Peptide bonds can have two conformations. The torsion angle ω (Cαi-1-Ci-1-Ni-Cαi) can be around 0°, cis, or around 180°, trans. The peptides in the protein structures in the Protein Data Bank (PDB) almost exclusively contain the trans conformation. However, some of those conformations are incorrect and should actually be cis. Other trans peptide planes should stay trans but need to be rotated ~180° about the Cα-Cα axis. This website details the methodology used to the predict cis↔trans flips and peptide plane flips in the backbone of protein structures, and contains information supplementary to the manuscript 'Detection of transcis flips and peptide plane flips in protein structures' by Wouter G. Touw, Robbie P. Joosten and Gert Vriend. The method predicts ~70K peptide plane flips and ~5K trans → cis flips.

Predict flips

There are several options to predict flips:

The FLPCHK validation routine of WHAT IF's CHECK menu.

Web service
The ShowPepFlips WHAT IF Web Services (WIWS) option allows SOAP and REST requests.

Web server
The flip checks are on the WHAT IF web servers

cis ↔ trans and peptide plane flips

We have adopted a systematic naming system for flips and non-flips. The flip type is indicated by three characters. The first character indicates the starting omega conformation (either t for trans or c for cis). The second character indicates the correct omega conformation (again t or c). The third character indicates whether the carbonyl 'flips' (+) or not (-). For cis-trans flips the third character implies the reverse for the N-H, i.e. tc- involves a NH-flip. Both an NH-flip and CO-flip occur when peptide planes are flipped (tt+). Theoretically there are 6 possible flip types and tt- and cc- designate correct trans and cis peptides:

Flip type Explanation
tt- The peptide conformation is correct and should stay trans
tt+ A peptide plane flip (~180° crankshaft motion about Cα-Cα axis); both CO-flip and NH-flip.
tc- trans to cis with NH-flip
tc+ trans to cis with CO-flip
cc- The peptide conformation is correct and should stay cis
cc+ The entire Cα-C-N-Cα unit theoretically would rotate ~180°
ct- cis to trans with NH-flip
ct+ cis to trans with CO-flip

We found all flip types except cc+ in the PDB. The different flip classes are best illustrated by examples found in structures deposited to the PDB. Click on the pictures to get detailed information ( help ). The residues mentioned are the ones directly after the peptide bond that needs to be flipped.







Training data

Pairs of X-ray structures solved at 3.5 Å or better, containing at least 25 amino acids in at least one chain, for which a DSSP file (Kabsch & Sander, 1981) exists, and that contained at least one trans - cis or peptide plane flip between PDB and PDB_REDO were obtained from the releases of 20-10-2014. From these PDB files stretches of four canonical residues were selected that had all atoms present with non-zero B-factor and full occupancy; no covalently bound atoms were allowed other than the continuation of the chain; all torsion angles and the DSSP secondary structure must be determinable; the four amino acids were neither N- or C-terminal, nor adjacent to a chain break. A training set was obtained by comparing peptide conformations in the pairs of PDB and PDB_REDO structures. The procedure calculates three values 1) ΔC=O, which is the angle between the PDB carbonyl and the PDB_REDO carbonyl after optimal superposition; 2) ΔN-H, which is the angle between the N-H pair; 3) Δω which is the ω torsion angle difference . If Δω is big, a cis-trans NH-flip (tc- or ct-) is assigned when ΔN-H is big and ΔC=O is small, and a CO-flip (tc+ or tc-) is assigned when ΔC=O is big and ΔN-H is small. If Δω is small and both ΔN-H and ΔC=O are big a peptide plane flip is assigned. It was found that the best assignments were obtained when ‘big’ was defined as being greater than 120° and ‘small’ was defined as being less than 60°. Irregular cases were excluded from the training examples. For irregular cases a) ΔN-H or ΔC=O or Δω is big but other criteria are not met; b) either one of the Cα atoms flanking the peptide plane has been superposed with more than 1 Å displacement. Click on the bar to show/hide the pseudo-code for determining the flip types.

