Predicting membrane protein contacts from non-membrane proteins by deep transfer learning

Zhen Li, Sheng Wang, Yizhou Yu, Jinbo Xu

Abstract Membrane proteins (MPs) are critical drug targets, yet their structural determination remains a significant challenge due to experimental difficulties and a lack of homologous templates. We present a deep transfer learning method that leverages the structural data of thousands of non-membrane proteins to predict the internal contacts of MPs. By treating contact maps as images and employing an ultra-deep residual network architecture, our model captures complex, long-range dependencies. The method significantly outperforms traditional evolutionary coupling analysis and state-of-the-art supervised methods, successfully folding nearly 40% of a non-redundant test set of 510 MPs from sequence alone.

Introduction

Membrane proteins (MPs) serve as essential gatekeepers for cellular processes, acting as receptors and transporters. Despite their medical importance—targeting approximately half of all current drugs—their 3D structures are notoriously difficult to determine experimentally. MPs are hard to crystallize for X-ray studies and often exceed the size limits for NMR. Consequently, high-resolution structures exist for only a small fraction of known MPs.

Computational approaches like homology modeling often fail due to a lack of known templates. Evolutionary coupling analysis (DCA) requires large protein families, which many MPs lack. This research addresses the "data scarcity" problem by investigating whether the fundamental rules of protein folding learned from common globular (non-membrane) proteins can be transferred to predict the structures of membrane-bound ones.

Methodology

The proposed method treats protein contact prediction as a pixel-level image labeling problem. For a protein of length $L$, the contact map is an $L \times L$ grid where each pixel represents whether two residues are physically close ($< 8\text{Å}$).

The core of the approach is an ultra-deep transfer learning architecture utilizing two concatenated Residual Networks (ResNets):

  1. 1D ResNet: Processes sequential features such as secondary structure and solvent accessibility.
  2. Dimension Transformation: Converts 1D sequential features into a 2D matrix via an outer-product-like operation.
  3. 2D ResNet: Merges the transformed sequential features with pairwise features (e.g., co-evolutionary information from CCMpred) to predict the final contact map.
The deep transfer learning architecture for protein contact prediction, featuring 1D and 2D Residual Networks.

Unlike previous methods that predicted contacts independently, this architecture predicts the entire map simultaneously, allowing the network to learn complex, long-range patterns and local contact textures.

Experimental Setup

The model was trained on a dataset of 9,627 non-membrane proteins (non-MPs) from the PDB25 set, ensuring no significant similarity ($< 25\%$ identity) with the test set. A validation set of 600 non-MPs was used for hyperparameter tuning. The final evaluation was performed on 510 non-redundant MPs from the PDBTM database.

Key Insights

A pivotal finding of this study is the universality of protein folding. A model trained exclusively on non-membrane proteins achieved higher accuracy (0.69) on MP contact prediction than a model trained only on membrane proteins (0.63). Analysis of local contact patterns revealed that the most frequent $3 \times 3$ and $4 \times 4$ submatrices are nearly identical between membrane and globular proteins, validating the effectiveness of transfer learning. Furthermore, the deep ResNet implicitly learned environmental constraints, as adding MP-specific features like lipid accessibility yielded negligible improvements.

Results

The "Mixed" model, trained on both non-MPs and MPs, achieved a top L/10 long-range contact accuracy of 0.72, significantly surpassing MetaPSICOV (0.55) and CCMpred (0.47).

The predicted contacts were used as restraints to guide 3D folding via the CNS suite. The method correctly folded 200 out of 510 MPs (defined by a TMscore $> 0.6$), nearly tripling the success rate of MetaPSICOV (77 MPs) and CCMpred (56 MPs).

Comparison of TMscores for the Top1 models generated by the proposed mixed method versus CCMpred.

The method demonstrated a particular advantage for proteins with small sequence families, where traditional evolutionary methods typically underperform.

Conclusion

This deep transfer learning approach represents a major advancement in membrane protein structure prediction. By bridging the gap between data-rich globular proteins and data-poor membrane proteins, the model provides a powerful tool for structural biology. Future work will focus on integrating these contact predictions with more sophisticated 3D modeling protocols to further enhance structural accuracy.

Summary updated

Abstract

Computational prediction of membrane protein (MP) structures is very challenging partially due to lack of sufficient solved structures for homology modeling. Recently direct evolutionary coupling analysis (DCA) sheds some light on protein contact prediction and accordingly, contact-assisted folding, but DCA is effective only on some very large-sized families since it uses information only in a single protein family. This paper presents a deep transfer learning method that can significantly improve MP contact prediction by learning contact patterns and complex sequence-contact relationship from thousands of non-membrane proteins (non-MPs). Tested on 510 non-redundant MPs, our deep model (learned from only non-MPs) has top L/10 long-range contact prediction accuracy 0.69, better than our deep model trained by only MPs (0.63) and much better than a representative DCA method CCMpred (0.47) and the CASP11 winner MetaPSICOV (0.55). The accuracy of our deep model can be further improved to 0.72 when trained by a mix of non-MPs and MPs. When only contacts in transmembrane regions are evaluated, our method has top L/10 long-range accuracy 0.62, 0.57, and 0.53 when trained by a mix of non-MPs and MPs, by non-MPs only, and by MPs only, respectively, still much better than MetaPSICOV (0.45) and CCMpred (0.40). All these results suggest that sequence-structure relationship learned by our deep model from non-MPs generalizes well to MP contact prediction. Improved contact prediction also leads to better contact-assisted folding. Using only top predicted contacts as restraints, our deep learning method can fold 160 and 200 of 510 MPs with TMscore>0.6 when trained by non-MPs only and by a mix of non-MPs and MPs, respectively, while CCMpred and MetaPSICOV can do so for only 56 and 77 MPs, respectively. Our contact-assisted folding also greatly outperforms homology modeling.
BibTeX
@misc{li2017predicting,
  author = {Zhen Li and Sheng Wang and Yizhou Yu and Jinbo Xu},
  title = {{Predicting membrane protein contacts from non-membrane proteins by deep transfer learning}},
  year = {2017},
  eprint = {1704.07207},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1704.07207},
}
Figure