2023-04-19

Presentation by Tatsuhiko Naito

Tatsuhiko works on decoding splicing regulatory mechanisms of Alzheimer’s disease microglia using deep learning. The aim is to construct prediction models from brain (microglia) RNA-Seq, train splicing DL model, score to predict variant effect and apply models to ADSP WGS data (n~36k) for rare variant analysis, fine mapping, etc.

Excerpts from the talk.

Microglia are a type of specialized immune cells and plays a key role in A\(\beta\) aggregation. Katia et al. Nat Genet 2022 showed that there is colocalization between AP/PD risk and microglia QTL effects.

Rare variant test using annotations. Deep learning outperforms conventional methods.

Pangolin

Similar architecture as SpliceAI, but improved the prediction accuracy by making predictions for multiple tissues simultaneously. Pangolin NN consist of 16 stacked residal blocks (dilated convolutional layer with skip connections). Takes an input block of 15,000 base pairs and outputs predictions for middle 5000 sites (nucleotides) simultaneously.

SSE (Splice-Site Strength Estimate)

\[ SSE = \frac{\alpha}{(\alpha + \beta_1 + \beta_2)} \]

Pangolin used 4 tissues from 4 different species to improve prediction. He is using microglia data (both in-house and other data) to further improve the learning. He is trying to benchmark his new model against the original Pangolin model.

• our data vs original model vs original model + our data
• accuracy improvement by increading sample size
• including other species

Expression Modifier Score (EMS): predicting gene expression altering effects of variants trained with fine-mapped eQTL variants (\(y\)) and thousands of genomic features (\(x\)), \(y \sim f(x)\).