Introduction

Installation

To run Ibex, open R and install Ibex from GitHub:

devtools::install_github("BorchLab/Ibex")

or via Bioconductor with

if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("Ibex")

Load Libraries

suppressPackageStartupMessages({
  library(bluster)
  library(dplyr)
  library(ggplot2)
  library(Ibex)
  library(kableExtra)
  library(mumosa)
  library(patchwork)
  library(Peptides)
  library(scater)
  library(viridis)
})

The Data Set

The data used here are derived from 10x Genomics’ 2k BEAM-Ab Mouse HEL data set, consisting of splenocytes from transgenic mice engineered to recognize Hen Egg Lysozyme (HEL). These splenocytes were labeled with a small antigen panel: SARS-TRI-S, gp120, H5N1, and a negative control.

To illustrate the Ibex framework, we subset to a smaller set of 200 cells (including some dominant clones) and convert the Seurat object into a SingleCellExperiment. The resulting “ibex_example” object stores all the necessary data—RNA expression, antigen capture (BEAM) features, BCR contig annotations, and computed dimensional reductions—ready for downstream Ibex analyses. The object is saved (ibex_example.rda), along with the contig information (ibex_vdj.rda), ensuring that the integrated data set can be readily reloaded and explored in subsequent steps. More information on the processing steps are available in the inst/scripts directory of the package.

Loading the processed data

Getting Expanded Sequences

The function combineExpandedBCR() extends the functionality of combineBCR() from the scRepertoire package by first concatenating the CDR1, CDR2, and CDR3 sequences into a single expanded variable. This approach retains additional information from the BCR variable regions before calling combineBCR() to consolidate BCR sequences into clones. This will allow for use of expanded sequence models which we will detail below.

Function Parameters

The combineExpandedBCR() function supports the following parameters:

Parameter	Description	Default
input.data	List of data frames containing BCR sequencing results.	Required
samples	Character vector labeling each sample.	Required
ID	Additional sample labeling (optional).	NULL
call.related.clones	Whether to group related clones using nucleotide sequences and V genes.	TRUE
threshold	Normalized edit distance for clone clustering.	0.85
removeNA	Remove chains without values.	FALSE
removeMulti	Remove barcodes with more than two chains.	FALSE
filterMulti	Select highest-expressing light and heavy chains.	TRUE
filterNonproductive	Remove nonproductive chains if the column exists.	TRUE

combined.BCR <- combineExpandedBCR(input.data = list(ibex_vdj),
                                   samples = "Sample1",
                                   filterNonproductive = TRUE)
head(combined.BCR[[1]])[,c(1,11)]

##                      barcode
## 1 Sample1_AAACCTGTCAACGGGA-1
## 2 Sample1_AAACGGGAGACAGGCT-1
## 3 Sample1_AAAGATGAGTCCGGTC-1
## 4 Sample1_AAAGATGGTAGAGGAA-1
## 5 Sample1_AAAGATGGTATCACCA-1
## 6 Sample1_AAAGATGGTATGGTTC-1
##                                                      CTaa
## 1 GDSITSD-SYSGS-CANWDGDYW_RASQSIGNNLH-YASQSIS-CQQSNSWPYTF
## 2 GDSITSD-SYSGS-CANWDGDYW_RASQSIGNNLH-YASQSIS-CQQSNSWPYTF
## 3 GDSITSD-SYSGS-CANWDGDYW_RASQSIGNNLH-YASQSIS-CQQSNSWPYTF
## 4                              GDSITSD-SYSGS-CANWDGDYW_NA
## 5 GDSITSD-SYSGS-CANWDGDYW_RASQSIGNNLH-YASQSIS-CQQSNSWPYTF
## 6 GDSITSD-SYSGS-CANWDGDYW_RASQSIGNNLH-YASQSIS-CQQSNSWPYTF

We can attach the expanded sequences to the Seurat or Single-Cell Experiment objects using the scRepertoire combineExpression() function.

