ANTICP3 — Anticancer Protein Prediction

ANTICP3 is a LLM-based tool for binary classification of proteins into Anticancer or Non-Anticancer classes, based solely on their primary amino acid sequences. It leverages the powerful ESM2-t33 transformer model, fine-tuned specifically for anticancer protein prediction.

Developed by Prof. G. P. S. Raghava's Lab, IIIT-Delhi
📄 Please cite: ANTICP3

---

Zenodo

https://doi.org/10.5281/zenodo.19911136

Dataset folder

Conatain all the datasets

Features

• Fine-tuned ESM2 model for accurate prediction.
• Accepts input in FASTA format.
• Outputs CSV with predicted labels and probabilities.
• Supports CPU and CUDA for faster inference.
• Easy to integrate into pipelines and large-scale datasets.

---

Model Details

• Base Model: facebook/esm2_t33_650M_UR50D
• Fine-Tuned On: Anticancer protein dataset
• Classification Type: Binary (Anticancer / Non-Anticancer)
• Output Format: CSV with prediction scores and labels

---

Command-Line Arguments

Parameter	Accepted Values	Description
-i, --input	Path to .fasta file	(Required) Input file containing protein sequences in FASTA format.
-o, --output	Any valid filename (e.g. results.csv)	Output CSV file to save predictions. Default is output.csv.
-t, --threshold	Float (0 to 1)	Classification threshold for deciding Anticancer vs Non-Anticancer. Default is 0.5.
-m, --model	1 Finetuned ESM2 + BLAST or 2 Finetuned ESM2	Classification model for predicting Anticancer vs Non-Anticancer. Default is 1.
-d, --device	cpu or cuda	Device to run inference on. Defaults to cpu. If cuda is specified and available, inference runs on GPU.

Usage - Standalone

Download the standalone version and set up the environment. Download ANTICP3 Standalone Package

Option 1: Using Conda

Recommended if you're using a Conda environment.

conda env create -f environment.yml
conda activate anticp3

Option 2: Using Pip

pip install -r requirements.txt

HuggingFace

🤗 Inference via Hugging Face

You can also run predictions using the fine-tuned model directly from Hugging Face Hub:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load tokenizer and fine-tuned model
tokenizer = AutoTokenizer.from_pretrained("raghavagps-group/anticp3")
model = AutoModelForSequenceClassification.from_pretrained("raghavagps-group/anticp3")

# Example protein sequence
sequence = "MANCVVGYIGERCQYRDLKWWELRGGGGSGGGGSAPAFSVSPASGLSDGQSVSVSVSGAAAGETYYIAQCAPVGGQDACNPATATSFTTDASGAASFSFVVRKSYTGSTPEGTPVGSVDCATAACNLGAGNSGLDLGHVALTFGGGGGSGGGGSDHYNCVSSGGQCLYSACPIFTKIQGTCYRGKAKCCKLEHHHHHH"

# Tokenize and predict
inputs = tokenizer(sequence, return_tensors="pt", truncation=True)

with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.nn.functional.softmax(logits, dim=-1)
    prediction = torch.argmax(probs, dim=1).item()

labels = {0: "Non-Anticancer", 1: "Anticancer"}
print("Prediction:", labels[prediction])