Improving Antibody Humanness Prediction using Patent Data

Authors: Talip Ucar, Aubin Ramon, Dino Oglic, Rebecca Croasdale-Wood, Tom Diethe, Pietro Sormanni

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results demonstrate that the learned model consistently outperforms the alternative baselines and establishes new state-of-the-art on five out of six inference tasks, irrespective of the used metric.
Researcher Affiliation Collaboration 1Centre for AI, Bio Pharmaceuticals R&D, Astra Zeneca 2Centre for Misfolding Diseases, Yusuf Hamied Department of Chemistry, University of Cambridge 3Biologics Engineering, Oncology R&D, Astra Zeneca.
Pseudocode No The paper describes the Self PAD framework and its training processes using explanatory text and diagrams (Figure 1), but it does not include formal pseudocode blocks or algorithms.
Open Source Code Yes The code for Self PAD is available at: https://github.com/AstraZeneca/SelfPAD
Open Datasets Yes We use patented antibody database (PAD) (Krawczyk et al., 2021)... 553 Therapeutics is a dataset from Prihoda et al. (2022)... 217 immunogenicity refers to the dataset obtained from Prihoda et al. (2022)... 25 humanization data refers to the dataset... Marks et al. (2021).
Dataset Splits Yes The training set is then split into two folds, 90% for training and 10% for validation, and the model is fine-tuned with cross-entropy loss for 25 epochs.
Hardware Specification Yes We used a compute cluster consisting of A10G GPUs throughout this work.
Software Dependencies Yes We implemented our work using Py Torch (Paszke et al., 2019).
Experiment Setup Yes We pre-trained the model with a batch size of 100 for 1000 epochs (see Figure 5 in the appendix)... we used cross-entropy loss with label smoothing (configured to 0.5) and a batch size of 512 for 25 epochs... Table 9 lists hyperparameters used for pre-training and fine-tuning stages.