Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2022 Jul 13.
Published in final edited form as: ACM BCB. 2020 Sep;2020:110. doi: 10.1145/3388440.3414708

Combine Cryo-EM Density Map and Residue Contact for Protein Structure Prediction – A Case Study

Maytha Alshammari 1, Jing He 1,†
PMCID: PMC9279007  NIHMSID: NIHMS1724531  PMID: 35838376

Abstract

Cryo-electron microscopy is a major structure determination technique for large molecular machines and membrane-associated complexes. Although atomic structures have been determined directly from cryo-EM density maps with high resolutions, current structure determination methods for medium resolution (5 to 10 Å) cryo-EM maps are limited by the availability of structure templates. Secondary structure traces are lines detected from a cryo-EM density map for α-helices and β-strands of a protein. When combined with secondary structure sequence segments predicted from a protein sequence, it is possible to generate a set of likely topologies of α-traces and β-sheet traces. A topology describes the overall folding relationship among secondary structures; it is a critical piece of information for deriving the corresponding atomic structure. We propose a method for protein structure prediction that combines three sources of information: the secondary structure traces detected from the cryo-EM density map, predicted secondary structure sequence segments, and amino acid contact pairs predicted using MULTICOM. A case study shows that using amino acid contact prediction from MULTICOM improves the ranking of the true topology. Our observations convey that using a small set of highly voted secondary structure contact pairs enhances the ranking in all experiments conducted for this case.

Keywords: protein structure, cryo-electron microscopy, secondary structure, contact, amino acid, topology, image, constraints

1. INTRODUCTION

Cryo-electron microscopy (cryo-EM) is a central technique for molecular structure determination, particularly for large molecular or membrane-bond complexes. Although a growing number of large molecular complexes have been resolved to atomic resolutions using this technique [1–3], it is often not possible to derive atomic structures from medium resolution density maps (5–10 Å) without knowledge of known atomic structures to use as templates. When a template structure is available, fitting is used to derive atomic structure [4–6]. When no suitable template structures are available, matching secondary structures that are detected from the density map with those predicted from the sequence of the protein may suggest possible topologies of the secondary structures [7–12].

Protein secondary structures, such as α-helices and β-sheets are the most distinguishable characteristics in a medium resolution cryo-EM density map, but characteristic details of amino acids are not discernible at such a resolution. Multiple methods have been developed to detect secondary structure elements such as α-helices and β-sheets from cryo-EM density maps [13–20]. An α-helix resembles a cylinder in the density map, and a β-sheet appears as a thin layer of density. Four helices were identified using DeepSSETracer, a secondary structure detection method that uses convolutional neural networks (CNN) (shown as red rods in Figure 1 (A)). DeepSSETracer also detected a β-sheet region from the density map.

Figure 1. Secondary structures, topology, and contact.

Figure 1.

Open in a new tab

(A) The cryo-EM density map (gray, EMDB ID 6810) that correspond to chain H of atomic structure 5y5x (PDB ID) was superimposed with secondary structure traces of helices (shown in red) and β-strands (shown in blue). Black arrows indicate order of the true topology. Green arrows indicate correctly predicted secondary structure contact pairs. Orange arrows indicate wrongly predicted secondary structure contact. (B) An example of a wrong topology superimposed with the atomic structure (yellow ribbon) of 5y5x chain H. (C) An illustration of the amino acid sequence of protein 5y5x chain H annotated with the location of helices (red rectangles) and β-strands (blue rectangles) predicted using JPred [23].

We have previously shown that it is possible to predict the orientation of β-strands through the analysis of twist [21]. Given a β-sheet density region segmented from a cryo-EM map, StrandTwister produces a small set of β-traces, each representing the central axis of a β-strand. Using DeepSSETracer and StrandTwister, each detected secondary structure trace is represented by a line. As an example, seven secondary structure traces were detected and labeled from L0 to L6 (Figure 1 (A)). Four traces are helices (shown in red), and three are β-strands (shown in blue).

