DeepMir

DeepMir classifies microRNAs into characterized families such as those in Rfam and miRBase using convolutional neural networks to support functional annotation.

Key Features:

• Deep Learning-Based Classification: Convolutional neural networks (CNNs) are used to automatically learn features for miRNA classification.
• Secondary Structure Encoding: Predicted secondary structures are encoded using a matrix-based approach to capture structural conservation of noncoding RNAs.
• Primary Sequence Utilization: Primary sequences are represented as one-hot encoding matrices to provide sequence-level information.
• Softmax Output for Sample Discrimination: A softmax output layer is used to distinguish in-distribution miRNAs from out-of-distribution non-miRNAs for transcriptomic datasets.
• Threshold-Based Exclusion: A threshold derived from the softmax output is applied to exclude sequences not belonging to targeted miRNA families.
• Benchmarking Against Infernal: Performance is compared with the ncRNA classification tool Infernal, showing comparable sensitivity and accuracy with significantly faster processing times.

Scientific Applications:

• MiRNA family classification and functional annotation: Assigns newly discovered miRNAs to characterized families (Rfam, miRBase) to support functional inference.
• Transcriptomic dataset screening: Discriminates miRNAs from non-miRNAs in transcriptomic datasets using softmax-based out-of-distribution detection and thresholding.
• Large-scale genomic data analysis: Faster processing times enable application to large-scale or high-throughput genomic datasets and to both small and large miRNA family sizes.

Methodology:

Convolutional neural networks trained on different feature learning and encoding methods using one-hot encoded primary sequences and matrix-encoded predicted secondary structures as inputs; a softmax output layer with a derived threshold for out-of-distribution exclusion; evaluated by comparison to Infernal reporting sensitivity, accuracy, and processing time.