# The angle between the carbonyls
oang = calcAngle(CO_PDB, CO_REDO)

# The angle between the amides
hang = calcAngle(NH_PDB, NH_REDO)

# Omega difference
opdb = calcTorsion(CACNCA_PDB)
oredo = calcTorsion(CACNCA_REDO)
odif = abs(opdb - oredo)

# C-alpha displacement
cadif_i-1 = distance(CA_i-1_PDB, CA_i-1_REDO)
cadif_i = distance(CA_i_PDB, CA_i_REDO)
cadif = max(cadif_i-1, cadif_i)

if oang > 120 or hang > 120 or odif > 120:
    if cadif > 1.0:

    if odif > 120:
        # cis-trans flip
        if abs(oredo) > abs(opdb):

        # CO-flip, NH-flip, or hard to determine automatically?
        if oang > 120 and hang < 60:
        else if hang > 120 and oang < 60:
    else if odif < 60:
        # peptide plane flip
        if oang > 120 and hang > 120:

Obviously, the different flip classes and the correct classes (tt- and cc-) are not equally distributed since the overwhelming majority of the PDB peptides have the correct conformation. It is well known that it is notoriously hard for machine learning algorithms to deal with highly skewed data, for example because always predicting the majority class still gives almost perfect prediction accuracy. This is also known as the imbalance problem. Popular strategies include under-sampling of the majority class, over-sampling of the minority class, a combination of both. We found that randomly downsampling the majority class to the size of the minority class to obtain a balanced training set worked well for training Random Forests (see below). When repeated with different random seeds, very similar results were obtained. Training with unbalanced data using estimated class priors did not work well in our hands.

Download training data

Test data

The test cases were manually validated and re-refined when necessary. Note that some of the validated peptides do not conform to all of the training set criteria. For example, some residues in the tetra peptides may be incomplete, are bound to something, etc. When determining the performance in the validation process, these cases have not been included.

Download test data
Feature changes

An incorrectly modeled peptide usually causes problems for the local backbone. The peptide has to be accomodated, also if little space is available, causing a fight between the X-ray data and refinement restraints. This causes strain that shows up in several features describing the local backbone conformation. If the peptide conformation is corrected by a cis↔trans flip or a peptide plane flip and the structure is re-refined, the strain will be relieved and the backbone parameters will adopt their normal values. The figures below show the change in continuous features (described in the Features section) upon flipping and re-refinement by PDB_REDO for x-X-Xnpg-x (X: any residue; Xnpg: any residue except Pro and Gly) tt-/tt+/tc-/tc+, x-X-Pro-x tt-/tc-/tc+, and x-X-Gly-x tt-/tt+ peptides in the test data. The lines are Gaussian kernel density estimates for the number of cases indicated in the legend for each flip class. The feature distributions in PDB structures are indicated with solid lines. The PDB_REDO distributions are indicated with dotted lines. For example, in the ψi (psi) plot for X-Xnpg the PDB_REDO distributions have been shifted for the 56 tt+, 49 tc- and 10 tc+ cases with respect to the PDB ψi distribution. The 'difference' plots show the feature distributions for PDB_REDO - PDB (i.e. 0 means no difference before and after correction). Click on the bar to show/hide the figures.

Test data before and after correction


The features that can capture the difference between an incorrect and correct peptide bond conformation belong to several feature groups and have been calculated for four amino acids surrounding the peptide bond. The feature groups are angles, torsion angles, distances, chiral volumes, B-factors, secondary structure, and a few other groups explained in the table. Rather than B-factors from PDB files, we used B-factors from BDB files. The BDB is a databank that contains PDB files with consistent B-factors (Touw & Vriend, 2014). Most features were calculated by WHAT IF. Click on the bar to show/hide the comprehensive list of all features and their explanation. Note that the definition of ω for residue i in WHAT IF is equal to the standard definition of ω for residue i+1.

Feature explanation

WH indicates whether the feature is part of the Weiss & Hilgenfeld (1999) algorithm. The angle subcript indicates which residue contributes most atoms.