Most large secondary structures can be detected from a cryo-EM density map of medium resolution and are represented as traces in 3-dimensional space. However, it is not known which trace is close to the N-terminal of the protein sequence and the order of the traces along the protein sequence. Existing work has shown that it is possible to derive a small set of possible secondary structure topologies when the detected traces are combined with the predicted sequence segments [12, 22]. Given N secondary structure traces detected from a cryo-EM density map, a topology describes the order of the N traces and the direction of each trace with respect to the protein sequence. In total, there are N!2N topologies, and only one of them is the correct one. The correct topology is indicated in Figure 1 (A), in which the order of the seven detected secondary structure traces is (L4, L0, L1, L5, L2, L6, L3) starting from N-terminal to C-terminal of the sequence. Each trace has two ends, and one of the two directions of the trace aligns with the direction of the protein sequence. An example of an incorrect topology is shown in Figure 1 (B), with an incorrect order of the seven detected secondary structure traces (L6, L1, L0, L5, L2, L4, L3).

The topology of secondary structure traces defines the overall folding framework of the protein sequence. It is a critical piece of information for protein structure prediction. Typical protein structure prediction is based entirely on the protein sequence. The topology information utilizes the cryo-EM density map that provides information complementary to the protein sequence. Existing methods for deriving topologies are based on matching the geometrical information of the secondary structure traces with the predicted sequence segments [7, 8, 10–12]. The accuracy for deriving topology is still yet to be improved. We propose a method, in this paper, to incorporate the amino acid contact information in the topology prediction. We performed a case study and observed improvement in topology ranking even when only a small number of secondary structure contact pairs is utilized.

2. METHOD

Given the amino acid sequence of a protein, existing methods are available to predict amino acid pairs that are in contact [24–27]. Such information is critical for protein structure prediction.

2.1. Protein secondary structure contact

Amino acid contact prediction was performed using DNCON2, which is a tool of MULTICOM software [28, 29]. Given an amino acid sequence, DNCON2 produces a list of all amino acid contacts, each with a p-value (between 0 and 1) for the probability for the two residues being in contact. In order to extract significant long-range contacts, screening was conducted to 1) remove all pairs with near zero p-values; 2) remove short-range pairs with less than or equal to 6 amino acids separating them; 3) extract those pairs that have p-values larger than three standard deviation of the p-values of the protein. As a result of the screening, 58 pairs of long-range significant residue contacts were extracted, and 46 pairs involve two secondary structures.

2.2. Secondary structure traces from Cryo-EM density maps

The cryo-EM density map (EMD ID 6810) was downloaded from the Electron Microscopy Data Bank (EMDB). The corresponding atomic structure (PDB ID 5y5x) was used to extract the density region of chain H. The region of α-helices and β-sheets were detected from the density map using DeepSSETracer that uses a convolutional neural network to detect secondary structures. Traces of α-helices were derived using Principle Component Analysis (PCA) of the segmented volume of helices. The traces of β-strands were derived using StrandTwister [21].

2.3. Deriving possible topologies

Secondary structure traces (SSTs), refer to the set of α-traces and β-traces detected from the Cryo-EM density map. The secondary structure sequence segments refer to α-helices or β-strands predicted using existing software such as JPred or SYMPRED [23, 30]. MultiTopoDP is a graph-based dynamic programming method to match between the secondary structure traces with secondary structure sequence segments [12]. In this application, four α-traces and three β-traces (Figure 1 (A)) were used to match with eight predicted segments (Figure 1 (C)) to produce a list of possible topologies. Although MultiTopoDP allows input of secondary structure segments from multiple predictors, only JPred was used in this application. The skeleton of the cryo-EM density map was derived using SkelEM [31]. MultiTopoDP produces a list of all possible topologies and indicates the rank of the true topology.

2.4. Re-rank Topologies using Contact Pairs

The contact information about secondary structures was used to evaluate each possible topology and those topologies that satisfy the contact constraints were ranked higher (Figure 2). Firstly, contacts at the amino acid level were converted to secondary structure level based on predicted sequence segments from JPred. As an example, 11 significant long-range amino acid contact pairs were extracted from the output of DNCON2; they belong to S0 and S4 β-strands (Table 1). Secondly, the secondary structure contact pairs are mapped to their corresponding α-trace or β-trace according to each topology. Thirdly, the shortest distance between the pair of traces supposedly in contact was evaluated for satisfaction of the distance requirement. Two secondary structures that are in contact are expected to be closer than 12Å from central-line to central-line, as measured by the shortest distant between two traces. Finally, the satisfaction percentage ((Number of satisfied pairs/total number of pairs in contact) * 100) was calculated for each possible topology to re-rank the topologies.

