This repository will host the source code accompanying the paper “Orphan toxin selectivity is accurately predicted by neural networks”.
The project presents an AI-driven pipeline for predicting the activity and isoform selectivity of peptide toxins targeting voltage-gated potassium channels (Kv1.1–Kv1.3). The approach combines:
- sequence-based classifiers built on fine-tuned ESM-2 protein language models;
- structure-based models leveraging AlphaFold-Multimer confidence metrics, AFToolkit, and Rosetta energy scores;
- a unified ranking strategy to prioritize orphan peptide toxins for experimental validation.
The proposed framework was used to screen hundreds of thousands of peptide sequences, leading to the discovery of several previously uncharacterized toxins with high potency and selectivity.
🔜 Coming soon:
- full source code for training and inference;
- scripts for dataset preparation and screening;
- pretrained model weights;
- instructions to reproduce the results from the paper.
Please stay tuned for updates.