Feature Code Type Explanation WH
∠φi-2 phi_m2 backbone torsion
∠ψi-2 psi_m2 backbone torsion
∠ωi-2 omega_m2 backbone torsion
∠φi-1 phi_m1 backbone torsion y
∠ψi-1 psi_m1 backbone torsion y
∠ωi-1 omega_m1 backbone torsion y
∠φi phi backbone torsion y
∠ψi psi backbone torsion y
∠ωi omega backbone torsion
∠φi+1 phi_p1 backbone torsion
∠ψi+1 psi_p1 backbone torsion
∠ωi+1 omega_p1 backbone torsion
∠N-Cα-Ci-2 ncac_m2 bond angle
∠Cα-C-Ni-2 cacn_m2 bond angle
∠Cα-C-Oi-2 caco_m2 bond angle
∠O-C-Ni-2 ocn_m2 bond angle
∠N-Cα-Cβi-2 ncacb_m2 bond angle a pseudo-Cβ is calculated for Gly
∠C-Cα-Cβi-2 ccacb_m2 bond angle a pseudo-Cβ is calculated for Gly
∠C-N-Cα i-2 cnca_m3 bond angle
∠N-Cα-Ci-1 ncac_m1 bond angle y
∠Cα-C-Ni-1 cacn_m1 bond angle y
∠Cα-C-Oi-1 caco_m1 bond angle y
∠O-C-Ni-1 ocn_m1 bond angle y
∠N-Cα-Cβi-1 ncacb_m1 bond angle a pseudo-Cβ is calculated for Gly
∠C-Cα-Cβi-1 ccacb_m1 bond angle a pseudo-Cβ is calculated for Gly
∠C-N-Cαi-1 cnca_m2 bond angle
∠N-Cα-Ci ncac bond angle y
∠Cα-C-Ni cacn bond angle
∠Cα-C-Oi caco bond angle
∠O-C-Ni ocn bond angle
∠N-Cα-Cβi ncacb bond angle a pseudo-Cβ is calculated for Gly
∠C-Cα-Cβi ccacb bond angle a pseudo-Cβ is calculated for Gly
∠C-N-Cαi cnca_m1 bond angle y
∠N-Cα-Ci+1 ncac_p1 bond angle
∠Cα-C-Ni+1 cacn_p1 bond angle
∠Cα-C-Oi+1 caco_p1 bond angle
∠O-C-Ni+1 ocn_p1 bond angle
∠N-Cα-Cβi+1 ncacb_p1 bond angle a pseudo-Cβ is calculated for Gly
∠C-Cα-Cβi+1 ccacb_p1 bond angle a pseudo-Cβ is calculated for Gly
∠C-N-Cαi+1 cnca bond angle
N-Cαi-1 nca_m1 bond length y
Cα-Ci-1 cac_m1 bond length y
C-Oi-1 co_m1 bond length y
C-Ni cn bond length y
N-Cαi nca bond length y
Cα-Ci cac bond length y
Feature Code Type Explanation WH
i-1-Cαi-1 cam2_cam1 backbone distance
i-1-Cαi cam1_ca backbone distance y
i-Cαi+1 ca_cap1 backbone distance
i-2-Cβi-1 cbm2_cbm1 distance a pseudo-Cβ is calculated for Gly
i-1-Cβi cbm1_cb distance a pseudo-Cβ is calculated for Gly
i-Cβi+1 cb_cbp1 distance a pseudo-Cβ is calculated for Gly
Oi-2-Oi-1 om2_om1 distance
Oi-1-Oi om1_o distance
Oi-Oi+1 o_op1 distance
i-2 cv imp_m2 improper dihedral chiral volume
i-1 cv imp_m1 improper dihedral chiral volume
i cv imp improper dihedral chiral volume
i+1 cv imp_p1 improper dihedral chiral volume
∠COi-2-COi-1 ooang1 angle
∠COi-1-COi ooang2 angle
∠COi-COi+1 ooang3 angle
∠COi-2-COi+1 foang1 angle
B Ni-2 bn_m2 B-factor
B Cαi-2 bca_m2 B-factor
B Ci-2 bc_m2 B-factor
B Oi-2 bo_m2 B-factor
B Ni-1 bn_m1 B-factor
B Cαi-1 bca_m1 B-factor
B Ci-1 bc_m1 B-factor
B Oi-1 bo_m1 B-factor
B Ni bn B-factor
B Cαi bca B-factor
B Ci bc B-factor
B Oi bo B-factor
B Ni+1 bn_p1 B-factor
B Cαi+1 bca_p1 B-factor
B Ci+1 bc_p1 B-factor
B Oi+1 bo_p1 B-factor
δBi-1 bdif1 B-factor gradient backbone to side-chain (0.0 for Gly)
δBi bdif2 B-factor gradient backbone to side-chain (0.0 for Gly)
DSSPi-2 dssp_m2 secondary structure
DSSPi-1 dssp_m1 secondary structure
DSSPi dssp secondary structure
DSSPi+1 dssp_p1 secondary structure
Dφ+ phip WH 0.0 if φi<0° else sin(φi) y
Dtotal dtot WH final WH score y
Oi-1 bump score om1_bump bump WHAT IF bump score
Oi-1 aligned helali other whether the carbonyl of the peptide plane is close to a helix and aligned with the hydrogen bonding helix carbonyls. Why?