Figure 2. Evaluation of possible topologies using amino acid contact pair.

Figure 2.

Open in a new tab

A list of possible topologies was produced using MultiTopoDP.

Table 1.

Secondary structure contact pairs derived from MULTICOM amino acid contact prediction.

Contact Secondary Structure Pairs AA pairs
(S0, S4) (β, β) 11
(S2, S3) (β, α) 2
(S3, S5) (α, α) 8
(S3, S4) (α, β) 2
(S4, S6) (β, β) 20
(S4, S7) (β, α) 2
(S5, S6) (α, β) 1
Open in a new tab

The seven secondary structure contact pairs (labeled in Figure 1C) and the number of significant long-range amino acid pairs extracted are shown in the two columns respectively. The type of each secondary structure is indicated as α-helix or β-strand.

3. RESULTS

Amino acid contact pairs were produced using the DNCON2 tool of MULTICOM [28, 29]. 46 significant long-range contact pairs of amino acids that involve two secondary structures were extracted (Table 1). The 46 pairs were mapped to seven pairs of secondary structures (Table 1). The secondary structure sequence segments were predicted using JPred [23]. Among the seven pairs of secondary structures, a pair of β-strands (S4, S6) has 20 pairs of significant long-range pairs of amino acids in contact. This suggests the existence of contact between secondary structures S4 and S6. This pair has the highest number of predicted amino acid contact pairs among the seven (Table 1). Although a contact pair of amino acids with the highest p-value may not be the closest pair in the atomic structure, we observed that they are often in the proximity of the shortest-distance pair of amino acids. Many amino acid contact pairs predicted for the same secondary structure pair may suggest the existence of contact for the secondary structures. Another pair of β-strands (S0 and S4) has the second largest number of amino acid contact pairs. Among the seven contact pairs of secondary structures, six are correctly predicted after cross-check with the atomic structure. One pair (S4, S7) is not correct, with two significant long-range pair of amino acid predicted. However, the other three correct pairs have one or two significant long-range contact predicted. It appears that it is more difficult to identify all correct pairs, but it is easier to identify a subset of likely secondary structure pairs that are in contact. We noticed that three of the seven pairs have 20, 11, and 8 predicted amino acid contact pairs respectively, many more than the other four pairs have. The analysis of the amino acid contact prediction suggests that those three pairs of secondary structures are most likely to be in contact.

MultiTopoDP method uses two dynamic programming algorithms to rank possible topologies when provided with a set of secondary structure traces and multiple sets of secondary structure sequence segments [12]. We applied MultiTopoDP in four different settings of contact pairs. In the first setting, no secondary structure contact pairs were used, and the ranking of topologies is only based on the detected traces from the cryo-EM density map and the predicted secondary structure sequence segments using JPred. In the 2nd, 3rd, and 4th settings, we evaluated the effect of using six, two, and three pairs of secondary structure contact pairs. In each setting, seven secondary structure traces (four α-traces and three β-traces) and eight sequence segments were used (Figure 1, (A) and (C)). The rank of the true topology on the list of possible topologies was used to evaluate the effectiveness of the method, since ideally the true topology is ranked top 1. When no secondary structure contact pair was incorporated, the true topology was ranked the 5th in the first set of secondary structure traces. We also examined two alternative sets of traces in which the three β-traces are shifted or have about 15° difference in orientation, since the precise location of β-strands is hard to detect. Only the rough orientation and position can be correctly detected. Using three alternative sets of secondary structure traces, the true topology was ranked 5th, 6th, and 16th respectively, when no contact pair information was applied. These ranks serve a baseline to evaluate the effect of using contact pairs.

To evaluate the effectiveness of using contact information, three experiments were conducted using six pairs, two most-voted pairs, and three most-voted pairs of secondary structures. Although eight sequence segments of secondary structure were predicted though JPred, only seven traces were detected from the cryo-EM density map. In particular, a short segment that was predicted as a β-strand (S2) was not detected from the density map. This leaves six out of seven pairs of secondary structure contacts available for distance measurement in the density map (Table 1). The rank of the true topology is improved from the 5th to top 1 when all six measurable and available pairs of secondary structure contacts were used. This is because those topologies that satisfy higher percentages of contact constraints are re-ranked higher. The contact constraints of secondary structures improve the rank of the true topology in all three experiments in which three alternative structure traces were used (Table 2).