Available Models

Ibex offers a diverse set of models built on various architectures and encoding methods. Currently, models are available for both heavy and light chain sequences in humans, as well as heavy chain models for mice. Models for CDR3-based sequences have been trained on sequences of 45 residues or fewer, while models for CDR1/2/3-based sequences are specific to sequences of 90 amino acids or fewer.

A full list of available models is provided below:

model.meta.data <-  read.csv(system.file("extdata", "metadata.csv", 
                                               package = "Ibex"))[,c(1:2,8)]
model.meta.data %>%
  kable("html", escape = FALSE) %>%
  kable_styling(full_width = FALSE) %>%
  scroll_box(width = "100%", height = "400px")

All the models are available via a Zenodo repository, which Ibex will pull automatically and cache for future use locally. There is no need to download the models independent of the runIbex() or Ibex_matrix() calls.

Choosing Between CNN and VAE

Convolutional Neural Networks (CNNs)

• Pros: Detect local sequence motifs effectively; relatively straightforward and quick to train.
• Cons: Can struggle to capture global context

Variational Autoencoders (VAEs)

• Pros: Model sequences within a probabilistic, continuous latent space; suitable for generating novel variants.
• Cons: Training can be more complex (balancing reconstruction and regularization losses); interpretability may be less direct.

Which to choose?

• Use CNNs if local motif detection and simpler training are priorities.
• Use VAEs if you want a generative model capturing broader sequence structures.

Choosing Encoding Methods

One-Hot Encoding: Represents each amino acid as a binary vector (e.g., a 20-length vector for the 20 standard residues).

• Pros: Simple and assumption-free.
• Cons: High-dimensional and doesn’t capture biochemical similarities.

Atchley Factors: Uses five numerical descriptors summarizing key physicochemical properties.

• Pros: Compact and embeds biochemical information.
• Cons: May overlook some residue-specific nuances.

Cruciani Properties: Encodes amino acids via descriptors that reflect molecular shape, hydrophobicity, and electronic features.

• Pros: Captures rich chemical details.
• Cons: More complex to compute and less standardized.

Kidera Factors: Provides ten orthogonal values derived from a broad set of physical and chemical properties.

• Pros: Offers a balanced, low-dimensional representation.
• Cons: Derived statistically, potentially averaging out finer details.

MSWHIM: Derives descriptors from 3D structural data, summarizing overall shape and surface properties.

• Pros: Provides robust, rotation-invariant structural insight.
• Cons: Requires 3D information and can be computationally intensive.

tScales: Encodes amino acids based on topological and structural features reflective of protein folding and interactions.

• Pros: Captures contextual information from the overall sequence structure.
• Cons: Less commonly used, making standardization and tool support a challenge.

Running Ibex

The idea behind Ibex is to combine BCR CDR3 amino acid information with phenotypic RNA/protein data to direct the use of single-cell sequencing towards antigen-specific discoveries. This is a growing field - specifically TESSA uses amino acid characteristics and autoencoder as a means to get a dimensional reduction. Another option is CoNGA, which produces an embedding using BCR and RNA. Ibex was designed to make a customizable approach to this combined approach using R.

Ibex_matrix Function

Ibex includes two primary functions: Ibex_matrix() and runIbex(). The Ibex_matrix() function serves as the backbone of the algorithm, returning encoded values based on user-selected parameters. In contrast to runIbex(), which filters input to include only B cells with attached BCR data, Ibex_matrix() operates on all provided data. Additionally, it is compatible with the list output from the combineBCR() function (from the scRepertoire package), whereas runIbex() is designed for use with a single-cell object.

Parameters

• chain:
  Specifies the chain type. Options:

  • "Heavy" for Ig Heavy Chain
  • "Light" for Ig Light Chain

• method:
  Chooses the transformation method. Options:

  • "encoder": Applies a CNN/VAE-based transformation.
  • "geometric": Uses a geometric transformation.

