How to Use JADBio

WORKING in the JADBio application

CREATING A PROJECT

All of the datasets and analyses within the JAD Use Cases project are read only. In order to start working with data, you must create a new project. You can create a new project and upload data from the JAD Use Cases project.
On the MENU sidebar at the left of the JADBio window, click on Projects.
This will bring you to the Projects window, which includes the Create a new Project function.
Click on the CREATE button.
Name your Project adding a Description, if you wish to. These can be altered at any time.
Click on the CREATE button.

Your new project will now appear in the Projects window.

UPLOADING A DATASET

Under ACTIONS, in the top left corner, click Add Data  to add a dataset to this project.

You will have two options. Either working with an external file, where you would select the Upload a file option, or use an existing dataset.

This opens the Assign feature types window. JADBio automatically assigns feature types based on the content of each column, but you can change the feature types here or in a transformation step to suit you understanding of your data. Here, you can change the type of features i.e. to assign an event or a time-to-event feature type. Also, as in the Preview dataset window, you are able to navigate across the dataset, search and filter the dataset, and view the metrics of each feature in the Metrics view.

Click SKIP to move the data, without changing any feature type, into your project.

Under the Actions header on the three dots, you have three options: Preview Dataset, Perform Analysis and Detach Dataset from project.

TRANSFORMATIONS

JADBio provides numerous transformation tools today and we will continue to augment this functionality to support your needs. For this demonstration, you will split the data into two datasets, one for training and one for testing.

Click on Transform Data in the ACTIONS sidebar.

Data Transformation takes place in two steps:

1. Select transformation type and
2. Transformation configuration, and JADBio currently provides four different transformation types: Filter, Merge, Split and Change dataset. You are going to create two datasets, test and training, from the imported dataset, so you will use the Split function.

Select the Split option – Note: Split samples randomly into two datasets.
Click NEXT.

This opens the wizard, Configure sample splitting.

Click on the TRANSFORM button.

BACKGROUND

Up to now, the prediction of complex phenotypes in plants like potatoes was based on growing plants and assaying the organs of interest in a time intensive process. Conventional genetic biomarkers are commonly applied in quality assessment and progeny selection but their application is still problematic for complex traits such as yield, disease resistances or stress tolerance.

A simple and accurate predictive test for potatoes yield quality could be derived from the analysis of potato metabolic biomarkers. In this demonstration, you will create, train and test a model based on the data from on 20 different potato cultivars grown in two Northern German potato farms, as described in the Steinfath et al. publication referenced below. These data include 203 metabolites measured on the tubers, Soil (loamy versus sandy) and climate (coastal versus inland) information and two traits ‘susceptibility to black spot bruising’ and ‘chip quality”. You will see that with JADBio you will produce accurate predictive models that are on par with the ones presented in the original publication, regardless of your expertise in advanced machine learning techniques.

Original publication: Steinfath et al, Plant Biotech Journal 2010

ANALYSIS

With both, a training dataset and a test dataset, you are able to train and test a predictive signature based —in our example, the potato metabolic profiles and climate/soil information.

JADBio will automatically create a descriptive name for your analysis based on your selections, which you can change.

Name the Analysis: Potatoes_training_preliminary

While there are many other options, the default values for the advanced settings are the following:
There is no requirement for an Identifier feature, to group the samples of the dataset.
The analysis will optimize performance based on the AUC (area under the ROC curve).
Maximum signature size is 25 features.
Maximum multiple signatures to visualize is 5.
JADBio will create four plots, PCA, UMAP, ICE, and Probabilities.

Click RUN ANALYSIS to start the analysis.

As soon as you begin the analysis, JADBio reports Progress in the Analyses table.

Click on the blue progress bar to view learning progress, you can watch each step in the analysis progression.

Click on the main window to return to close the Analyses window.

You do not have to wait for the learning progress to finish in order to move away from the window in your browser. You can also monitor when it will finish through your Dashboard.

Finally, you can view the results by selecting Show Results in the pop-up Learning Progress window, if it’s completed or by selecting the final analysis.

ANALYSIS RESULTS

Select the finalized analysis (under the Analyses tab)

In the new window you will find a sidebar menu, on the left, for ANALYSIS ACTIONS. This provides options to:

• Summarize the analysis (as an Analysis Report),
• Show Model (if it is an interpretable one)
• Download Predictions of the samples
• Apply Model to new samples

Click on Analysis Report  

Here, JADBio provides a summary audit of the analysis including: Dataset, Outcome to Predict, Analysis Type, etc. and a description of the selected options as well as the JADBio version.

