PDeepPP: revised-study checkpoints
This repository contains task-specific checkpoints for the revised PDeepPP study, including the PDeepPP models, reproduced baselines, controls, ablations, and loss comparisons. PDeepPP is a sequence-based framework for bioactive-peptide (BP) classification and post-translational modification (PTM) site prediction.
The release contains 435 PyTorch checkpoints. These are task-specific model weights, not a single Transformers AutoModel or a ready-to-use Hugging Face pipeline. Model architecture, preprocessing and embedding requirements vary by experiment and task.
Files
All checkpoints are under revised/. The directory structure is:
revised/
βββ main_comparison/ Main PDeepPP comparison models
βββ baselines/ Reproduced baseline models
βββ controls/ ESM-based control models
βββ ablation/ Component and embedding ablations
βββ loss_comparison/ Loss-function comparison models
βββ index.csv Checkpoint metadata and experiment references
βββ checksums.txt SHA-256 checksums, relative to revised/
βββ README.md Release directory notes
Within each experiment, paths identify the model or variant, task family (BP or PTM) and task, ending in model.pt. The index identifies each file's task, seed, associated experiments, size and SHA-256 checksum. Some checkpoints are used in multiple experiments and are stored only once; consult the experiment_references column rather than expecting a separate file for every experiment.
Download and use
For example, to download the main-comparison checkpoint for the BP ACE task:
from huggingface_hub import hf_hub_download
checkpoint_path = hf_hub_download(
repo_id="fondress/PDeepPP",
filename="revised/main_comparison/pdeeppp/BP/ACE/model.pt",
)
print(checkpoint_path)
Checkpoints have different PyTorch formats: some contain a state dictionary directly and others wrap it with additional metadata. Instantiate the matching model architecture and use the task-specific input and embedding pipeline before loading the weights. The ESM model and embedding caches are not included in the checkpoints. Consult the PDeepPP code and reproduction instructions for dependencies, data preparation and model implementations. Download and load only files you trust.
For a complete local download of revised/, run sha256sum -c checksums.txt from inside that directory to verify the checkpoint files. Some checkpoints retain training settings such as maximum epochs and early-stopping patience; these values are not necessarily the actual number of epochs trained or the selected best epoch.