---
task_categories:
- text-generation
tags:
- biology
- protein
- mutation
---

# RipplePLM Dataset

This dataset accompanies the paper [RipplePLM: Structural and Property Decoupling for Protein Mutation Effect Generation](https://huggingface.co/papers/2610.01891). It contains protein mutation data for training and evaluating models that generate natural-language descriptions of protein mutation effects.

Code and usage instructions are available at [https://github.com/Lyu6PosHao/RipplePLM](https://github.com/Lyu6PosHao/RipplePLM).

The dataset is organized into two splits (`structural_split/` and `temporal_split/`), each containing `train_with_label.csv`, `val_with_label.csv`, and `test_with_label.csv`.

The columns include:
- `entry`: Protein ID and mutation, e.g. `P12345-A123B`; residue positions are 1-based.
- `protein1`: Wild-type amino acid sequence.
- `protein2`: Mutant sequence; reconstructed from `entry` if absent.
- `function`: Wild-type protein function used as input context.
- `uniprot_description`: UniProt mutation-effect description.
- `GPT_description`: Additional generated mutation-effect description.
- `all_description`: Description field used to filter samples with missing annotations.
- `ephod_label`: Integer class label for pH.
- `pptstab_label`: Integer class label for thermostability.

The reference text combines `uniprot_description` and `GPT_description`. Samples missing `function` or `all_description` are omitted.