JADBio also presents a list of all of the configurations that were tested in order to produce the model and selected features. The list is not long in this example, but more complex analyses may test up to 1000s of configurations. Specifically, for this analysis 5 configurations were tested which you can download by clicking on the DOWNLOAD CSV button.

Click on return button to return to the main analysis results page.

BEST PERFORMING MODEL

The Analysis page provides an overview of the analysis process and a description of the Best Performing Model and Best Interpretable Model.

The methods include the optimal configurations for: Preprocessing, Feature selection, and Predictive algorithm.

PERFORMANCE OVERVIEW

Optimize classification threshold as: 0.82147636, and note the selection of the position on the ROC curve.

Hover over the highlighted ROC curve to see the full range of metrics at this threshold.

Click on the ROC plot | Prediction recall plot button to view the Precision recall plot.

FEATURE SELECTION

Feature Selection is a process that identifies a minimal-size subset of features that is maximally predictive of the outcome of interest, the selected target feature.

Scroll back to the top of the page, and Select the Feature Selection tab.

The Individual Conditional Expectation (ICE) plots further reveal the nature of the contribution of each metabolite feature to the model.

Click on the thumbprint ICE plot adjacent to Fructose-A187002 feature to enlarge the ICE plot.

This opens the ICE plot for the prediction of “high” quality classification. In this plot, you can see that as the metabolite level increases, the likelihood of a “high” quality classification decreases.

Use the pulldown to select the Class low.

As one would expect for a binary classification, the level of Fructose-A187002 metabolite has the inverse correlation to a low-quality classification.

ANALYSIS VISUALIZATION

Change the Display as selection from UMAP to PCA.

Principal component analysis (PCA) is another dimensionality reduction technique that seeks the linear combinations (Principal Components) of the original features such that the derived features capture maximal variance.

The appropriate visualizations for this model are UMAP and PCA plot, in which the low and high quality predictions are visualized. Other types of analysis and other types of models would likely result in different visualizations. For instance, a survival (time to event) analysis would have resulted in a Kaplan-Meier curve.

Scroll down to see the Probabilities plot.

The Probabilities plot shows, for both classes, the probability of the prediction resulting in the high class. An ideal plot would have complete separation between the two classes. You can also visualize the probabilities in a box plot.

APPLY MODEL

Usually the analysis is performed on a training dataset. Once the analysis of the trained model is complete, you can run the model on your test dataset or any other data.

In Select outcome check the feature box Quality.

There are four steps to the validation process:

Select Analysis
Check the only analysis option available in the Project, Potatoes_training_preliminary, which you just created and reviewed.
Click NEXT.

2. Select model and signature

• Select model: There is only model available, which is already selected.
• Select signature: Place a check in the second row of the four statistically equivalent signatures.

Click NEXT.

3. Select how to apply the model

Select the option, Validate model against labeled samples.

Click NEXT

This selection allows JADBio to use the test dataset you created, when you split the original dataset into training and test datasets. If you choose to Predict outcomes for Unlabelled samples, JADBio will give you the option to upload another dataset. The only requirement you will have for this dataset is that it includes the Predictors in the selected signature with the same names as in the training data. If you select to Predict outcomes for manually entered predictor values, JADBio will provide a dialog for manually entering values for each of the Predictors in the Signature to generate a prediction. If you choose to Download model to make predictions off-line will give you the options to download a standalone version of the chosen model that can be applied on new data on a local machine.

4. Select labeled dataset

Select the dataset: Potatoes_test
Click APPLY MODEL

JADBio will automatically bring you to the Applied Models window, where you can view the results of your Applied Model on the test data set.

Click on the Actions function, View results to open the results window.

The results window is similar to the original training dataset window, but with some exceptions. In the VALIDATION ACTIONS sidebar, there are buttons to Download Predictions of the samples and Apply Model to new samples.

Define Class High as the positive class, using the radio buttons on the top.

Now, JADBio displays the AUC and the Average Precision of the validation and the train datasets

The new ROC and Precision recall curves include the results from the original Train — data and from the test data, Validation —.

JADBio also displays the confusion matrix for the Validation and the Model (training) datasets and the 13 threshold dependent metrics.

Click on the Analysis Visualization tab.

Here, the Supervised Principal Component Analysis displays the segregation of the low and high class test data based on the model from the training data.