teal.slice Classes

NEST CoreDev

This vignette provides a deep dive into the design of the teal.slice package. It is intended for advanced developers. If you want to merely use teal.slice, the Filter panel for developers vignette should satisfy all your needs.


The teal.slice package is composed of multiple classes, whose primary purpose is to provide a shiny module for managing and displaying filters. The top class in the structure is FilteredData. Other modules in the app can interact with FilteredData to make filter calls and obtain filtered data.

The FilteredData object contains one or more FilteredDataset objects. While FilteredData manages the entire filter panel, each FilteredDataset is responsible for filtering a single data set. FilteredDataset contains one or more FilterStates objects, each corresponding to one data structure like a data.frame, a SummarizedExperiment, or a MultiAssayExperiment. FilterStates holds a collection of FilterState objects, each representing a single filter condition applied on one variable.

FilterStates can also hold FilterStateExpr objects. This is a variation of FilterStates that focuses on the filter expression regardless of the underlying data. The expression can refer to one or several variables.

FilteredData, FilteredDataset, FilterStates, and FilterState/FilterStateExpr are all R6 classes and objects of those classes form a hierarchy but there is no inheritance between them.

Each FilterState/FilterStateExpr object contains one teal_slice object. teal_slice stores all information necessary to define a filter. It is not an R6 class, it does not have any innate methods or behaviors and does not produce any shiny modules. Its sole purpose is to store filter information.


As part of the teal workflow, FilteredData, FilteredDataset, and FilterStates are created instantly when the data is loaded and remain unchanged. One FilteredData object is initialized on a list of objects (e.g. data.frame, MAE). A FilteredDataset is initialized for each data set. One or more FilterStates are initialized, depending on the type of data set.

On the other hand, a FilterState is initialized each time a new filter is added. The values of the FilterState can be changed, and it can also be removed and added again.

The key mechanism in the new filter panel is in FilterStates class. One can think of FilterStates as equivalent to a single, possibly compound, subset call made on one data structure. While FilterState represents a logical predicate on one vector, e.g SEX == "F", FilterStates will compose all predicates of its member FilterState objects into a call to a subsetting function, e.g. data <- subset(data, SEX == "F" & RACE == "CAUCASIAN").

In the case of a data.frame, a single dplyr::filter call is sufficient to subset the whole data set. A MultiAssayExperiment on the other hand contains patient data in the @colData slot and multiple experiments in the @ExperimentList slot, and all of these objects have to be filtered by separate subsetting calls. Therefore, subclasses of FilterStates exist to handle different kinds of data structures and they use different subsetting functions.

Class Description

This section provides a detailed description of all classes that make up the filter panel structure.


FilteredData is an object responsible for managing the filter panel. It sits on top of the class structure and handles the shiny modules of the subclasses.

FilteredData provides several API methods that can be used to access reproducible subsetting calls and the resulting filtered data. It also allows external modules to manage filter states through functions such as get_filter_state, set_filter_state, remove_filter_state, and clear_filter_state.


FilteredDataset is a class that keeps unfiltered data and returns filtered data based on the filter call derived from the contained FilterStates. FilteredDataset class objects are initialized by FilteredData, one for each data set. FilteredDataset contains a single data object and one-to-many FilterStates depending on the type of that object. FilteredDataset stores data set attributes, joins keys to other data sets, and also combines and executes the subsetting calls taken from FilterStates.

The following FilteredDataset child classes are currently implemented:


When the app starts, FilteredDataset initializes one or more FilterStates objects, one for each component of the underlying data set. Every FilterStates object is responsible for making one subset call. For example, a MAEFilteredDataset will create one FilterStates for its colData and one FilterStates for each of its experiments. Every FilterStates will return a separate subsetting call, which will be used to subset the entire MultiAssayExperiment.

The following FilteredStates child classes are currently implemented:

FilterStates serves two shiny-related purposes:


This class controls a single filter card and returns a condition call that depends on the selection state. A FilterState is initialized each time FilterStates adds a new filter. Different classes of data require different handling of choices so several FilterState subclasses exist and each of them presents a different user interface and generates a different subsetting call. A FilterState is created as follows:

  1. FilterStates creates teal_slice with dataname based on the parent data set and varname based on the selected variable
  2. the teal_slice is wrapped in teal_slices and passed to FilterStates$set_filter_state
  3. FilterStates$set_filter_state_impl is called
  4. FilterStates$set_filter_state_impl calls init_filter_state, passing the appropriate variable as x
  5. init_filter_states is a generic function that dispatches x, teal_slice, and other arguments to the respective child class constructor:

All child classes handle missing values, and RangedFilterState also handles infinite values.

The FilterState constructor also takes the extract_type argument, which determines how the call is constructed extract_type can be unspecified, "matrix" or "list" and its value corresponds to the type of the variable prefix in the returned condition call. If extract_type is "list", the variable in the condition call is <dataname>$<varname>, while for "matrix" it would be <dataname>[, "<varname>"].


