TOPO: Improving remote homologue recognition via identifying common protein structure framework

TOPO: Improving remote homologue recognition via identifying common protein structure framework


(Extended Abstract)

1 Motivation

Protein structure prediction plays an important role in the fields of bioinformatics and biology. Traditional protein structure prediction approaches include template-based modeling (TBM, including homology modeling, and threading), and free modeling (FM). In particular, a threading algorithm takes a query protein sequence as input, recognizes the most likely fold, and finally reports the alignments of the query sequence to structure-known templates as output. The existing threading approaches mainly utilizes the information of protein sequence profile, solvent accessibility, contact probability, etc. The threading strategy has been shown to be successful in structure prediction of a great amount of proteins; however, the existing threading approaches show poorly performance for remote homology proteins. How to improve the fold recognition for remote homology proteins remains a challenge to protein structure prediction.

The sequences of proteins in remote homology generally show relatively weak signal of structure. However, this does not mean that there is no sequence conservation hints for structure. The success of multiple-templates strategy implies the existence of common frameworks, i.e. some regions of proteins are conservative both in the structure and sequence. Such common frameworks should be responsible to the structural stability and then conservative in the evolution.

Based on this we proposed a novel threading approach in three steps. First, for each template, the common structural frameworks shared by its homologous proteins were calculated. Second, unlike in traditional threading methods where the alignment is made against the whole template, we aligned the query protein sequence against a common framework first. This strategy avoids the drawback of the traditional threading approach, i.e. the alignment of variable regions beyond conserved motifs is prone to bringing in error. Third, the final alignments were generated via aligning query sequence against candidate full-length templates in the family. Briefly speaking, we run TreeThreader[2] to build alignments of query against the new template database, and ranked alignments by E-value for model generation. Finally, we generated models by MODELLER based on candidate alignments. The generated models are ranked according to dDFIRE[3] energy function.

2 Methods

For each template with known structure, all of its remote homology proteins are first identified based on structure alignment. Then, a linear programming was designed to identify the common framework shared by these remote homology proteins.

The common framework identification problem can be described as: given a collection of homologous proteins with length , the objective is to find segments with length with high sequence conservation and structural similarity. As an example, Fig. 1 shows the common frameworks shared by protein 3gxr_A and its homologous proteins.

Figure 1: Common frameworks shared by protein 3gxr_A and its homologous proteins. The common framework consists of three dispersed segments (in yellow, cyan, and green). At the conserved segments, the homologous proteins display significant sequence conservation and structural similarity.

2.1 Basic idea of the linear program

The common framework poses double-fold requirements, i.e., significantly high sequence conservation and structural similarity. In the linear program, the objective function was designed to describe structural similarity, and the constraints were designed to describe sequence similarities.

Specifically, the linear program utilizes a set of boolean variables to represent the location of conserved segments, i.e., denotes that in the th protein, the th segment is located at the -th residue. Then, the structural similarity objective and sequence similarity constraints can be described using .

The constraints were designed to represent the following requirements.

  • For any sequence, the th segment in common framework is unique;

  • No segment in a common framework overlaps nor crosses.

  • The segments should have significantly high sequence similarity.

The integer linear programming model can be described as:


where denotes the length of the th protein, and denotes the pre-calculated sequence similarity matrix. In particular, the cell denotes the sequence similarity between of the segment starting from in the th protein and the segment starting from in the th protein.

2.2 Refining the ILP model

In our model, the structure similarity is described using Dscore [1].

where and denote the C distance of residue and residue in protein and respectively.

The final integer linear programming can be formulated as:

where the indicator equals to 1 iff all the four item , , , equal to 1, and 0 otherwise. The indicator equals to 1 iff both and equal to 1. The and matrix are calculated in advance. The cell denotes the approximated Dscore of segments start from and in the th protein and segments start from and in the th protein.

2.3 An example

Fig. 2 shows the common frameworks shared by two domains in SCOP family c.37.1.11. The common frameworks has a Dscore of 8.35 and an RMSD of 1.9Å, implying a significantly high structure conservation.

Figure 2: Common frameworks shared by two domains in SCOP family c.37.1.11.

3 Experiments before CASP11

For a total of over 27,000 proteins in PDB70, updated at Apr. 19, 2014, the common frameworks were identified to yield a database called TOPO. The test set consists of 142 pairs of protein structures similar in structure but with low sequence identity. Traditional threading approaches, say HHpred, fail to build an accurate alignment between such protein pairs. In contrast, our alignment method successfully build accruate alignment (TMscore) for seven protein pairs, and generate accurate contact information for 45 protein pairs. Take a pair of protein 3dz1_A vs. 1twd_A as an example. The two proteins share similar protein structure (TMscore=0.56); however, the alignment generated by HHpred has a TMscore of only 0.22. In contrast, our alignment method generates an alignment with TMscore=0.43.

4 Conclusions

Unlike close homology proteins, remote homology proteins show weakly overall sequence signals of structure similarity. However, they still share common frameworks which carry strong sequence signals of structure similarity. Aligning against the common frameworks instead of whole protein sequences improves the fold recognition.


The study was funded by the National Basic Research Program of China (973 Program) under Grant 2012CB316502, the National Nature Science Foundation of China under Grants 11175224 and 11121403, 31270834, 61272318, 30870572, and 61303161 and the Open Project Program of State Key Laboratory of Theoretical Physics (No.Y4KF171CJ1). This work made use of the eInfrastructure provided by the European Commission co-funded project CHAIN-REDS (GA no 306819).


  1. Correspondence should be addressed to Wei-Mou Zheng ( and Dongbo Bu (


  1. Zhang J, Xu D. Fast algorithm for population‐based protein structural model analysis[J]. Proteomics, 2013, 13(2): 221-229.
  2. Wu W, Chen G, Kan W, et al. Harness public computing resources for protein structure prediction computing[C]. The International Symposium on Grids and Clouds (ISGC). 2013, 2013.
  3. Yang Y, Zhou Y. Specific interactions for ab initio folding of protein terminal regions with secondary structures[J]. Proteins: Structure, Function, and Bioinformatics, 2008, 72(2): 793-803.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description