PIP-BERT

[] (https://colab.research.google.com/drive/1zccF8lGrF5rNQaSFPTd4wI-xvIJr-A78?usp=sharing)

ProtBERT-PI enables rapid screening of potential small secreted protease inhibitors in large-scale genomic, transcriptomic, or proteomic datasets.

The model assigns each input sequence to one of two classes:

Positive (Potential PI): Predicted to exhibit protease inhibitor activity
Negative (Non-PI): Predicted to lack protease inhibitor activity

Output includes:

Probability of the positive class (prob_class_1): ranges from 0 (low likelihood) to 1 (high likelihood of PI activity)
Confidence score: probability of the predicted class

Model Architecture and Training

ProtBERT-PI is a fine-tuned sequence classification model built on ProtBERT (BertForSequenceClassification):

Base model: Rostlab/prot_bert
Pre-trained on large corpora of protein sequences using masked language modeling

Fine-tuning was performed on a curated dataset of known protease inhibitors and non-protease inhibitor negative set. Sequences are tokenized by inserting spaces between amino acids (standard for ProtBERT), enabling effective representation learning. Maximum sequence length is configurable (default: 250 AA); longer sequences are truncated.

Positive examples: known protease inhibitors (<250 AA) from the MEROPS database
Negative examples: non-inhibitors selected from UniProt using sequence similarity and Pfam domain analysis

---

license: creativeml-openrail-m