Similarly to FilterState, FilterStateExpr controls a single filter card and returns logical expression. However, while FilterState generates the call based on the selection state, in FilterStateExpr the call must be specified manually and it must be a valid R expression.


teal_slice is a simple object for storing filter information. It can be thought of as a quantum of information. A teal_slice object is passed directly to FilterState$initialize and is stored inside of the FilterState to keep the current state of the filter. Technically, all values used to generate a call are in teal_slice. FilterState can be described as a wrapper around teal_slice that provides additional methods to handle filter state. It also contains the actual data (a single column).

While teal_slice is not an R6 object and does not encode any behaviors, it is implemented around a reactiveValues object to allow shared state in advanced teal applications.

See ?teal_slice for a detailed explanation.

Making a reproducible filter call


The diagram above presents the filter panel classes and their responsibilities when composing filter calls. Code is generated by nested execution of get_call methods.

FilteredData$get_call calls FilteredDataset$get_call, which calls FilterStates$get_call which in turn calls FilterState$get_call.

FilterState$get_call() returns a single subsetting expression (logical predicate).

FilterStates$get_call() returns a single complete subsetting call by extracting expressions from all FilterState objects and combining them with the & operator.

FilteredDataset$get_call() returns a list of calls extracted from all FilterStates objects.

FilteredData$get_call(<dataname>) returns a list of calls extracted from the specified FilteredDataset.


Calling datasets$get_call(<dataname>) in teal modules executes a chain of calls in all filter panel classes. Consider the following scenario:

  1. FilteredData contains three data sets: ADSL, ADTTE, MAE, each stored in its own FiteredDataset object

  2. ADSL is a data.frame so it can be filtered with a single dplyr::filter call. This data set is stored in DataframeFilteredDataset, which holds a single FilterStates object.

  3. FilterStates constructs a dplyr::filter call based on the FilterState objects present in its private$state_list.

  4. When FilterStates$set_filter_state adds a new teal_slice, a FilterState is created and added to private$state_list in FilterStates. FilterStates gathers logical expressions from all of its FilterState objects and composes them into a dplyr::filter(ADSL, ...) call.

  5. Two new filters have been added: SEX and AGE. This causes initialization of appropriate FilterState subclasses: ChoiceFilterState and RangeFilterState. Each FilterState produces a subsetting expression: SEX == "F" and AGE >= 20 & AGE <= 60. The expressions are combined with & and passed to dplyr::filter, producing ADSL <- dplyr::filter(ADSL, SEX == "F" & AGE >= 20 & AGE <= 60). DataframeFilteredDataset puts this call in a list and returns it to FilteredData.

  6. ADTTE is also a data.frame so the FilteredDataset that stores it works much the same as the one for ADSL. The one difference is that the dplyr::filter call for ADTTE is followed by a dplyr::inner_join call to merge the filtering result with the parent data set (ADSL) so that key columns remain consistent. Note that this is only done when join_keys is specified - otherwise ADTTE would be treated as a separate data set and filtered independently.

  7. The MAE data set is a MultiAssayExperiment, which contains multiple sub-objects which can be filtered on. One of them is ADSL-like patient data, stored as a DataFrame in MAE@colData, and others are experiments, typically SummarizedExperiment objects, stored in MAE@ExperimentList, which can be extracted using MAE[["experiment name"]]. Therefore, MAEFilteredDataset has multiple FilterStates objects: one for subject data and one for each experiment.

  8. A MAEFilterStates object is initialized for subject data and for this object MultiAssayExperiment::subsetByColData function is applied. MultiAssayExperiment::subsetByColData has two arguments: x (data) and y (conditions). Since all filter expressions are passed to one argument, MAEFilterStates only has one state_list, just like DFFilterStates. Adding new filters triggers the same actions as described in (4).

  9. A SummarizedExperiment is more complicated as observations can be filtered based on its rowData and colData slots, both contain DataFrames. Subsetting is done by a dedicated S4 subset method, which takes two key arguments: subset takes logical expressions that will be applied to rowData, and select takes logical expressions that will be applied to colData. teal_slice objects that specify filters in a SummarizedExperiment must contain an arg element, either "subset" or "select", to reflect which slot of the experiment they refer to. The SEFilterStates gathers logical expressions of its member FilterState objects, groups them by the arg element, and builds a call so subset with two combined logical expressions passed to the subset and select arguments.

Filter panel modules

The FilteredData object uses the filter_panel_ui and filter_panel_srv methods to put up a filter panel that can be used in any application. In teal applications it will be placed on the right-hand side. The filter panel module does not return anything. Data, subsetting calls, and filter states are accessible by specific public methods: get_data, get_call, and get_filter_state, respectively. Typically, the filter panel consists of three modules:

FilteredData does not handle data sets directly because they may be of different types, rather, it calls respective methods in lower-level classes.

When a new filter is added using the “Add Filter Variable” module in FilterStates, a new FilterState object is initialized and added to private$state_list. FilterStates$srv_active observes private$state_list (which is a reactiveVal) and when the state of the list changes (a filter is added or removed), it calls FilterState$server and uses renderUI to display FilterState$ui.