Required - Job name: you must input a valid job name in the text box labeled as "Job name". A valid job name can not have any space(s) and/or special characters except dash (-) and underscore (_). The job name should be between 4 and 10 characters long. Outputs, including refined models and post-refinement analyses files, will be generated based on your provided job name.

Required - Starting structure: you must provide a starting structure for the refinement. You can either copy and paste a valid starting structure in the textarea labeled as "Copy and paste the starting structure" or upload a valid starting structure using the upload button labeled as "Or upload the starting structure from your local computer". The starting structure must conform to the valid PDB format and the length must be between 10 and 500 residues. Here you may find the information about valid PDB format.

For testing purpose, you may either populate the textarea with an example starting structure by clicking the "Example" button located at the top right corner of the textarea or download the "example.pdb" by clicking the "Example" button located at the bottom left corner of the "Choose file" button and subsequently upload as a starting structure for processing.

Optional - Email address: although an email address is not required, it is recommended that you provide a valid email address in the field labeled as "Email address" for receiving the job-related update. Your email address will not be public and only be used for communicating the job status.

DeepRefiner provides a range of options that allows you to calibrate your job by customizing the deep learning model, refinement mode, post-refinement analysis, and privacy.

Deep learning model: DeepRefiner utilizes two different deep learning models, Residual Neural Networks (ResNets) and Deep Convolutional Neural Fields (DeepCNF), in guiding the refinement with a default set to ResNet. You may either select ResNet or DeepCNF guided refinement. However, if the starting structure contains gap(s), the refinement will be set to DeepCNF.

Refinement mode: DeepRefiner offers two different modes of refinements, "Adventurous" and "Conservative" with a default set to "Adventurous". "Adventurous" mode performs a higher degree of refinement using non-cumulative restraints. On the other hand, the "Conservative" mode performs more consistent refinement using cumulative restraints.

Post-refinement analysis: DeepRefiner by default performs numerous post-refinement analyses for effectively evaluating the stereochemical qualities of the refined models. They include,

You may selectively customize the post-refinement analyses from the section, labeled as "Post-refinement analysis".

Privacy: your submitted job and corresponding results will be accessible publicly unless you choose to keep your job private by selecting the checkbox marked as "Keep my job private". If you keep your job private, no user including you, will be able to access the results directly from the job queue. However, you may still limitedly track the progress of your job from the job queue. In such a scenario, you can access the corresponding results only by using the prompted URL at the time of job submission. The URL will also be sent to your valid email address if you provide one. However, if you choose not to provide a valid email address, you should record the job URL from the job submission confirmation window.

DeepRefiner allows you to track your submitted job status as it is being processed. Keeping your job public enables you to directly track the status from the job queue as well as access the results. Therefore, your job will be clickable from the "Job queue" panel as shown below,


Currently, "Job queue" promptly shows one of the following status for a valid job,

Once you click the job name, it will navigate you to the job status page, showing real-time job status as follows,


Additionally, in the "Submission detail" panel, you can see the user-specified "Job parameters" and the starting structure visualization.

Once your job has finished processing, you can navigate to the result analysis page by clicking the "Show the prediction" button. Meanwhile, you can download the results by clicking the "Download" button.

Tracking a job using the URL

The results page contains the visualization of the top 5 refined models as shown below for the example results.


Refined models are ordered from model 1 to model 5 based on the global quality scores, predicted either by ResNet or DeepCNF, specified deep learning model.
For a better view, you may see the individual model by clicking the respective tab marked as "Model 1" to "Model 5" as shown below,


Additionally, the model viewed through the specific tab, offers multiple options to visually inspect the models that you can access from the control panel at the bottom.

Stereochemical qualities appear next to the model view panel. You can see the model-specific stereochemical quality in the panel marked as "Stereochemical quality" as shown below,


Similarly, you may see the stereochemical qualities of all models or individual model through a specific selection.

A brief tabulated description of each of the stereochemical qualities is provided below,

Stereochemical quality	Overview	Score
Predicted global quality score	Predicted by either ResNet or DeepCNF, specified by the users.	Ranges between [0, 1] with a higher value indicating the better global predicted quality
Rosetta energy score	Calculated by Rosetta's optimized scoring function "ref2015".	A lower score indicates a better model quality.
MolProbity score	A log-weighted combination of the clash score, percentage of Ramachandran not favored, and the percentage of bad side-chain rotamers	A lower score indicates a better model quality.
GOAP score	Energy score based on Generalized Orientation-Dependent, All-atom Statistical Potential.	A lower score indicates a better quality.
OPUS-PSP score	Orientation-dependent statistical all-atom potential derived from side-chain packing.	A lower score indicates a better quality
DFIRE score	Energy score derived by finite ideal-gas reference state.	A lower score indicates a better quality
RWplus score	Distance-dependent atomic potential based energy score.	A lower score indicates a better quality

DeepRefiner provides the comparisons of secondary structure and solvent accessibility similarity between the starting structure, and refined models for in-depth model analysis in the panel labeled as "Sequence, secondary structure and solvent accessibility".


Here the top row represents sequence number based on pose residue numbering. Colon (:) denotes a match between the secondary structure and solvent accessibility of a residue in the starting structure and corresponding residue in the refined model, space (' ') otherwise.

DeepRefiner additionally provides comparisons of refined models with respect to the starting structure. This further helps to evaluate the degree of changes through refinement. You may see the comparisons in the panel labeled as "Comparison to the starting structure", as shown below.


The comparisons are reported based on the following metrics,

Metric	Overview	Score
GDC-sc score	Global Distance Calculation for Side-Chains (GDC-SC), indicating the correctness of the side-chain positioning	Ranges between [0, 100] with a higher value indicating the better side-chain positioning with respect to the starting structure
GDT-HA score	Global Distance test calculated as the proportion of aligned residues between the refined model and the starting structure at 0.5, 1, 2, and 4Å thresholds.	Ranges between [0, 100] with a higher value indicating the better backbone positioning with respect to the starting structure
GDT-TS score	Global Distance test calculated as the proportion of aligned residues between the refined model and the starting structure at 1, 2, 4, and 8Å thresholds.	Ranges between [0, 100] with a higher value indicating the better backbone positioning with respect to the starting structure
Cα RMSD score	Root-mean-square deviation (RMSD) is calculated as the average distance between the Cα atoms in the residues of the refined model and the corresponding residues in the starting structure.	A lower value indicates better similarity in terms of Cα positioning with respect to the starting structure

DeepRefiner analyzes the models' quality at the residue-level by providing interactive graphical representations of the residue-level error estimates in the panel labeled as "Predicted local quality" as shown below,


The x-axis in the graph shows the sequence number based on the pose residue numbering and the y-axis shows the predicted local quality scores. Here a lower score represents better quality. Reliable and unreliable regions can readily be identified through this graphical visualization.