Table 2.

Improved ranks of true topology using contact pairs of secondary structures.

Alternative sets of β-tracesa Rank of True Topology
No Contact Pairs Six Paris Two Pairs Three Pairs
1 5 1 4 1
2 6 3 6 2
3 16 6 11 7
Open in a new tab
a

Three alternative sets of traces among which β-traces are slightly different; The rank of the true topology after zero pair, six pairs, two pairs, and three pairs of contacting secondary structures were used (2nd to 5th column) respectively. Six pairs: all pairs in Table 1 except (S2, S3); two pairs: (S0, S4) and (S4, S6); three pairs: (S0, S4), (S4, S6) and (S3, S5).

Since it is challenging to predict all secondary structure contacts correctly, it is important to study the effectiveness of a subset of pairs, particularly those that are highly voted. One of the six pairs of secondary structure contacts (S4, S7) is predicted incorrectly, and three pairs are voted higher than the others. In the true topology, the predicted contact (S4, S7) corresponds to (L5, L3). However, L5 and L3 are not close enough to make contact (orange arrows in Figure 3 (A)), and the atomic structure shows that the helix is far from the β-strand (Figure 3 (B), (C)). A measure of the shortest distance between L5 and L3 is between 14.93Å and 17.42Å for the three sets of alternative traces. The three highly-voted pairs of secondary structure contacts are (S4, S6), (S0, S6), and (S3, S5) with 20, 11, and 8 pairs of significant long-range amino acids in contact respectively (Table 1). We observe that the rank of the true topology is 1, 2, and 7 respectively for three sets of alternative traces when only three highly-voted pairs are used. The experiments suggest that the use of three highly-voted secondary structure contact pairs (Column 5 of Table 2) is as effective as the use of all six pairs (Column 3 and 5 of Table 2).

Figure 3. Detected β-traces and two alternative atomic structures.

Figure 3.

Open in a new tab

(A) Secondary structure traces detected from cryo-EM density map of 6810 (EMD ID) corresponding to chain H of 5y5x (PDB ID) and the true topology. See annotation in Figure 1 (A). (B) The atomic structure (yellow ribbon), chain H of 5y5x. (C) The atomic structure (spring green ribbon) of chain H in 6r0z.

In this case study, we observed two alternative annotations of the atomic structure in chain H of 5y5x and chain H of 6r0z. The two chains have identical sequence, but annotation of the two structures is slightly different, particularly in the β-sheet region. Two loops were annotated in this region in 5y5x (Figure 3 (B)), but two β-strands were annotated at the approximate the same region in 6r0z (Figure 3 (C)). The detection of secondary structure traces was based on the density map EMD-6810 that has 5Å resolution, from which 5y5x atomic structure was derived. DeepSSETracer and StrandTwister detected only three β-strands (L4, L5, L6 in Figure 3A) that agree well with the recent structure 6r0z, but not as well with 5y5x. Two of the three β-traces (L5 and L6) are annotated as loops in 5y5x (Figure 3B), but they are annotated as β-strands in 6r0z (Figure 3 (C)). The helix at the detected trace L1 shows a slight difference in the number of turns and the orientation when 5y5x and 6r0z are compared (Figure 3 (B) and (C)). It is not clear why the two structures are annotated differently, even though the detected secondary structure traces agree more with the structure in 6r0z in the β-sheet region. The predicted sequence segments using JPred also agrees with the structure in 6r0z. The structure in 5y5x appears to be less folded nearby the β-sheet region. There is slight difference in the resolution of the two cryo-EM density maps. 6810 (EMDB ID) from which 5y5x was derived has 5.0 Å resolution, and 4702 (EMDB ID) from which 6r0z was derived has 3.8 Å resolution. It would be interesting to see if the difference in annotation is because of the limitation of the density map resolution or due to biologically relevant factors.

