Evaluating predictive patterns of antigen-specific B cells by single- cell transcriptome and antibody repertoire sequencing

This repository contains code to perform the analysis performed by L. Erlach, et al.

Abstract

The field of antibody discovery typically involves extensive experimental screening of B cells from immunized animals. Machine learning (ML)-guided prediction of antigen-specific B cells could accelerate this process, but requires sufficient training data with antigen-specificity labeling. Here, we introduce a dataset of single-cell transcriptome and antibody repertoire sequencing of B cells from immunized mice, which are labeled as antigen-specific or non-specific through experimental selections. We identify gene expression patterns associated with antigen-specificity by differential gene expression analysis and assess their antibody sequence diversity. Subsequently, we benchmark various ML models, both linear and non-linear, trained on different combinations of gene expression and antibody repertoire features. Additionally, we assess transfer learning using features from general and antibody-specific protein language models (PLMs). Our findings show that gene expression-based models outperform sequence-based models for antigen-specificity predictions, highlighting a promising avenue for computational-guided antibody discovery.

Table of contents

1. Working Environment
2. Datasets
3. Feature generation
4. Model evaluations
5. Visualization
6. Citing This Work

Working environment

Setup with Conda

conda env create -f environment_scabpred.yml
conda activate abpred

Datasets

The raw sequencing data is deposited in SRA under the BioProject number: PRJNA1124428.

1. The sequencing files were processed with cellranger (v5.0.0) and the scripts for the alignment of the files is in scSeq_preprocess/1_Cellranger_alignment_GEX.sh and scSeq_preprocess/2_Cellranger_alignment_VDJ.sh
2. Preprocessed single cell sequencing data was further processed in R, mainly utilizing the Playtpus and Seurat packages. The analysis, including the differential expression analysis is in scSeq_preprocess/
3. Preprocessing of the datasets for ML model evaluations is in notebooks/ML_preprocess/

Feature generation

Features for the ML model evaluations were generated in the Jupyter notebooks in notebooks/ML_preprocess/

1. Gene expression data was processed in 003_GEX_dataprep.ipynb
2. Antibody sequencing data was processed in 001_VDJ_OVA_seq_preprocessing.ipynb and 001.2_VDJ_RBD_seq_preprocessing.ipynb
3. The PLM embeddings were generated with the notebooks in notebooks/ESM_embed/001_Generate_ESM_embeddings.ipynb, notebooks/ESM_embed/002_Extract_CDR3Embeddings.ipynb, notebooks/Antiberty_embed/Embed_seqs_antiberty.ipynb and notebooks/ESM3_embed/Embed_seqs_ESM3.ipynb

Model evaluations

Scripts for training and evaluating the ML models are in

• src/Spec_classification/Specificity_classification_script.py
• src/GEX/Spec_classification/Specificity_classification_script.py

These scripts can be executed as shown in the example below.

./Specificity_classification_script.py --config path_to_config --simsplit_thresh 0.05 --chaintype VH_VL --outpath path_to_save_results

Visualization

Visualization and summarization of the results of the model evaluation are performed with the jupyter notebooks notebooks/model_evaluations/Metrics_visualization.ipynb and notebooks/model_interpretation/LogReg_Koefficient_analysis.ipynb. The latter also contains the biological analysis of the LogReg models.