DAWO is a specialized Variational Autoencoder (VAE) designed to predict drug responses in cancer cell lines by integrating gene expression data with drug and cell line features. The model leverages multi-modal representation learning to capture complex interactions between drugs and cells, enabling more accurate prediction of drug responses across diverse conditions.
Motivation
Understanding and predicting how cancer cells respond to different therapeutic compounds is crucial for advancing precision medicine approaches in oncology. Traditional methods often fail to capture the complex relationships between drugs, cell lines, and their molecular profiles. DAWO addresses this challenge by combining a VAE architecture with drug-aware and cell-line-aware components to model these interactions effectively.
Methods
DAWO incorporates a multi-modal architecture with the following key components:
Gene Expression Encoder: Processes normalized gene expression data from cancer cell lines (input dimension: 5000)
Drug Feature Encoder: Processes drug features combining:
Cell Line Feature Encoder: Processes cell line features focusing on driver gene mutations and other genomic characteristics (input dimension: 113)
Latent Space: A 50-dimensional latent representation combining drug, cell line, and gene expression information
Decoder: Reconstructs gene expression profiles from the latent representation
Classifier: Predicts drug response categories from the latent representation (379 classes)
The model was trained using a combined loss function that balances reconstruction accuracy, latent space regularization, and classification performance.
Results
DAWO demonstrates strong performance in predicting drug responses across multiple cancer cell lines, with particular strength in:
Distinguishing between responsive and non-responsive cell lines for specific drugs
Generalizing to new drug-cell line combinations not seen during training
Capturing meaningful biological signals in the latent space that reflect known drug mechanisms and cellular pathways
Discussion
Our model provides a powerful framework for drug response prediction that could accelerate drug discovery and repurposing efforts. The integration of multi-modal data (gene expression, drug features, cell line characteristics) enables DAWO to capture complex interaction patterns that simpler models miss.
Limitations include the need for comprehensive feature sets for new drugs and cell lines, and potential biases from the training data distribution. Future work will focus on incorporating additional molecular modalities and expanding the training data to improve generalization across diverse drug classes.
Model Description
Using a variational autoencoder (VAE) approach, DAWO learns latent representations of these data sources and combines them to predict drug responses and identify potential drug-cell line interactions.
This model was developed using the Tahoe-100M dataset as part of the Tahoe-DeepDive Hackathon 2025.
License
MIT License
Copyright (c) 2023 Team DAWO
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