4 different training data sets were constructed: X-Xnpg tc-, X-Xnpg tt+, X-Pro tc+ and X-Gly tt+. For all these training sets the negative class is tt-. For all 4 training sets a Random forest (RF; Breiman, 2001) classifier was constructed. The flip type-specific classifiers were later combined into one classifier per residue class. This strategy optimally made use of the available training examples (e.g. many more tt- and tt+ cases could be used for X-Xnpg flips than could have been used with a multi-class training with a balanced data set). A description of RF is available here. In short, an RF is an ensemble of decision trees. Each classification tree is constructed using a random subset of training examples and a random subset of features. The individual trees are so-called weak learners; the correlation between them is relativey small. Every tree predicts the class of a training example. The ensemble individual trees, the RF, classifies new cases by collecting votes from each tree. A threshold determines the fraction of votes that is needed to predict the class. The combination of enough sufficiently uncorrelated weak learners increases the classification strength of the ensemble and usually leads to a robust classifier. The RF models were trained using the randomForest and caret R packages. The RF parameter mtry was tuned using 5-fold cross-validation that was repeated 20 times with different random seeds. The classification performance on the training examples was measured by and/or optimized against the cutoff-independent area under the ROC curve , against the area under the precision-recall curve, and/or against the Matthews correlation coefficient and euclidian distance to perfect specificity and sensitivity at the optimal threshold determined by tuning across 40 different threshold values between 0.5 and 1 (code examples here and here).

The final classifiers were constructed using all training data with parameters from the best-performing cross-validation models. The resampling performance for the final classifiers is shown in the table below (click on the bar to show/hide the table). The AUC under the ROC curve has been calculated by the ROCR package. The rocplus package was used to obtain the AUC under the precision-recall curve.

Training performance
Training performance
minimum 1st quartile median mean 3rd quartile maximum
X-Xnpg tc- 0.9060.9500.9640.9620.9750.997
X-Xnpg tt+ 0.9760.9820.9840.9840.9850.989
X-Pro tc+ 0.8580.9330.9540.9530.9751
X-Gly tt+ 0.9430.9590.9650.9640.9710.983
Precision-recall AUC
X-Xnpg tc- 0.9060.9500.9630.9620.9760.997
X-Xnpg tt+ 0.9760.9820.9840.9840.9850.989
X-Pro tc+ 0.8580.9320.9540.9530.9751
X-Gly tt+ 0.9430.9590.9650.9640.9710.983
Matthews correlation coefficient
X-Xnpg tc- 0.6260.7530.7890.7940.8390.921
X-Xnpg tt+ 0.8430.8690.8780.8770.8860.899
X-Pro tc+ 0.520.8350.8690.8640.8971
X-Gly tt+ 0.740.7880.8110.8100.8310.873
Topleft distance
X-Xnpg tc- 0.06080.1390.1690.1740.2180.390
X-Xnpg tt+ 0.07580.08870.09500.09530.1020.120
X-Pro tc+ 00.07860.1150.1260.1670.404
X-Gly tt+ 0.09810.1270.1420.1410.1550.192
X-Xnpg tc- 0.7920.8750.8900.8930.9170.959
X-Xnpg tt+ 0.9210.9340.9380.9380.9420.949
X-Pro tc+ 0.750.9170.9320.9290.9461
X-Gly tt+ 0.8690.8930.9050.9040.9140.936
False positive rate (fallout)
X-Xnpg tc- 0.00000.11100.1670.1550.1940.389
X-Xnpg tt+ 0.02300.03260.03630.03650.04000.0496
X-Pro tc+ 00.05560.1110.1110.1670.278
X-Gly tt+ 0.03330.05710.06670.06980.08100.129
False discovery rate
X-Xnpg tc- 0.00000.10700.1450.1380.1820.286
X-Xnpg tt+ 0.02450.03440.03820.03850.04230.0518
X-Pro tc+ 00.05560.1000.09960.1430.217
X-Gly tt+ 0.03550.06110.07300.07330.08430.127
True positive rate (sensitivity/recall)
X-Xnpg tc- 0.8060.9170.9460.9420.9721
X-Xnpg tt+ 0.8900.9060.9120.9120.9190.933
X-Pro tc+ 0.6110.9441.0000.96911
X-Gly tt+ 0.8240.8620.8760.8790.8950.933
False negative rate (miss rate)
X-Xnpg tc- 00.02780.05410.05770.08330.194
X-Xnpg tt+ 0.06650.08100.08830.08780.09430.110
X-Pro tc+ 000.00000.03080.05560.389
X-Gly tt+ 0.06670.1050.1240.1210.1380.176
True negative rate (specificity)
X-Xnpg tc- 0.6110.8060.8330.8450.8891.000
X-Xnpg tt+ 0.9500.960.9640.9630.9670.977
X-Pro tc+ 0.7220.8330.8890.8890.9441
X-Gly tt+ 0.8710.9190.9330.9300.9430.967
Positive predictive value (precision)
X-Xnpg tc- 0.7140.8180.8550.8620.8931.000
X-Xnpg tt+ 0.9480.9580.9620.9620.9660.976
X-Pro tc+ 0.7830.8570.9000.9000.9441
X-Gly tt+ 0.8730.9160.9270.9270.9390.964
Negative predictive value
X-Xnpg tc- 0.8110.9140.9420.9390.9681
X-Xnpg tt+ 0.8960.9110.9160.9170.9220.935
X-Pro tc+ 0.6960.9441.0000.97011
X-Gly tt+ 0.8420.8740.8840.8850.8990.932
X-Xnpg tc- 0.8170.8800.8970.8990.9210.961
X-Xnpg tt+ 0.9180.9320.9370.9360.9410.948
X-Pro tc+ 0.710.9190.9340.9320.9481
X-Gly tt+ 0.8640.8910.9020.9020.9120.934

