Quadtree

The Quadtree is a gradient-boosted decision tree model used to predict guanine quadruplexes in DNA sequences. It's developed on top of the LightGBM python library. Each sequence base is encoded based on a given encoding prescription. The model was trained to be used with a sliding window and analyses the whole sequence. Machine learning model can be used as python script or thru preview website quadtree.vercel.app

Repository structure

quadtree
    └─ web -> preview website source code
    └─ python
          └─ model -> lightgbm model params
          └─ train -> example files how training was performed
          └─ quadtree.py -> predictor

Requirements

• lightgbm==3.3.2
• numpy==1.21.2

Install dependencies

Before using install the requirements:

  pip install -r requirements.txt

Usage

Create model instance

  from quadtree import Quadtree
  
  model = Quadtree()

Run analysis - algorithm inputs

• sequence as a string (maximum length is not limited)
• threshold (recommended values is 0.2)
• quadnet model file path

result = quadtree.analyse(
    sequence='ATTAATACTTTTAACAATTGTAGTATATAAAAAAGGGAGTAACC...', 
    model_path='/path/to/quadnet_model.txt',', 
    score_threshold=0.1
)

Results are then returned in given form which can be loaded into pandas DataFrame.

import pandas as pd

df = pd.DataFrame(result)

	index	position	sequence	length
0	0	907	GCAACAATGGCTGATCCAGAAGGTACAGACGGGGAGGGCACGGGTTGTAACGGCTGGTTTTATGTACAAGCTATTGTAGACAAAAAAACAGGAGATGTAATATCA	105
1	1	1184	GAGGCAGCACAGAAAACAGTCCATTAGGGGAGCGGCTGGAGGTGGATACAGAGTTAAGTCCACGGTTACAAGAAATATCTTTAAATAGTGGGCAGA	96
2	2	1389	ATGTAGTGGCGGCAGTACGGAGGCTATAGACAACGGGGGCACAGAGGGCAACAACAGCAGTGTAGACGGTACAAGTGACAATAGCAATATAGAAAATGTAAATCCAC	107
3	3	1635	AGATTGGGTTACAGCTATATTTGGAGTAAACCCAACAATAGCAGAAGGATTTAAAACACTAATACAGCCATTTAT	75
4	4	2229	AATAGATGAAGGGGGAGATTGGAGACCAATAGTGCAATTCCTGCGATACCAACAAATAGAGTTTATAACATTTTTAG	77

Model scheme

Training parameters

These parameter were used to train lightgbm model

LGBM Classifier	value
colsample bytree	0.817574864502621
learning rate	0.03744835808549148
max bin	127
min child sample	3
number of estimators	1000
number of leaves	74
regularization alpha	0.0033803043003857677
regularization lambda	0.7013136087939289
objective	binary

Authors

• Patrik Kaura - Main developer - patrikkaura

License

This project is licensed under the MIT License - see the LICENSE file for details. # quadtree