Reproduce “PeptideBERT: A Language Model based on Transformers for Peptide Property Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chakradhar Guntuboina, Adrita Das, Parisa Mollaei, Seongwon Kim, Amir Barati Farimani
PeptideBERT is a transformer-based language model developed by researchers at Carnegie Mellon University that predicts crucial peptide properties solely from their amino acid sequences. It achieved state-of-the-art accuracy for hemolysis (86.051%) and non-fouling (88.365%) prediction, and competitive accuracy for solubility (70.018%) by fine-tuning a pre-trained protein language model.
Key insights
- Transformer-based language models effectively learn complex sequence-to-property relationships in peptides without needing explicit structural data.
- Fine-tuning pre-trained protein language models like ProtBERT transfers learned biological knowledge, leading to robust performance on diverse downstream prediction tasks.
- Data augmentation, specifically random swapping, can improve prediction accuracy for challenging peptide properties such as solubility.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.