Characterization and Integration of Driven Omics
CHiDO is a no-code application developed by the Jarquin Lab at the University of Florida's Department of Agronomy. The application aims to remove technical and financial barriers associated with modeling Genotype-by-Environment (GxE) interactions.
There is a demo version of CHiDO hosted as a free web application at: https://jarquinlab.shinyapps.io/chido/
- Francisco González - University of Florida (UF)
- Julián García - Universidad Politécnica de Madrid (UPM)
- Vitor Sagae - University of Florida (UF) and Federal University of Viçosa (UFV)
- Dr. Diego Jarquin - University of Florida (UF)
The following packages, listed in requirements.txt, are necessary to
run CHiDO locally: - shiny - shinyjs - shinyjqui - shinydashboard -
dplyr - ggplot2 - viridis - scales - gridExtra - DT - zip - markdown -
BGLR - Matrix - lme4 - mvtnorm - rstiefel - tmvtnorm
Once you clone/download the repository, ensure you are in the CHiDO directory and run the following command in the terminal:
Rscript -e 'packages <- readLines("requirements.txt"); for(package in packages) if (!require(package, character.only = TRUE)) install.packages(package, dependencies = TRUE)'This will ensure all prerequisite libraries and any of their dependencies are installed in your local environment.
- Clone the repository:
git clone https://github.com/jarquinlab/CHiDO.git - Navigate to the project directory:
cd CHiDO - Install required packages by using the command mentioned in the Prerequisites section.
- Run the application:
shiny::runApp()or pressing "Run App" in RStudio.
There are four main steps for using CHiDO: (1) loading the data, (2) building models with the data, (3) training/testing the models, and (4) viewing/downloading results.
Data format supported: CSV
Data size limits: 50MB
CHiDO can fit a genotype and/or a parent level model (combining ability).
For genotype level model at minimum, users must upload a file with the following columns:
- Genotype ID: Identify the genotypic lines that were examined in the field trials, this can be a numeric or non-numeric identifier.
- Environment ID: Identify the environmental unit (e.g. plot) through which the field trials were split, this can be a numeric or non-numeric identifier.
- Target trait: The phenotypic trait to predict. There can be multiple columns for each phenotypic trait but only one can be defined as the target trait per run.
For parent level model at minimum, users must upload a file with the following columns:
- Genotype ID: Identify the genotypic lines that were examined in the field trials, this can be a numeric or non-numeric identifier.
- Parent Group 1 ID: Identify the first group of parents, this can be a numeric or non-numeric identifier.
- Parent Group 2 ID: Identify the second group of parents, this can be a numeric or non-numeric identifier.
- Environment ID: Identify the environmental unit (e.g. plot) through which the field trials were split, this can be a numeric or non-numeric identifier.
- Target trait: The phenotypic trait to predict. There can be multiple columns for each phenotypic trait but only one can be defined as the target trait per run.
Additional datasets, or "omics", can be uploaded on the right-hand box of the Data Upload page (image below). These files should have at least one ID column that can be linked back to the file containing the IDs and phenotypic data.
When you upload an omic file, you must specify a label (e.g., G for genomics) and the column where the IDs are located. If you select "other" as the data type or using the combining ability model, a dropdown menu will be displayed where you can select how the ID column should be linked back to the first file.
The combining ability model is intended for hybrid modeling. In this case, users can upload omics data specifically for each parental group. For example, the genomic marker matrices where available genotypic information corresponds to the parents of phenotyped individuals, can be uploaded for each of the parental groups (Parent Group 1 ID and Parent Group 2 ID), or for both groups using an unique genomic marker matrix containing males and females genotypes (Parents Groups ID). The user can also provide a unique Genomic Kinship Matrix containing both parent groups instead of entering the molecular marker information (genomic relationship matrix as data type). instead of molecular information. IDs should corresponds to the chosen option in a single linkage column.
- Check data consistency function will remove those rows whose ID in the omic data are absent in the phenotypic file following the linkage columns.
Every omic uploaded will have a label (e.g., G for genomics) and this label will be displayed as a draggable icon in the Model Assembly page. By default, the environment ID and genotype ID columns are used to create two random effects displayed with the labels E and L, respectively.
In the top-left box (image below), all omics available will be displayed via their labels and can be added as terms to a linear mixed model as:
- Main effects by dragging a single label into the top-right box and pressing "Add".
- Interaction terms by dragging two or more labels into the top-right box and pressing "Add" (e.g., G and E are added together and shown as G*E in the model preview).
There is also the option of writing the model directly into a text entry box.
There are three buttons in the model preview section
- Save: Saves the current model and launches a pop-up box to define a name for it.
- Previous: Removes the last term added to the model.
- Clear: Removes all terms
Multiple models can be created and saved.
In this section, you can define the parameters and configurations to train and test your saved models.
Hyperparameter Tuning
CHiDO uses the BGLR library to train and fit the models, which uses a Markov Chain Monte Carlo (MCMC) sampling process. You can adjust this by using the following two parameters in interface.
-
Number of iterations: The total number of MCMC iterations that the algorithm will perform. During these iterations, the algorithm samples from the posterior distribution of the model parameters.
-
Burn-in rate: The number of initial iterations that will be discarded from the MCMC sampling process, allowing the algorithm to reach a stable state by discarding early samples that might be biased due to the starting values. Usually, this value ranges anywhere from 10% to 50% of the total iterations.
You can also set the seed using any fixed number of your choice.
Data Pre-processing
In this section, you can define the number of folds you will split your data into to perform cross-validations.
Additionally, you can define the NaN Threshold -- this variable corresponds to the limit of NaN values in a column for an omic before discarding that column prior to creating the covariance matrices that are used as model terms. For example, if you have a file containing genomic markers and set the NaN Threshold to 20%, any marker with more than 20% missing values will be discarded.
Currently, we only have two data transformation processes to choose from
- Centering: Subtracting the mean of the variable from each individual value so that the mean of the variable becomes zero.
- Standardizing: Centering the variable and scaling it by its standard deviation to have a mean of zero and a standard deviation of one.
Cross-Validation
Finally, in this section you can begin training and testing your model(s) by choosing the model from a drop-down menu and selecting one or more of four cross-validation (CV) schemes.
In a general sense, these CV schemes are used for the following:
- CV1 to predict untested genotypes in observed environments
- CV2 to predict tested genotypes in observed environments
- CV0 to predict tested genottypes in new environments
- CV00 to predict untested genotypes in new environments
For each cross-validation, the data will be sorted following the appropriate methods for it and it will be split into the number of defined folds.
There is a terminal box that, once you press "Run", displays the progress in creating, training, and testing the models. Upon completion, a pop-up box will be displayed notifying that results from the cross-validation are ready.
You can view your results in two methods, as tables or graphs. In addition to the CV results, you also have the following evaluation metrics:
- Prediction accuracy
- Root-mean-square-error (RMSE)
- Variance components
These results can be downloaded as a zip folder and the contents are split by model tested. For each model, there will be CSV files containing the raw results for each CV selected, PNG files with images of the plots generated, an RDA object containing the fitted model, and RDS files for each plot (created with ggplot2). A sample output can be located in the example folder.
This project is licensed under the GNU Affero General Public License - see the LICENSE file for details.
For questions about using CHiDO, please visit our documentation page for guidance.
If you encounter any issues with the application or have an idea to extend its capabilities, please submit it to our GitHub issues page, found here.
For any additional concerns or interest in collaborating with the project, please email Dr. Diego Jarquin (jhernandezjarqui@ufl.edu{.email}).