The official repository for thermostability prediction of single fragment variable chain antibodies (scFvs) and monoclonal antibodies (mAbs) using deep learning.
The code from this work is made available here. The google colab to try this out will be out soon! The pre-trained weights will be released for commercial and non-commercial use soon.
We have trained both unsupervised pre-trained language models (PTLMs) and supervised convolutional neural networks (SCNNs) in this work. The PTLMs are in the ./ptlm directory and the SCNNs are in the ./thermD directory respectively. The antibody-specific pre-trained language model (trained with AntiBERTy) is in the ./ab-ptlm directory.
Further details on how to use each of these models is described in the subsequent README files.
For the PTLMs, the following dependencies are required:
pytorchpytorch-lightningtorchmetricsesmevo
For the SCNNs, the following dependencies are required:
pytorchDeepAbstructure prediction software available via RosettaCommons-DL (This can be updated withIgFoldfor better structural accuracy)torchmetrics
TherML is a proof-of-concept study for general pre-trained language models (UniRep, ESM), antibody-specific language models (AntiBERTy), and supervised energy-based models (thermD). The three directories have code for the three different models being evaluated in this work.
ab-ptlm: AntiBERTy model (zero-shot and fine-tuned)ptlm: General pre-trained language models (zero-shot and fine-tuned)thermD: Supervised convolutional networks trained with energies and sequence information Each of the directories has further information about respective models.
If you run into any problems while using therML, please create a Github issue with a description of the problem and the steps to reproduce it. You can also reach out to us at ameya@jhu.edu and let us know if we can improve our code/help you troubleshoot the bug/s.
Please read the terms and conditions as listed in the licensing document provided by Amgen.
author = {Harmalkar, Ameya and Rao, Roshan and Honer, Jonas and Deisting, Wibke and Anlahr, Jonas and Hoenig, Anja and Czwikla, Julia and Rau, Doris and Rice, Austin and Riley, Timothy P and Li, Danqing and Catterall, Hannah B and Tinberg, Christine E and Gray, Jeffrey J and Wei, Kathy Y and Science, Computer and Discovery, Therapeutic and Oaks, Thousand and Francisco, South San},
title = {{Towards generalizable prediction of antibody thermostability using machine learning on sequence and structure features}},
journal = {KMAB-mAbs},
doi = {10.1080/19420862.2022.2163584}
year = {2023}
}