• encoder.model:
  When using the "encoder" method, selects the specific model variant. Options:

  • "CNN": CDR3 Convolutional Neural Network-based autoencoder
  • "VAE": CDR3 Variational Autoencoder
  • "CNN.EXP": CDR1/2/3 CNN
  • "VAE.EXP": CDR1/2/3 VAE

• encoder.input:
  Specifies the encoding input method. Options:

  • "atchleyFactors"
  • "crucianiProperties"
  • "kideraFactors"
  • "MSWHIM"
  • "tScales"
  • "OHE"

• theta:
  For the geometric transformation, defines the value of theta (default is π/3).

Ibex_vectors <- Ibex_matrix(ibex_example, 
                            chain = "Heavy",
                            method = "encoder",
                            encoder.model = "VAE", 
                            encoder.input = "OHE", 
                            species = "Mouse",
                            verbose = FALSE)

## Using Python: /home/biocbuild/.pyenv/versions/3.9.24/bin/python3.9
## Creating virtual environment '/var/cache/basilisk/1.21.5/Ibex/0.99.32/IbexEnv' ...

## Done!
## Installing packages: pip, wheel, setuptools

## Installing packages: 'keras==3.6.*', 'tensorflow==2.18.*', 'h5py==3.13', 'numpy==1.26'

## Virtual environment '/var/cache/basilisk/1.21.5/Ibex/0.99.32/IbexEnv' successfully created.
## 
1/7 ━━━━━━━━━━━━━━━━━━━━ 0s 133ms/step
7/7 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step

ggplot(data = as.data.frame(Ibex_vectors), aes(Ibex_1, Ibex_2)) + 
  geom_point(color = "grey", alpha = 0.7, size = 2) + 
  theme_classic()

Ibex_vectors2 <- Ibex_matrix(ibex_example, 
                             chain = "Heavy",
                             method = "geometric",
                             geometric.theta = pi, 
                             verbose = FALSE)

ggplot(as.data.frame(Ibex_vectors2), aes(x = Ibex_1, y = Ibex_2)) + 
  geom_point(color = "grey", alpha = 0.7, size = 2) + 
  theme_classic()

runIbex

Additionally, runIbex() can be used to append the Seurat or Single-cell Experiment object with the Ibex vectors and allow for further analysis. Importantly, runIbex() will remove single cells that do not have recovered BCR data in the metadata of the object.

ibex_example <- runIbex(ibex_example, 
                        chain = "Heavy",
                        encoder.input = "kideraFactors", 
                        reduction.name = "Ibex.KF", 
                        species = "Mouse",
                        verbose = FALSE)

## 
1/7 ━━━━━━━━━━━━━━━━━━━━ 0s 58ms/step
7/7 ━━━━━━━━━━━━━━━━━━━━ 0s 7ms/step

Using Ibex Vectors

After runIbex() we have the encoded values stored under “Ibex…”. Using the Ibex dimensions, we can calculate a UMAP based solely on the embedded heavy chain values. Here we will visualize both the Heavy/Light Chain amino acid sequence (via CTaa) and normalized counts associated with the Anti-Hen-Egg-Lysozyme antigen.

set.seed(123)
#Generating UMAP from Ibex Neighbors
ibex_example <- runUMAP(ibex_example, 
                        dimred = "Ibex.KF",
                        name = "ibexUMAP")
#Ibex UMAP
plot1 <- plotUMAP(ibex_example, color_by ="Anti-Hen-Egg-Lysozyme", dimred = "ibexUMAP") + 
            theme(legend.position = "bottom")
plot2 <- plotUMAP(ibex_example, color_by = "CTaa", dimred = "ibexUMAP") + 
  scale_color_viridis(discrete = TRUE, option = "B") + 
  guides(color = "none")

plot1 + plot2

In this workflow, we can combine these three dimension reductions into a single, integrated UMAP embedding using the runMultiUMAP() function with a cosine metric. To further refine this integration, we apply rescaleByNeighbors() to align the nearest neighbors across modalities, followed by clustering with clusterRows(), resulting in a “combined.clustering” that reflects all data types. Finally, we visualize this joint embedding as “MultiUMAP,” coloring points by expression of a specific protein marker (e.g., Anti-Hen-Egg-Lysozyme), the integrated cluster assignments, or other relevant annotations. The result is a holistic representation of cellular diversity that leverages shared and unique signals from RNA, protein, and Ibex IGH latent features.

