Epigenetic Gene Regulation Model
The Multi-Attention based Deep Learning model aims to learn regulatory latent space of epigenetic/transcriptional markers such as histone modifications, methylations, and transcription factors.
- Embedding different epigenetic/transcriptional markers into a regulatory latent space using marker-specific architectures such as CNN & RNN with attention (histone modification), RNN with attention (methylation), self-attention network (transcription factor).
- Integrating each latent feature by multi-attention block to interpret relationships between multiple regulatory features.
- In experiments with 18 cell lines, the proposed model predicted the gene expression level more accurately than the state-of-the-art model.
- Reveal the cell-type-specific gene regulation mechanisms and enriched genes in terms of their functions and epigenetic regulation.
Pathway Attention Model
The pathway attention model is an explainable deep learning model for predicting cancer subtypes using gene expression data and biological pathways.
- Capturing localized gene expression patterns in pathways by graph convolutional networks (GCNs)
- Reproducing biological mechanisms by combining results of GCN with multi-attention based ensemble
- Identifying transcription factors as regulator of pathways to elucidate subtype-specific biological functions from network propagation algorithm
RKSS Kernel
RKSS kernel aims to measure similarity of DNA sequences in terms of evolutionary distance.
- Extension of the k-spectrum string kernel by utilizing two features for comparative and evolutionary sequence comparison
- Building a common k-mers template “landmark” to mimic common ancestors and reduce features
- Using rank information instead of frequency of k-mers to robust outliers
- Relatively well reconstruct phylogenetic trees of 10 mammalian species on three genomic regions (exon, intron, CpG island)
- Landmark space that is constructed using RKSS kernel effectively represents the genetic properties of the three regions: Order of three regions in terms of evolutionary information (exon > CpG island > intron)
MIDAS
MIDAS determines condition specific subpaths, each of which has different activities across multiple phenotypes.
- Utilizing explicit gene expression quantity information from RNA-seq data
- Addressing subpath mining problem on multi-class by adopting statistical approaches
- Using a greedy subpath extension method with exponentially increasing criteria to mining complex interaction of genes
