Pablo Gaínza and Sarah Wehrle and Alexandra Van Hall‐Beauvais and Anthony Marchand and Andreas Scheck and Zander Harteveld and Stephen Buckley and Dongchun Ni and Shuguang Tan and Freyr Sverrisson and Casper A. Goverde and Priscilla Turelli and Charlène Raclot and Alexandra Teslenko and Martin Pačesa and Stéphane Rosset and Sandrine Georgeon and Jane Marsden and Aaron S. Petruzzella and Kefang Liu and Zepeng Xu and Yan Chai and Pu Han and George F. Gao and Elisa Oricchio and Beat Fierz and Didier Trono and Henning Stahlberg and Michael M. Bronstein and Bruno E. Correia},
  title = {{<i>De novo</i> design of site-specific protein interactions with learned surface fingerprints}},
  year = {2022},
  doi = {10.1101/2022.06.16.496402},
  howpublished = {bioRxiv},
}
Figure