This model was trained on approximately 141 million human antibody variable heavy chain (VH) sequences for humanization tasks. It achieves an accuracy of 91.46% and can be applied to antibody sequence analysis, humanization, and the study of VH chain patterns.

Usage

python

from transformers import RobertaTokenizer, RobertaForMaskedLM

# Load tokenizer and model from Hugging Face
tokenizer = RobertaTokenizer.from_pretrained("hemantn/roberta-base-humAb-vh")
model = RobertaForMaskedLM.from_pretrained("hemantn/roberta-base-humAb-vh")

Using AnthroAb Python Package

For easier antibody humanization, you can use the AnthroAb Python package which provides a high-level interface for antibody humanization tasks. This package is available on PyPI and includes both VH and VL chain models.

Installation

bash

conda create -n anthroab python=3.10
conda activate anthroab
pip install anthroab

Quick Usage

python

import anthroab

# Humanize a heavy chain sequence (VH)
vh_sequence= "**QLV*SGVEVKKPGASVKVSCKASGYTFTNYYMYWVRQAPGQGLEWMGGINPSNGGTNFNEKFKNRVTLTTDSSTTTAYMELKSLQFDDTAVYYCARRDYRFDMGFDYWGQGTTVTVSS"
humanized_vh = anthroab.predict_masked(vh_sequence, 'H')
print(f"Humanized VH: {humanized_vh}")

# Humanize a light chain sequence (VL)
vl_sequence = "DIQMTQSPSSLSASV*DRVTITCRASQSISSYLNWYQQKPGKAPKLLIYSASTLASGVPSRFSGSGSGTDF*LTISSLQPEDFATYYCQQSYSTPRTFGQGTKVEIK"
humanized_vl = anthroab.predict_masked(vl_sequence, 'L')
print(f"Humanized VL: {humanized_vl}")

Features

Easy Installation: Install directly from PyPI with pip install anthroab
High-Level API: Simple functions for antibody humanization
Dual Chain Support: Separate models for VH and VL chains
Sequence Infilling: Fill masked positions with human-like residues
Mutation Suggestions: Get humanizing mutations for frameworks and CDRs
Embedding Generation: Create vector representations of antibody sequences

The AnthroAb package uses this RoBERTa model (hemantn/roberta-base-humAb-vh) for VH chain humanization along with a companion VL model for light chain processing.

Model Details

Architecture

Model: RoBERTa (trained from scratch)
Architecture: RobertaForMaskedLM
Model Type: Masked Language Model for antibody sequences

Specifications

Hidden Size: 768
Number of Layers: 12
Number of Attention Heads: 12
Intermediate Size: 3072
Max Position Embeddings: 192
Vocabulary Size: 25 tokens
Model Size: ~164 MB