Ibex Using BCR sequences for graph embedding

Introduction

Single-cell sequencing is an integral tool in immunology and oncology, enabling researchers to measure gene expression and immune cell receptor profiling at the level of individual cells. We developed the scRepertoire R package to facilitate the integration of immune receptor and gene expression data. However, leveraging clonal indices for more complex analyses—such as using clonality in cell embedding—remains challenging.

Ibex addresses this need by using deep learning to vectorize BCR sequences based on amino acid properties or their underlying order. Ibex is the sister package to Trex, which focuses on TCR sequence data.

---

System Requirements

Ibex has been tested on R versions >= 4.0. For details on required R packages, refer to the package’s DESCRIPTION file. It is designed to work with single-cell objects containing BCR data generated using scRepertoire. Ibex has been tested on macOS and Linux.

keras3 is required for the immApex-based (deep learning) components of Ibex. Follow the steps below to install and configure keras3 and TensorFlow:

```r

Install keras3

install.packages("keras3")

Setting up Tensor Flow

library(reticulate) usecondaenv(condaenv = "r-reticulate", required = TRUE) library(keras3) installkeras() ```

Installation

Ibex relies on the immApex API can be installed directly from GitHub:

r devtools::install_github("BorchLab/immApex")

You may also install immApex from Bioconductor:

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

BiocManager::install("immApex") ```

After immApex installation, you can install Ibex with:

r devtools::install_github("BorchLab/Ibex")

The main version of Ibex is submitted to Bioconductor (installation instructions will be updated after review). By default, Ibex will automatically pull deep learning models from a Zenodo repository and cache them locally.

Alternatively, to install Ibex and all the required models at once: r devtools::install_github("BorchLab/Ibex@fullstack")

Usage/Demos

Ibex integrates smoothly into most popular R-based single-cell workflows, including Seurat and Bioconductor/SingleCellExperiment.

Quick Start

See the vignette for a step-by-step tutorial.

Autoencoded Matrix

The Ibex algorithm allows users to select BCR-based metrics to return autoencoded values to be used in dimensional reduction. If single-cell objects are not filtered for B cells with BCR, Ibex.matrix() will still return values, however IBEX_1 will be based on the disparity of BCR-containing and BCR-non-containing cells based on the Ibex algorithm.

r library(Ibex) my_ibex <- Ibex.matrix(singleObject)

Seurat or Single-Cell Experiment

You can run Ibex within your Seurat or Single-Cell Experiemt workflow. Importantly runIbex() will automatically filter single-cells that do not contain BCR information in the meta data of the single-cell object.

```r seuratObj_Bonly <- runIbex(seuratObj, #The single cell object chain = c("Heavy", "Light"), # "Heavy" or "Light" method = c("encoder", "geometric"), # Use deep learning "encoder" or "geometric" transformation encoder.model = c("CNN", "VAE", "CNN.EXP", "VAE.EXP"), # Types of Deep Learning Models encoder.input = c("atchleyFactors", "crucianiProperties", "kideraFactors", "MSWHIM", "tScales", "OHE"), # Method of Encoding geometric.theta = pi/3, # theta for Geometric Encoding species = "Human") # "Mouse" or "Human"

seuratObj_Bonly <- runIbex(seuratObj, reduction.name = "Ibex") ```

After Running Ibex

Once the Ibex embeddings are part of your Seurat object, you can use these embeddings to generate a t-SNE or UMAP:

r seuratObj <- RunTSNE(seuratObj, reduction = "Ibex", reduction.key = "Ibex_") seuratObj <- RunUMAP(seuratObj, reduction = "Ibex", reduction.key = "Ibex_")

If using Seurat package, the Ibex embedding information and gene expression PCA can be used to find the Weighted Nearest Neighbors. Before applying the WNN approach, best practice would be to remove the BCR-related genes from the list of variable genes and rerunning the PCA analysis.

Recalculate PCA without BCR genes with quietBCRgenes() function in Ibex.

r seuratObj <- quietBCRgenes(seuratObj) seuratObj <- RunPCA(seuratObj)

Running WNN approach

```r seuratObj <- FindMultiModalNeighbors(seuratObj, reduction.list = list("pca", "Ibex"), dims.list = list(1:30, 1:20), modality.weight.name = "RNA.weight")

seuratObj <- RunUMAP(seuratObj, nn.name = "weighted.nn", reduction.name = "wnn.umap", reduction.key = "wnnUMAP_") ```

Bug Reports/New Features

If you run into any issues or bugs please submit a GitHub issue with details of the issue.

• If possible please include a reproducible example. Alternatively, an example with the internal ibex_example would be extremely helpful.

Any requests for new features or enhancements can also be submitted as GitHub issues.

Pull Requests are welcome for bug fixes, new features, or enhancements.

Citation

More information on Ibex is available at our Biorxiv preprint.