Features important for correct classification of training examples are shown below. These figures show either the overall mean and sd permutation importance or scaled class-specific importance.

X-Xnpg tc-

X-Xnpg tc-

X-Xnpg tt+

X-Xnpg tt+

X-Pro tc+

X-Pro tc+

X-Gly tt+

X-Gly tt+

The classifiers have been converted (automatically) to FORTRAN IF/ELSE statements for inclusion in WHAT_CHECK:

X-Pro prediction
All X-Pro tc- cases in the test set derived from the PDB-PDB_REDO comparision had a positive φi angle and could be separated perfectly from tt- and tc+ cases, for which the average φ is always around -60°. The rule φ > 0° misclassifies two tc+ instances . When this rule was applied to the entire PDB we also found examples of X-Pro with positive φ angles other than tc- cases. Incorrect chirality of the nitrogen atoms resulted for example from strain in the local backbone because residues i+1 or i-1 needed to be flipped. The class of trans X-Pro residues with N-chirality problems was called nch. The average φ is 96° for tc- and 12° for nch. The nch cases could be separated from tc- cases by a simple rule: if the angle N-Cα-C is large (> 112.47°) and the bump score of the oxygen in the peptide plane is large (> 0.26 WHAT IF bump score units), then the X-Pro with a positive φ is not a cis peptide in need of a tc- flip but a trans-Pro with N-chirality problems. This rule was found by visual inspection of the data using RFScout, a tool that allows the creation of Simple Decision Models. The WHAT_CHECK code for this classifier has been hand-written and not generated automatically.

Cis → trans
We detected only 44 cis → trans flips in the entire PDB. This means not enough data was available to create accurate classifiers. Our method therefore does not include cis → trans flip prediction. Nevertheless we observed that the Cαi-1-Cαi distance and the Ci-1-Ni-Cαi angle tend to be larger for ct+ and ct- cases than for normal cc- tetramers. This can be expected because the data works against the cis restraints. For X-Pro we also observed that the angle τ (Ni-C&alphai-Ci) may help to separate ct+, ct-, and cc- cases:


The five remaining classifiers were tested against an independent test set. The performance of both individual and combined prediction models are shown in the tables below. These tables also show the performance on the subset of test cases that does not include NCS-related flip examples (only the first example in a structure is retained). We deliberately included a few NCS-related examples to observe how sensitive the RF are to small inter-chain variation.

Test performance for all flip-types.

The performance of dual-class (vs. tt-) RF classifiers is shown for the all test cases and for a subset of non-redundant cases without NCS-related flips (separated by a forward slash). The threshold-dependent metrics are at the highest MCC. Note that the metrics are sensitive to class imbalance, except for the AUC and MCC values. WH: Weiss & Hilgenfeld (1999) method with original threshold for Dtot (143.10). WH’: WH with cut-off re-determined in this study (82.256)

X-Xnpg tt+ X-Xnpg tc- X-Xnpg WH’ X-Xnpg WH X-Pro tc+ X-Gly tt+
ROC AUC 0.995/0.9940.983/0.9760.966/0.9790.9660.941/0.9380.982/0.977
Precision-recall AUC 0.995/0.9940.983/0.9760.966/0.9790.9660.941/0.9380.982/0.977
Matthews correlation coefficient 0.911/0.8970.892/0.8290.822/0.7790.3150.852/0.8440.928/0.907
Accuracy 0.979/0.9780.963/0.9680.935/0.9540.8030.924/0.9260.971/0.965
True positive rate (sensitivity/recall) 0.952/0.9380.884/0.8370.909/0.8840.1240.824/0.7950.964/0.952
False positive rate (fallout) 0.012/0.0160.014/0.0160.057/0.0380.00.0/0.00.026/0.031
True negative rate (specificity) 0.983/0.9840.986/0.9840.943/0.9621.01.0/1.00.974/0.969
False negative rate (miss rate) 0.048/0.0620.116/0.1630.091/0.1160.8760.176/0.2050.036/0.048
Positive predictive value (precision) 0.894/0.8820.947/0.8570.821/0.7311.01.0/1.00.931/0.909
Negative predictive value 0.993/0.9920.967/0.9810.973/0.9860.7980.881.0.8950.987/0.984
X-Xnpg test set confusion table.