#Multimodal UMAP
ibex_example <- mumosa::runMultiUMAP(ibex_example, 
                                     dimreds=c("pca", "apca", "Ibex.KF"))
#Multimodal Clustering
output <- rescaleByNeighbors(ibex_example, 
                             dimreds=c("pca", "apca", "Ibex.KF"))
ibex_example$combined.clustering <- clusterRows(output, NNGraphParam())

plot3 <- plotUMAP(ibex_example, 
                  dimred = "MultiUMAP", 
                  color_by = "Anti-Hen-Egg-Lysozyme") + 
            theme(legend.position = "bottom")
plot4 <- plotUMAP(ibex_example, 
                  dimred = "MultiUMAP", 
                  color_by = "combined.clustering") + 
            theme(legend.position = "bottom")
plot5 <- plotUMAP(ibex_example, 
                  dimred = "MultiUMAP", 
                  color_by = "CTaa") + 
  scale_color_manual(values = viridis_pal(option = "B")(length(unique(ibex_example$CTaa)))) +
  guides(color = "none")

plot3 + plot4 + plot5

Comparing the outcome to just one modality

We can also look at the differences in the UMAP generated from RNA, ADT, or Ibex as individual components. Remember, the clusters that we are displaying in UMAP are based on clusters defined by the weighted nearest neighbors calculated above.

ibex_example <- runUMAP(ibex_example, 
                        dimred = 'pca', 
                        name = "pcaUMAP")

ibex_example <- runUMAP(ibex_example, 
                        dimred = 'apca', 
                        name = "beamUMAP")

plot6 <- plotUMAP(ibex_example, 
                  dimred = "pcaUMAP", 
                  color_by = "combined.clustering") 
plot7 <- plotUMAP(ibex_example, 
                  dimred = "beamUMAP", 
                  color_by = "combined.clustering") 
plot8 <- plotUMAP(ibex_example, 
                  dimred = "ibexUMAP", 
                  color_by = "combined.clustering") 

plot6 + plot7 + plot8 + plot_layout(guides = "collect") &  
  theme(legend.position = "bottom")

CoNGA Reduction

Single-cell B-cell receptor (BCR) sequencing enables the identification of clonotypes, which are groups of B cells sharing the same BCR sequence. Often, you want to link clonotypes to their gene expression profiles.

A challenge arises, however, when a clonotype contains multiple cells (e.g., 10 cells sharing the same BCR). Including all cells for every clonotype can lead to over-representation of highly expanded clones or complicate analyses that require a one-to-one mapping between clonotypes and “cells.” Recent work Schattgen,2021 has proposed different strategies to summarize or represent a clonotype by a single expression profile. Two key strategies are common:

Distance Approach

• First, look at the PCA or count matrices
• Identify the cell that has the minimum summed Euclidean distance to all other cells in the clonotype.
• This approach can help ensure that your single representation is an actual cell, rather than a potentially non-biological average.

Mean Approach

• Simply take the average (mean) expression across all cells in the same clonotype.
• Conceptually, you collapse a multi-cell clone into one “virtual cell” representing its average expression.

