Welcome to TopSuite: A deep-learning meta-server for computational structural biology

TopScore predicts the structural similarity between a protein model and the native structure by combining model quality predictions from 15 different primary predictors using a two-stage deep neural network. TopScore was validated on over 780 different protein targets totaling ~1.5x105 models and ~2.3x107 residues.
    These models originated from diverse sources including homology models for different difficulties, ab-initio folded models, CASP models and models from different decoy data-sets. TopScore has a significantly better and more robust performance than any of its primary predictors. TopScore provides both a global and a residue-wise measure of the error in the protein.


TopModel performs template-based structure prediction using top-down consensus and deep neural networks. The input sequence is submitted to 12 primary servers for template identification. TopModel then uses a series of deep neural networks to combine primary threader scores and model quality scores from TopScore to predict the template quality in terms of TM-Score to the native structure. TopModel is much more sensitive at detecting distantly related templates than any of its primary predictors even when the target has no truly homologous structure.
    Following template identification, TopModel generates a large ensemble of different alignments for different combinations of the top ranked templates using 18 different primary alignment programs. Models built from this large alignment ensemble are then scored with TopScore.
    After constructing multi-template models, TopModel selects the best scoring single-template and multi-template models, and combines them into a single refined model. This is done by deleting regions in the models predicted by TopScore that contain errors and re-combining the remaining parts into a final refined model.


TopDomain performs domain boundary prediction using multi-source information from over 50 primary predictors combined using deep neural networks. TopDomain takes either an input structure or an input sequence and predicts boundaries using sequence-based, homology-based and structure-based information. For input sequences, TopDomain also predicts if parsing the sequence into domain at the predicted boundaries is required for structure-based prediction of the domain structures.


TopProperty is a meta-predictor that combines outputs of 27 primary predictors using two ensembles of deep neural networks to predict secondary structure, solvent accessibility, transmembrane topology and membrane exposure for both transmembrane and globular proteins. TopProperty is trained on datasets without bias towards a high number of sequence homologs and the predictions are significantly better than all primary predictors on all quality metrics. TopProperty eliminates the need for protein type- or property-tailored tools, especifically for transmembrane proteins.

Impressum - Datenschutzerklärung der Heinrich-Heine-Universität Düsseldorf