AntiCapt: A method for predicting, designing and scanning anticancer single-stranded (ss) DNA aptamers

Introduction

AntiCapt is developed for predicting, designing, and scanning anticancer (AC) single-stranded (ssDNA) aptamers. Two models are incorporated -- both use the SAME feature pipeline (fine-tuned HyenaDNA embeddings), with different classifiers trained on different datasets:

• Model 1 (main / default): Anticancer aptamer v/s random oligonucleotides -- fine-tuned HyenaDNA embeddings + LightGBM
• Model 2 (alternate): Anticancer aptamer v/s general aptamers -- fine-tuned HyenaDNA embeddings + LightGBM (separate training data, scaler, and classifier)

Modules/Jobs: This program implements three modules (job types):

1. Predict -- predicting anticancer-aptamer potential of input DNA sequences.
2. Design -- generating all single-point mutants (every position x every alternative base) of input sequences and computing the anticancer-aptamer potential (score) of each mutant. Useful for identifying which single substitutions most improve predicted potency.
3. Scan -- creating all overlapping windows of a given length from longer input sequences and computing the anticancer-aptamer potential of each window. Useful for localizing the active region of a longer sequence.

Installation

Option 1: Install via PyPI (recommended)

pip install anticapt

This installs the anticapt command directly, with both models (fine-tuned HyenaDNA checkpoints, LightGBM classifiers, and scalers) bundled in -- nothing further to set up.

Option 2: Install from source

git clone https://github.com/raghavagps/anticapt.git
cd anticapt
pip install -e .

Both models (Dataset1 and Dataset2 -- fine-tuned HyenaDNA checkpoints, LightGBM classifiers, and scalers) are already included under anticapt/models/ and ready to use immediately -- no additional setup needed. The layout is:

anticapt/models/
├── dataset1/
│   ├── finetuned_hyenadna/    <- fine-tuned HyenaDNA checkpoint
│   │                             (config.json, weights, tokenizer files)
│   ├── LGBM_model.joblib
│   └── LGBM_scaler.joblib
└── dataset2/
    ├── finetuned_hyenadna/    <- fine-tuned HyenaDNA checkpoint
    ├── LGBM_model.joblib
    └── LGBM_scaler.joblib

If you later want to update either model, see models/README.txt for the layout requirements. If both datasets should share one fine-tuned embedding model, copy/symlink it into both finetuned_hyenadna/ folders.

MAX_LENGTH in anticapt/hyenadna_features.py is set to 128, matching the original fine-tuning pipeline -- change it there if yours differs.

Minimum Usage

anticapt -i example/aptamer.fasta

This predicts the anticancer-aptamer potential of sequences in FASTA format, using Model 1 by default, threshold 0.5. Output saved to anticapt_output.csv.

Full Usage

anticapt [-h] -i INPUT [-o OUTPUT] [-j {1,2,3}] [-m {1,2}]
         [-t THRESHOLD] [-w WINLENG] [-d {1,2}]

optional arguments:
  -h, --help            show this help message and exit
  -i, --input           Input: ssDNA aptamer sequence(s) in FASTA format,
                         or one sequence per line (simple format)
  -o, --output          Output: file for saving results, default anticapt_output.csv
  -j, --job {1,2,3}     Job Type: 1:predict, 2:design, 3:scan, default 1
  -m, --model {1,2}     Model: 1:Anticancer aptamer v/s random
                         oligonucleotides (default), 2:Anticancer aptamer
                         v/s general aptamers; both use fine-tuned
                         HyenaDNA embeddings + LGBM
  -t, --threshold       Threshold: value between 0 and 1, default 0.5
  -w, --winleng         Window Length: scan mode only, default 30
  -d, --display {1,2}   Display: 1:AC-aptamers only, 2:all sequences,
                         default 2

Examples

Predict, using Dataset2's model, only showing positives:

anticapt -i example/aptamer.fasta -m 2 -d 1 -o predict_results.csv

Design mutants of your sequences and rank by predicted score:

anticapt -i example/aptamer.fasta -j 2 -o design_results.csv

Scan a long sequence with a 25 nt window:

anticapt -i long_sequence.fasta -j 3 -w 25 -o scan_results.csv

Input File

Two formats accepted:

1. FASTA format (standard) -- >header line followed by sequence.
2. Simple format -- one sequence per line, no headers; IDs are auto-generated as Seq_1, Seq_2, ...

Non-ACGT characters trigger a warning but are not stripped automatically (check your sequences if you see this warning -- it usually means stray whitespace, ambiguity codes, or an accidental protein/RNA sequence).

Output File

Results are saved in CSV format. Column sets differ slightly by job type:

• Predict: Sequence_ID, Sequence, Length, Score, Prediction
• Design: Sequence_ID, Position, Original_Base, Mutant_Base, Mutant_Sequence, Score, Prediction
• Scan: Sequence_ID, Start_Position, End_Position, Window_Sequence, Score, Prediction

Score is the predicted probability (0 to 1) of the anticancer-aptamer class; Prediction is the thresholded call at --threshold.

Package Files

setup.py                              : Package metadata / build config
README.md                             : This file provides information about this package
LICENSE.txt                           : GNU GPL v3.0 license text
CITATION.cff                          : Citation information
src/anticapt/
    __init__.py                        : Package entry point, exposes core functions
    python_scripts/
        anticapt.py                    : Main CLI program (predict/design/scan)
        seq_io.py                      : FASTA/simple-format reading, output writing
        design_scan.py                 : Point-mutant and sliding-window generation
        hyenadna_features.py           : Fine-tuned HyenaDNA embedding extraction (both models)
    models/
        README.txt                     : Model directory layout
        dataset1/
            finetuned_hyenadna/          : Model 1's fine-tuned HyenaDNA checkpoint
            LGBM_model.joblib            : Model 1 classifier
            LGBM_scaler.joblib           : Model 1 scaler
        dataset2/
            finetuned_hyenadna/          : Model 2's fine-tuned HyenaDNA checkpoint
            LGBM_model.joblib            : Model 2 classifier
            LGBM_scaler.joblib           : Model 2 scaler
    example/
        aptamer.fasta                   : Test example file containing ssDNA aptamer sequences in FASTA format

License

This project is licensed under the GNU General Public License v3.0 -- see the LICENSE file for details.

Address for Contact

In case of any query please contact:

Prof. G. P. S. Raghava, Head Department of Computational Biology,
Indraprastha Institute of Information Technology (IIIT),
Okhla Phase III, New Delhi 110020; Phone: +91-11-26907444;
Email: raghava@iiitd.ac.in  Web: http://webs.iiitd.edu.in/raghava/