Tahoeformer

Members

• Ryan Keivanfar GitHub: rylosqualo
• Min Dai GitHub: genecell
• Xinyu Yuan GitHub: KatarinaYuan
• Qichen Huang GitHub: qhuang20

Project

Title

Tahoeformer: Interpreting Cellular Context and DNA Sequence Determinants Underlying Drug Response

Overview

Tahoeformer is a deep learning model that integrates cellular context and DNA sequence information to predict drug responses. Built upon the Enformer architecture, our model aims to understand how genome variations influence drug effects in different cellular environments.

Motivation

Precision medicine requires understanding how genetic variations affect drug responses across different cellular contexts. Tahoeformer addresses this challenge by modeling:

• Cellular context (different transcriptional factor expression patterns)
• DNA sequence variations (transcriptional factor binding site mutations)

Methods

We fine-tuned the Enformer architecture using the Tahoe-100M dataset, incorporating:

• Morgan fingerprints for drug representation
• Pseudobulked gene expression data across 8 cell lines with 27 drugs at a single dosage
• DNA sequence information centered around TSS (transcription start sites) from a curated subset of 500 genes

Results

Our model demonstrates strong performance on top 20 curated genes in predicting gene expression changes in response to drug treatments across different cellular contexts, enabling better understanding of drug-genome interactions.

Code

• Tahoeformer

HuggingFace

• Tahoeformer

Datasets

• Tahoe-100M dataset

Acknowledgements

• Enformer
• GradientShap
• Weights & Biases