Combination of tt+ and tc- classifiers. Inclusion of a tc+ classifier would result in overfitting. The classification accuracy is 93.2% including tc+ cases and 95.0 % without tc+ cases. Note that all tc+ cases found in the PDB have been included in the test set.

Actual class
Predicted class tt- tt+ tc- tc+
tt- 405 3 12 6
tt+ 7 59 2 5
tc- 6 0 107 1
tc+ 0 0 0 0
X-Pro test set confusion table.

Combination of tc-/nch and tc+ classifiers. The classification accuracy is 93.3%. nch means an incorrect N chirality of trans-Pro. Note that all 40 tc- cases that could be corrected by PDB_REDO have been included. The classification accuracy is 93.1% when nch cases are excluded.

Actual class
Predicted class tt- tt+ tc- tc+ nch
tt- 89 0 0 12 0
tt+ 0 0 0 0 0
tc- 0 0 58 0 1
tc+ 0 0 0 54 0
nch 0 1 0 2 22
Table VII in the Weiss & Hilgenfeld paper lists the 20 peptides with the highest Dtot score in their 25% database. Our method agrees with half of those tc- preditions (5 cases could not be predicted because they did not pass our data selection and quality criteria). Even though some of the cases we predicted to be tt- were just below the tc- threshold, a detailed comparison of the predictions made by the two methods probably makes no sense because for none of the listed entries structure factors are available. The only entry for which structure factors became available at a later stage was 1aak, which was superseded by 2aak. The WH method predicted tc- for the Arg 6 - Lys 7 peptide bond in 1aak and our method predicted tt-. In 1aak the backbone around these residues is very distorted and φi is positive, which is probably the reason for the high Dtot score. In 2aak the peptide bond has a 'normal' trans and the electron density suggests that trans is the correct conformation. Furthermore, the residues are located inside an α-helix and have the expected hydrogen-bonding pattern in the trans conformation.


Flips can be predicted with the FLPCHK validation routines in WHAT_CHECK, the ShowPepFlips WHAT IF web services option, or the WHAT IF web server. The response of our RF classifiers was translated to qualitative measures of severity based on the validation set. All available methods therefore also report if it is ‘highly unlikely’, ‘unlikely’, ‘somewhat likely’, ‘likely’, or ‘highly likely’ that a peptide needs to be flipped or requires the attention of a structural biologist. The 'highly likely' category has no FP in the test set, the 'likely' category has between 0 and 2% FP in the test set, the 'somewhat likely' category has up to 10% FP, the 'unlikely' category has beteen 0 and 2% FN in the test set, and the 'highly unlikely' category has 0 FN in the test set.


In an effort to classify the small classes (and possibly address the imbalance problem further), X-Xnpg tc+, ct-, and ct+ errors were simulated. The errors were simulated for well-defined peptides selected from medium-sized single chain structures solved at a resolution better than 2.0 Å. WHAT IF (Vriend, 1990) flipped the peptides and relaxed the strain in a 20-residue window spanning the local backbone of the peptides. Subsequently, the flipped peptides were refined. After refinement only a small fraction of the peptide conformations still had a wrong conformation. Classifiers were constructed with 114 fabricated tc+ cases, but a two-class classifier could only correctly classify a third of the true tc+ cases, and none of the true tc+ cases could be distinguished by the four-class classifier. Furthermore, the classifiers did not pick up any previously unrecognized tc+ in the PDB. Finally, the simulated tc+ cases had a broad distribution and showed overlap with all flip classes when they were projected in the Principal Component Analysis space of validated tt-, tt+, tc-, and tc+ cases. Although the number of true tc+ cases very low, the simulated tc+ cases seemed to be more similar to tc- and tt- cases than to tc+ cases in most dimensions. In summary, the simulated flips were not representative of actual flips and could therefore not be used to train classifiers.

Re-building and re-refinement: changes in reciprocal- and real-space coefficients

Single peptide and cis-trans flips will probably have a small effect on the R-factors. The local changes in the protein backbone are expected to lead to an increase in local real-space correlation coefficient. We here present the flip correction and re-refinement of 1hi8, 1i6n, 2z81 and 1pe9 as examples.


