Dataset Configuration Format

This document describes the JSON configuration format used for protein localization datasets.

Configuration Structure

Each dataset is configured using a JSON file with the following structure:

Fields Description

Field	Description	Example Values
`dataset`	HuggingFace dataset path	`"tyang816/DeepLocMulti_ESMFold"`
`pdb_type`	Type of protein structure prediction	`"ESMFold"`, `"AlphaFold2"`
`num_labels`	Number of classification labels	`10`
`problem_type`	Type of machine learning problem	`"single_label_classification"`
`metrics`	Evaluation metric	`"accuracy"`
`monitor`	Metric to monitor during training	`"accuracy"`
`normalize`	Normalization method	`"None"`

Usage

Place your configuration files in the data/DeepLocMulti/ directory with the naming convention DeepLocMulti_[ModelType]_HF.json, where [ModelType] represents the structure prediction model used (e.g., ESMFold, AlphaFold2).

Notes

All datasets are hosted on HuggingFace
Currently supports single-label classification tasks
Accuracy is used as both the evaluation and monitoring metric
No normalization is applied by default