CoNGA.sce <- CoNGAfy(ibex_example, 
                     method = "mean", 
                     assay = c("RNA", "BEAM"))

CoNGA.sce <- runIbex(CoNGA.sce, 
                     encoder.input = "kideraFactors", 
                     encoder.model = "VAE",
                     reduction.name = "Ibex.KF", 
                     species = "Mouse",
                     verbose = FALSE)

## 
1/2 ━━━━━━━━━━━━━━━━━━━━ 0s 58ms/step
2/2 ━━━━━━━━━━━━━━━━━━━━ 0s 33ms/step

CoNGA.sce <- CoNGA.sce %>%
                  runUMAP(dimred = "Ibex.KF", 
                          name = "ibexUMAP")

plot9 <- plotUMAP(CoNGA.sce, 
                  dimred = "ibexUMAP", 
                  color_by = "Anti-Hen-Egg-Lysozyme", 
                  by.assay.type = "counts") 

plot10 <- plotUMAP(CoNGA.sce, 
                   dimred = "ibexUMAP", 
                   color_by = "H5N1", 
                   by.assay.type = "counts") 

plot9 + plot10 & 
  theme(legend.position = "bottom")

Conclusion

This has been a general overview of the capabilities of Ibex for incorporating BCR information into the embedding space of single-cell data. If you have any questions, comments, or suggestions, feel free to visit the GitHub repository.

Session Info

sessionInfo()

