IPIAbDev is a highly flexible AI-ML Framework, designed for high-throughput prediction of antibody developability and biophysical properties, with a primary focus on polyreactivity (PSR) and SEC developability, while being extensible to SPR binder, HIC, etc. It integrates multiple antibody-specific protein language models (AbLang2, AntiBERTy, AntiBERTa2, AntiBERTa2-CSSP) for embedding generation and supports a diverse set of classifiers, including XGBoost, Random Forest, 1D-CNN with residual blocks, and Transformer architectures. Key features include automated embedding generation, HCDR3-cluster-stratified k-fold cross-validation to prevent data leakage, model training and prediction for binary classification tasks, and built-in interpretability via Integrated Gradients for residue-level attribution. The package also provides publication-ready visualization of ROC curves, performance metrics (AUC, accuracy, F1, precision, recall), and attribution heatmaps

This open source software was developed at The Antibody Platform, Institut for Protein Innovation,Boston, USA

Machine Learning Architect and Designer: Hoan Nguyen, PhD

Authors and Contact: {Hoan.Nguyen, Andre.Teixeira }@proteininnovation.org

Sequence-based ML models and Encoding

Full heavy and light chain sequences, Heavy chain sequences, CDR3 sequences

Protein Lanuage model embedding, one hot encoding, sequence biophysical properties (molecular weight, charge, Isoelectric Point,..)

Trainset size vs Model performance and Generalization diagnostics

The ML data in figure (A,B) below was generated by IPIAbDev with Transformer-VH+VL+HCDR3 one hot encoding and public dataset #1 from (HT Chen, 2024)

Identifying optimal threshold for binary classification

From Prediction to Residue Interpretablity to Antobody Mutagenesis

Download and install: IPIAbDev package

git clone https://github.com/proteininnovation/IPIAbdev.git

#Create a new environment with Python 3.11 or 3.12 conda create -n ml python=3.11 -y conda activate ml

#python package install conda install -c bioconda anarci pip install -r requirements.txt

How to use IPIAbDev

prepare train set

filename: your_trainset_name.xlsx required columns: BARCODE,CDR3,HSEQ,LSEQ, any_biophysical_properties_column as sec_filter,psr_filter ,spr_filter any_biophysical_properties_column : this label should be annotated as 1 (pass/positive) or 0 (fail or negative)

Generate embeddings

python predict_developability.py --build-embedding data/test.xlsx --lm all

python predict_developability.py --build-embedding data/test.xlsx --lm ablang

Xgboost trainning evaluation with k-fold validation

python predict_developability.py --kfold 10 --target sec_filter --lm antiberta2 --model xgboost --db data/ipi_antibody.xlsx

Randomforest training evaluation with k-fold validation

python predict_developability.py --kfold 10 --target sec_filter --lm antiberty --model rf --db data/ipi_antibody.xlsx

CNN trainning evaluation with k-fold validation

python predict_developability.py --kfold 10 --target sec_filter --lm antiberta2 --model cnn --db data/ipi_antibody.xlsx

Transformer trainning evaluation with k-fold validation

python predict_developability.py --kfold 10 --target sec_filter --lm antiberty --model transformer_lm --db data/ipi_antibody.xlsx

Transformer with one hot encoding trainning evaluation with k-fold validation

python predict_developability.py --kfold 10 --target sec_filter --lm onehot --model transformer_onehot --db data/ipi_antibody.xlsx

Train final model with full dataset

#SEC trainning

python predict_developability.py --train --target sec_filter --lm antiberta2 --model xgboost --db data/ipi_antibody.xlsx

#PSR trainning

python predict_developability.py --train --target sec_filter --lm antiberta2 --model xgboost --db data/ipi_antibody.xlsx

Predict instantly

python predict_developability.py --predict data/new_lib.xlsx --target sec_filter --lm antiberta2

Predict on test set

python predict_developability.py --predict data/test.xlsx --target sec_filter --lm ablang --model cnn