4. CONCLUSION

We propose a method for protein structure prediction that combines three sources of information: the secondary structure traces detected from the cryo-EM density map, predicted secondary structure sequence segments, and amino acid contact pairs that can be predicted using MULTICOM. A case study suggests that amino acid contact prediction from MULTICOM improves the rank of true topology of secondary structure traces. In this case, using a small set of highly-voted secondary structure contact pairs enhances the ranking as well as using all available contact pairs. Our results show the potential of combining the cryo-EM density maps with well analyzed contact information in deriving protein structures for cryo-EM density maps at medium resolution. Our study investigated two alternative atomic structures, Chain H of 5y5x and 6r0z, both have the same sequence. We observed that the secondary structures that were detected from the cryo-EM density map EMD-6810 using DeepSSETracer and StrandTwister agree better with 6r0z, particularly in the region of two β-strands.

CCS CONCEPTS.

  • Applied computing → Life and medical sciences → Computational biology → Molecular structural biology

5. ACKNOWLEDGMENTS

The work in this paper is supported by NIH R01-GM062968 and the PhD scholarship to M.A. provided by the government of Saudi Arabia.

6. REFERENCES

  • [1].Liu Z, Guo F, Wang F, Li T-C and Jiang W 2.9 Å resolution cryo-EM 3D reconstruction of close-packed virus particles. Structure, 24, 2 (2016), 319–328. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [2].Liu H, Jin L, Koh SB, Atanasov I, Schein S, Wu L and Zhou ZH Atomic structure of human adenovirus by cryo-EM reveals interactions among protein networks. Science (New York, N.Y.), 329, 5995 (2010), 1038–1043. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [3].Wlodawer A, Li M and Dauter Z High-Resolution Cryo-EM Maps and Models: A Crystallographer’s Perspective. Structure, 25, 10 (2017), 1589–1597.e1581. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [4].Schröder GF, Brunger AT and Levitt M Combining efficient conformational sampling with a deformable elastic network model facilitates structure refinement at low resolution. Structure, 15, 12 (2007), 1630–1641. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [5].Chan KY, Trabuco LG, Schreiner E and Schulten K Cryo-electron microscopy modeling by the molecular dynamics flexible fitting method. Biopolymers, 97, 9 (2012), 678–686. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [6].Wriggers W and Birmanns S Using situs for flexible and rigid-body fitting of multiresolution single-molecule data. Journal of structural biology, 133, 2–3 (2001), 193–202. [DOI] [PubMed] [Google Scholar]
  • [7].Abeysinghe S, Ju T, Baker ML and Chiu W Shape modeling and matching in identifying 3D protein structures. Computer-Aided Design, 40, 6 (2008), 708–720. [Google Scholar]
  • [8].Al Nasr K, Ranjan D, Zubair M and He J Ranking valid topologies of the secondary structure elements using a constraint graph. Journal of bioinformatics and computational biology, 9, 03 (2011), 415–430. [DOI] [PubMed] [Google Scholar]
  • [9].Lindert S, Alexander N, Wötzel N, Karakaş M, Stewart PL and Meiler J EM-fold: de novo atomic-detail protein structure determination from medium-resolution density maps. Structure, 20, 3 (2012), 464–478. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [10].Al Nasr K, Ranjan D, Zubair M, Chen L and He J Solving the Secondary Structure Matching Problem in Cryo-EM De Novo Modeling Using a Constrained K Shortest Path Graph Algorithm. IEEE/ACM transactions on computational biology and bioinformatics, 11, 2 (2014), 419–430. [DOI] [PubMed] [Google Scholar]
  • [11].Biswas A, Ranjan D, Zubair M and He J A dynamic programming algorithm for finding the optimal placement of a secondary structure topology in Cryo-EM data. Journal of Computational Biology, 22, 9 (2015), 837–843. [DOI] [PubMed] [Google Scholar]
  • [12].Biswas A, Ranjan D, Zubair M, Zeil S, Al Nasr K and He J An effective computational method incorporating multiple secondary structure predictions in topology determination for cryo-EM images. IEEE/ACM transactions on computational biology and bioinformatics, 14, 3 (2016), 578–586. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [13].Kong Y and Ma J A structural-informatics approach for mining beta-sheets: locating sheets in intermediate-resolution density maps. Journal of molecular biology, 332, 2 (2003), 399–413. [DOI] [PubMed] [Google Scholar]