## R version 4.5.1 Patched (2025-08-23 r88802)
## Platform: x86_64-pc-linux-gnu
## Running under: Ubuntu 24.04.3 LTS
## 
## Matrix products: default
## BLAS:   /home/biocbuild/bbs-3.22-bioc/R/lib/libRblas.so 
## LAPACK: /usr/lib/x86_64-linux-gnu/lapack/liblapack.so.3.12.0  LAPACK version 3.12.0
## 
## locale:
##  [1] LC_CTYPE=en_US.UTF-8       LC_NUMERIC=C              
##  [3] LC_TIME=en_GB              LC_COLLATE=C              
##  [5] LC_MONETARY=en_US.UTF-8    LC_MESSAGES=en_US.UTF-8   
##  [7] LC_PAPER=en_US.UTF-8       LC_NAME=C                 
##  [9] LC_ADDRESS=C               LC_TELEPHONE=C            
## [11] LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C       
## 
## time zone: America/New_York
## tzcode source: system (glibc)
## 
## attached base packages:
## [1] stats4    stats     graphics  grDevices utils     datasets  methods  
## [8] base     
## 
## other attached packages:
##  [1] viridis_0.6.5               viridisLite_0.4.2          
##  [3] scater_1.37.0               scuttle_1.19.0             
##  [5] Peptides_2.4.6              patchwork_1.3.2            
##  [7] mumosa_1.17.0               SingleCellExperiment_1.31.1
##  [9] SummarizedExperiment_1.39.2 Biobase_2.69.1             
## [11] GenomicRanges_1.61.5        Seqinfo_0.99.2             
## [13] IRanges_2.43.5              S4Vectors_0.47.4           
## [15] BiocGenerics_0.55.3         generics_0.1.4             
## [17] MatrixGenerics_1.21.0       matrixStats_1.5.0          
## [19] kableExtra_1.4.0            Ibex_0.99.32               
## [21] ggplot2_4.0.0               dplyr_1.1.4                
## [23] bluster_1.19.0              BiocStyle_2.37.1           
## 
## loaded via a namespace (and not attached):
##   [1] splines_4.5.1             later_1.4.4              
##   [3] batchelor_1.25.0          filelock_1.0.3           
##   [5] tibble_3.3.0              polyclip_1.10-7          
##   [7] basilisk.utils_1.21.2     lifecycle_1.0.4          
##   [9] edgeR_4.7.6               globals_0.18.0           
##  [11] processx_3.8.6            lattice_0.22-7           
##  [13] MASS_7.3-65               magrittr_2.0.4           
##  [15] limma_3.65.5              rmarkdown_2.30           
##  [17] yaml_2.3.10               metapod_1.17.0           
##  [19] spam_2.11-1               sp_2.2-0                 
##  [21] reticulate_1.43.0         cowplot_1.2.0            
##  [23] chromote_0.5.1            RColorBrewer_1.1-3       
##  [25] ResidualMatrix_1.19.0     abind_1.4-8              
##  [27] rvest_1.0.5               purrr_1.1.0              
##  [29] ggraph_2.2.2              hash_2.2.6.3             
##  [31] rappdirs_0.3.3            tweenr_2.0.3             
##  [33] evmix_2.12                ggrepel_0.9.6            
##  [35] irlba_2.3.5.1             listenv_0.9.1            
##  [37] iNEXT_3.0.2               MatrixModels_0.5-4       
##  [39] scRepertoire_2.5.6        dqrng_0.4.1              
##  [41] parallelly_1.45.1         svglite_2.2.1            
##  [43] DelayedMatrixStats_1.31.0 codetools_0.2-20         
##  [45] DelayedArray_0.35.3       xml2_1.4.0               
##  [47] ggforce_0.5.0             tidyselect_1.2.1         
##  [49] farver_2.1.2              ScaledMatrix_1.17.0      
##  [51] base64enc_0.1-3           jsonlite_2.0.0           
##  [53] BiocNeighbors_2.3.1       tidygraph_1.3.1          
##  [55] progressr_0.17.0          ggalluvial_0.12.5        
##  [57] survival_3.8-3            systemfonts_1.3.1        
##  [59] tools_4.5.1               Rcpp_1.1.0               
##  [61] glue_1.8.0                gridExtra_2.3            
##  [63] tfruns_1.5.4              SparseArray_1.9.1        
##  [65] xfun_0.53                 websocket_1.4.4          
##  [67] withr_3.0.2               BiocManager_1.30.26      
##  [69] fastmap_1.2.0             basilisk_1.21.5          
##  [71] SparseM_1.84-2            digest_0.6.37            
##  [73] rsvd_1.0.5                R6_2.6.1                 
##  [75] textshaping_1.0.4         dichromat_2.0-0.1        
##  [77] tidyr_1.3.1               FNN_1.1.4.1              
##  [79] graphlayouts_1.2.2        httr_1.4.7               
##  [81] S4Arrays_1.9.1            whisker_0.4.1            
##  [83] uwot_0.2.3                pkgconfig_2.0.3          
##  [85] gtable_0.3.6              tensorflow_2.20.0        
##  [87] S7_0.2.0                  XVector_0.49.1           
##  [89] htmltools_0.5.8.1         dotCall64_1.2            
##  [91] SeuratObject_5.2.0        scales_1.4.0             
##  [93] png_0.1-8                 scran_1.37.0             
##  [95] ggdendro_0.2.0            knitr_1.50               
##  [97] rstudioapi_0.17.1         reshape2_1.4.4           
##  [99] rjson_0.2.23              cachem_1.1.0             
## [101] stringr_1.5.2             vipor_0.4.7              
## [103] parallel_4.5.1            pillar_1.11.1            
## [105] grid_4.5.1                vctrs_0.6.5              
## [107] promises_1.3.3            BiocSingular_1.25.0      
## [109] beachmat_2.25.5           cluster_2.1.8.1          
## [111] beeswarm_0.4.0            evaluate_1.0.5           
## [113] cli_3.6.5                 locfit_1.5-9.12          
## [115] compiler_4.5.1            rlang_1.1.6              
## [117] crayon_1.5.3              future.apply_1.20.0      
## [119] labeling_0.4.3            ps_1.9.1                 
## [121] immApex_1.3.7             ggbeeswarm_0.7.2         
## [123] plyr_1.8.9                stringi_1.8.7            
## [125] BiocParallel_1.43.4       gsl_2.1-8                
## [127] quantreg_6.1              Matrix_1.7-4             
## [129] dir.expiry_1.17.0         sparseMatrixStats_1.21.0 
## [131] future_1.67.0             statmod_1.5.1            
## [133] igraph_2.2.0              memoise_2.0.1