B3BPFN

B3BPFN is a reproducible repository for blood-brain barrier-penetrating peptide (BBBP) prediction. The current pipeline combines ESM2 sequence embeddings, iFeatureOmega physicochemical descriptors, mutual-information-based feature selection, and a TabPFN classifier.

The repository includes:

• the curated benchmark dataset used in the current experiments
• the training script for rebuilding the model pipeline
• the inference script for scoring new peptide sequences
• the exported preprocessing objects and trained classifier

Overview

The released workflow follows four stages:

1. Encode peptide sequences with facebook/esm2_t33_650M_UR50D.
2. Extract handcrafted descriptors with iFeatureOmegaCLI.
3. Fuse features, remove constant dimensions, standardize them, and keep the top 700 features with mutual information.
4. Train a TabPFNClassifier on an undersampled training set and apply the saved threshold for final prediction.

Current benchmark scale in this repository:

• Positive peptides: 426
• Negative peptides: 6865
• Feature space before selection: 2121
• Retained features: 700

Reported performance from the current internal benchmark:

• Sensitivity: 0.9294
• Specificity: 0.8824
• Accuracy: 0.9059
• MCC: 0.8127
• AUROC: 0.9460

Repository Layout

B3BPFN/
├── data/
│   ├── all_negative.fasta
│   └── all_positive.fasta
├── models/
│   ├── kbest_selector.pkl
│   ├── optimal_threshold.txt
│   ├── standard_scaler.pkl
│   ├── tabpfn_classifier.pkl
│   └── variance_selector.pkl
├── tests/
│   └── test_predict_cli.py
├── predict_peptide.py
├── train_peptide_final.py
└── README.md

Environment

This project is intended to run on a Linux server or a compute node with a configured Python environment. At minimum, you will need:

• Python 3.10 or newer
• numpy
• pandas
• torch
• joblib
• scikit-learn
• transformers
• tabpfn
• iFeatureOmegaCLI

Notes:

• The ESM2 backbone is loaded from Hugging Face on first use.
• Training and large-batch inference are best run on a GPU-enabled server.
• iFeatureOmegaCLI installation can vary by environment; follow your server setup conventions if you already maintain a shared bioinformatics stack.

Data

The benchmark data are stored in FASTA format:

• data/all_positive.fasta: curated BBB-penetrating peptides
• data/all_negative.fasta: curated non-BBB-penetrating peptides

The current training script builds a balanced test set by holding out 20% of the positive peptides and sampling the same number of negatives.

Training

Run training from the repository root:

python3 train_peptide_final.py

The script will:

• read the FASTA files under data/
• extract ESM2 and iFeatureOmega features
• perform variance filtering, scaling, and top-700 feature selection
• undersample the negative class to a 1:5 positive-to-negative ratio
• train the TabPFN classifier
• write the resulting artifacts to models/

Inference

Prepare an input FASTA file such as:

>pep_1
RRWQWRMKKLG
>pep_2
ACDEFGHIKLMNP

Then run:

python3 predict_peptide.py -i input.fasta -o bbb_predictions.csv

Optional arguments:

• -m, --model_dir: path to the saved model directory, default is models
• -o, --output: output CSV path, default is bbb_predictions.csv

The output CSV contains:

• ID
• Sequence
• Probability
• Prediction (BBB+ or BBB-)

Reproducibility Notes

• The training script uses a fixed random seed of 42.
• The exported objects in models/ are the exact preprocessing and classification artifacts used by the inference script.
• tabpfn_classifier.pkl is tracked with Git LFS because of its size.

Verification

This repository currently includes a lightweight regression test to ensure that the default inference model path stays aligned with the published repository layout:

python3 -m unittest tests/test_predict_cli.py

Public Release Status

This repository is being prepared as a code release first. Paper metadata, citation information, and author details are intentionally omitted at this stage.