The RNA dependent RNA polymerase from dsRNA bacteriophage φ6 PDB structure 1hi8 has been solved at 2.50 Å. The reported R-/R-free factors are 0.280/0.316.

Two cis-peptides are reported in the PDB file:

CISPEP   1 ILE A   96    PRO A   97          0         0.03
CISPEP   2 ILE B   96    PRO B   97          0         0.04
both Ile-Pro peptide planes fit the electron density well:

However, 8 tt+ and Pro tc+ flips are necessary in both chain A and B (a total of 16 flips):

GLY 92  - A tt+
GLY 92  - B tt+
ASP 137 - A tt+
ASP 137 - B tt+
PRO 154 - A tc+
PRO 154 - B tc+
ALA 210 - A tt+
ALA 210 - B tt+
LEU 391 - A tt+
LEU 391 - B tt+
MSE 406 - A tt+
MSE 406 - B tt+
ARG 584 - A tt+
ARG 584 - B tt+
LYS 627 - A tt+
LYS 627 - B tt+
These screenshots show for either the A or the B chain their PDB conformation.

The WHAT_CHECK validation report flags Asp 137 and Pro 154 for having unusual C-N-Cα bond angles and Leu 391 for having unusual torsion angles. Furthermore, Gly 92, Als 210 and Leu 391 have unusual φ/ψ combinations. The buried hydrogen bond donor Gly 92 N (see figure above) is picked up, as well.

10 cycles TLS refinement and 50 cycles restrained refinement in REFMAC with these parameters results in the following re-refinement statistics:

Re-refinement statistics
CC: reciprocal-space correlation coefficient
Structure R-work reported R-free reported R-work initial R-free initial R-work final R-free final Work CC initial Free CC initial Work CC final Free CC final Work CC Z-score Free CC Z-score
1hi8 0.280 0.316 0.2760 0.3082 0.2579 0.2868 0.8893 0.8636 0.9025 0.8836 11.84 3.46
1hi8 rebuilt N/A N/A 0.2749 0.3044 0.2556 0.2825 0.8903 0.8660 0.9043 0.8864 12.72 0.07
Even though REFMAC automatically performs the Ala 210 tt+ flip in 1hi8 in both chain A and B also without rebuilding, all global reciprocal space metrics improve when the incorrect peptides are corrected. The following image shows the improvement in real-space correlation coefficient (RSCC) of rebuilding/refinement over /refinement only, in the 5 residues before and 5 residues after the flipped peptide bonds in chain A. The RSCC is calculated by the Perl script edstats.pl, a wrapper around the CCP4 program edstats (Tickle, 2012). The RSCC values are weighted by the number of grid points that are covered by the groups of main-chain (solid lines) and main-chain + side-chain (dashed lines) atoms. The grey dashed line indicates zero change in RSCC between final re-built and re-refined 1hi8 and 1hi8 deposited in the PDB.

The corrections lead to positive RSCC difference around most flipped peptide bonds, which indicates that the local fit is improved. The improvement is dominated by the backbone atom increase in RSCC. As expected, the fit around Ala 210 is not better in the rebuilt/re-refined structure than in the re-refined structure because the conformation is correct in both structures.


The crystal structure of Bacillus subtilis ioli 1i6n has been solved at 1.80 Å. The reported R-/R-free factors are 0.201/0.238. CISPEP records are absent in the PDB file.

The following trans → cis NH-flip will correct many local problems:

1i6n ALA 67  - A tc-
The local backbone around Ala 67 is also wrong in 1i60 (see above), the model that was used to solve 1i6n by molecular replacement. The difference density shows many peaks around Asn 66, Ala 67 and Leu68 from the PDB model (first picture).

The WHAT_CHECK validation report flags the unusual backbone conformation and bond angles. Re-refinement in REFMAC of the original PDB structure with these parameters improves the backbone significantly (picture 2 and 3). However, ω is 72° after re-refinement, and some difference density is still present (click on the images above for a larger version), probably because REFMAC does not treat the peptide as a cis-peptide yet. Re-refinement of the flipped structure removes all difference density around the flipped peptide bond (picture 4). Before re-refinement, the better fit of the rebuilt structure is visible in the global reciprocal-space metrics. After re-refinement, the global metrics are virtually identical, while the local real-space fit is only optimal after re-refinement of the rebuilt structure (compare pictures 2 and 4). Clearly, a single flip doesn't make summer.
Re-refinement statistics
CC: reciprocal-space correlation coefficient
Structure R-work reported R-free reported R-work initial R-free initial R-work final R-free final Work CC initial Free CC initial Work CC final Free CC final Work CC Z-score Free CC Z-score
1i6n 0.201 0.238 0.1946 0.2280 0.1762 0.2089 0.9294 0.9039 0.9394 0.9170 10.19 2.26
1i6n rebuilt N/A N/A 0.1929 0.2257 0.1761 0.2089 0.9305 0.9051 0.9393 0.9169 9.03 2.05
This is also shown by the extra increase in real-space correlation coefficient:


