Deep Neural Network

Overview

This tool treats TSR key frequencies as a feature vector and trains a feed-forward neural network to classify structures from them.

It runs in two steps: build the feature table from triplet files, then train on it.

BioTSR and this model

BioTSR can compute the keys and write a key-frequency table from Python: biotsr.analysis.feature_matrix(samples).to_header_csv() gives a sample_id column followed by one column per key, as in tutorial 08, §3. That table is not the one this model reads: key_frequency_drug.py writes different columns, including the residue_type class column the model trains on, and the two have not been tested together. To reproduce the published model, use the research scripts as described below.

Installation

git clone https://github.com/KrishnaRauniyar/Nucleotide_Analysis.git
cd Nucleotide_Analysis/dnn

Create a virtual environment and install the dependencies:

python3 -m venv tsrenv
source tsrenv/bin/activate   # macOS / Linux
tsrenv\Scripts\activate      # Windows

pip install -r requirements.txt

Step 1 — build the feature table

python key_frequency_drug.py -p triplets_directory -H yes

The result is one row per structure, one column per key:

Each cell is how often that triplet key occurs in that structure.

Step 2 — train the model

python drug_model.py -p input_csv_file

Architecture

Five hidden layers, each ReLU, narrowing 128 → 64 → 32 → 16 → 8, with L2 regularisation on all but the first. The output is a softmax layer.

Categorical columns are label-encoded and features are standardised before training. The model uses the Adam optimiser with sparse categorical cross-entropy, trains for up to 50 epochs at batch size 32 against a held-out split, and stops early when validation stops improving.

Output

Three image files: