Genomics Deep Learning

MiniSpliceNet — Deep Learning for Splice-Site Classification

A PyTorch-based genome-annotation workflow for classifying DNA sequence windows into donor splice site, acceptor splice site, and non-splice background sequence classes.

PythonPyTorchNumPyscikit-learnMatplotlibDNA Encoding1D CNNBioinformatics

Sequence Modelling Pipeline

DNA window to splice class

1DNA sequence window
2One-hot nucleotide tensor
31D convolutional feature extractor
4Softmax splice-site classifier

3

Sequence classes

donor, acceptor, background

201 bp

Window length

centered genomic context

1D CNN

Model architecture

PyTorch sequence classifier

F1

Primary metric

class-aware evaluation

Problem Framing

Splice-signal modelling for genome annotation

Splice-site recognition is central to transcript annotation. MiniSpliceNet implements this as a reproducible supervised learning workflow: construct labelled sequence windows, encode nucleotides, train a neural sequence classifier, and evaluate class-specific performance.

Sequence Windows

Fixed-length DNA windows are organized around donor, acceptor, and background sequence regions for supervised splice-signal learning.

Nucleotide Encoding

A, C, G, and T are converted into one-hot vectors so each sample becomes a model-ready sequence tensor.

CNN Classifier

A 1D convolutional network learns local splice-signal patterns from motif-centred genomic windows.

Benchmark Report

The training run reports class-balanced metrics including precision, recall, F1-score, and a confusion matrix.

Architecture

The core model uses 1D convolutions over nucleotide channels, nonlinear activations, global pooling, and a dense classifier for three-way splice-site prediction.

Evaluation

Stratified train and validation splits keep donor, acceptor, and background classes balanced while reporting precision, recall, F1-score, and confusion-matrix results.

Research Context

The workflow connects splice-signal detection with genome annotation tasks such as FASTA/GTF-derived sequence modelling and transcriptomics-informed isoform analysis.