  • [14].Dal Palu A, He J, Pontelli E and Lu Y Identification of Alpha-Helices from Low Resolution Protein Density Maps. Proceeding of Computational Systems Bioinformatics Conference(CSB) (2006), 89–98. [PubMed] [Google Scholar]
  • [15].Baker ML, Ju T and Chiu W Identification of secondary structure elements in intermediate-resolution density maps. Structure, 15, 1 (2007), 7–19. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [16].Rusu M and Wriggers W Evolutionary bidirectional expansion for the tracing of alpha helices in cryo-electron microscopy reconstructions. Journal of structural biology, 177, 2 (2012), 410–419. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [17].Si D, Ji S, Nasr KA and He J A Machine Learning Approach for the Identification of Protein Secondary Structure Elements from Electron Cryo-Microscopy Density Maps. Biopolymers, 97, 9 (2012), 698–708. [DOI] [PubMed] [Google Scholar]
  • [18].Si D and He J Beta-sheet Detection and Representation from Medium Resolution Cryo-EM Density Maps. Proceedings of the International Conference on Bioinformatics, Computational Biology and Biomedical Informatics (2013), 764–770. [Google Scholar]
  • [19].Li R, Si D, Zeng T, Ji S and He J Deep convolutional neural networks for detecting secondary structures in protein density maps from cryo-electron microscopy. 2016 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 15–18 Dec. (2016), 41–46. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [20].Maddhuri SVS, Terashi G and Kihara D Protein secondary structure detection in intermediate-resolution cryo-EM maps using deep learning. Nature methods, 16, 9 (2019), 911–917. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [21].Si D and He J Tracing Beta Strands Using StrandTwister from Cryo-EM Density Maps at Medium Resolutions. Structure, 22, 11 (2014), 1665–1676. [DOI] [PubMed] [Google Scholar]
  • [22].Baker ML, Abeysinghe SS, Schuh S, Coleman RA, Abrams A, Marsh MP, Hryc CF, Ruths T, Chiu W and Ju T Modeling protein structure at near atomic resolutions with Gorgon. Journal of structural biology, 174, 2 (2011), 360–373. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [23].Cuff JA, Clamp ME, Siddiqui AS, Finlay M and Barton GJ JPred: a consensus secondary structure prediction server. Bioinformatics (Oxford, England), 14, 10 (1998), 892–893. [DOI] [PubMed] [Google Scholar]
  • [24].Hanson J, Paliwal K, Litfin T, Yang Y and Zhou Y Accurate prediction of protein contact maps by coupling residual two-dimensional bidirectional long short-term memory with convolutional neural networks. Bioinformatics (Oxford, England), 34, 23 (2018), 4039–4045. [DOI] [PubMed] [Google Scholar]
  • [25].Seemayer S, Gruber M and Söding J CCMpred—fast and precise prediction of protein residue–residue contacts from correlated mutations. Bioinformatics (Oxford, England), 30, 21 (2014), 3128–3130. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [26].Jones DT, Singh T, Kosciolek T and Tetchner S MetaPSICOV: combining coevolution methods for accurate prediction of contacts and long range hydrogen bonding in proteins. Bioinformatics (Oxford, England), 31, 7 (2015), 999–1006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [27].Jones DT, Buchan DW, Cozzetto D and Pontil M PSICOV: precise structural contact prediction using sparse inverse covariance estimation on large multiple sequence alignments. Bioinformatics (Oxford, England), 28, 2 (2012), 184–190. [DOI] [PubMed] [Google Scholar]
  • [28].Adhikari B, Hou J and Cheng J DNCON2: improved protein contact prediction using two-level deep convolutional neural networks. Bioinformatics (Oxford, England), 34, 9 (2018), 1466–1472. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [29].Cheng J, Li J, Wang Z, Eickholt J and Deng X The MULTICOM toolbox for protein structure prediction. BMC bioinformatics, 13, 1 (2012), 1–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • [30].Simossis VA and Heringa J The influence of gapped positions in multiple sequence alignments on secondary structure prediction methods. Computational Biology and Chemistry, 28, 5–6 (2004), 351–366. [DOI] [PubMed] [Google Scholar]
  • [31].Al Nasr K, Liu C, Rwebangira M, Burge L and He J Intensity-based skeletonization of CryoEM gray-scale images using a true segmentation-free algorithm. IEEE/ACM transactions on computational biology and bioinformatics, 10, 5 (2013), 1289–1298. [DOI] [PMC free article] [PubMed] [Google Scholar]

RESOURCES