The crystal structure of the Toll-like receptor 2 2z81 has been solved at 1.80 Å and refined to reported R-/R-free factors of 0.213/0.232. CISPEP records are absent in the PDB file.

1 Pro tc+ and 5 tt+ flips are necessary in 2z81:

2z81 SER 42  - A tt+
2z81 ARG 87  - A tt+
2z81 GLY 384 - A tt+
2z81 GLN 396 - A tt+
2z81 PRO 540 - A tc+
2z81 ARG 541 - A tt+
These screenshots show their PDB conformation.

The WHAT_CHECK validation report reports τ (N-Cα-C) angle problems around these bonds as well as poor φ/ψ combinations. Arg 87, Gly 384 and Gln 396 are listed as buried unsatisfied hydrogen bond donors.

The re-refinement statistics after 10 cycles of TLS refinement and 50 cycles restrained refinement in REFMAC with these parameters are slightly better for the rebuilt/re-refined structure than the re-refined-only structure:

Re-refinement statistics
CC: reciprocal-space correlation coefficient
Structure R-work reported R-free reported R-work initial R-free initial R-work final R-free final Work CC initial Free CC initial Work CC final Free CC final Work CC Z-score Free CC Z-score
2z81 0.213 0.232 0.2088 0.2224 0.1747 0.2135 0.9418 0.9381 0.9594 0.9435 31.30 1.84
2z81 rebuilt N/A N/A 0.2056 0.2181 0.1733 0.2094 0.9437 0.9403 0.9602 0.9457 30.12 1.91
The Arg 541 tt+ flip is automatically performed by REFMAC during refinement of the original 2z81 structure. Therefore, the real-space correlation coefficient is not higher for the rebuilt/re-refined Arg 541. The other RSCC values are however much higher if the necessary peptides are flipped in the 2z81 structure prior to re-refining it. The side-chain Pro 541 atoms are also much better modeled in the re-built and re-refined structure:


The crystal structure of pectate lyase A 1pe9 has been solved at 1.60 Å and refined to reported R-/R-free factors of 0.198/0.213. 1 CISPEP record is present in the PDB file:
CISPEP   1 ALA B  242    PRO B  243          0        -0.11

However, there are two molecules of pectate lyase A in the asymmetric unit, and the same Ala-Pro dipeptide in needs to be flipped in chain A:

1pe9 PRO 243  - A tc+
These screenshots show both chains. There are still some difference density peaks around B Pro 243, but they have vanished after re-refinement.

The WHAT_CHECK validation report reports an unusually small Cδ-N-Cα angle for Pro 243 A and unusually short τ (N-Cα-C) angles for Ala 242 and Arg 244. Furthermore, the improper dihedral of the Pro 243 nitrogen deviates almost 10 σ from normal values, indicating distorted chirality. The puckering amplitude is also very high and the puckering phase is unusual. The torsion angles around Pro 243 are also unusual and the many bumps are present.

The following re-refinement statistics were obtained after 10 cycles of TLS refinement and 50 cycles restrained refinement in REFMAC with these parameters:

Re-refinement statistics
CC: reciprocal-space correlation coefficient
Structure R-work reported R-free reported R-work initial R-free initial R-work final R-free final Work CC initial Free CC initial Work CC final Free CC final Work CC Z-score Free CC Z-score
1pe9 0.198 0.213 0.1906 0.2078 0.1619 0.1815 0.9482 0.9432 0.9598 0.9541 28.20 5.48
1pe9 rebuilt N/A N/A 0.1894 0.2065 0.1612 0.1816 0.9489 0.9439 0.9601 0.9544 27.51 5.32
The re-refinement statistics are virtually the same for the flipped and original 1pe9 structure. After re-refinement, the original structure is almost fixed (picture 1), but the ω is still 77°. Only re-refinement with the flipped Ala-Pro peptide resolves the difference density problems (picture 3):

There is no extra improvement in chain B, but in chain A there an extra increase in RSCC is visible at Ala 242 - Pro 243.