TUGDA: Task uncertainty guided domain adaptation for robust generalization of cancer drug response prediction from in in vitro to in vivo settings
TUGDA is a novel multi-task unsupervised domain adaptation method that leverages transfer learning from tasks/domains in a unified framework by quantifying uncertainty in predictors and weighting their influence on shared domain/task feature representations. TUGDA's ability to rely more on predictors with low-uncertainty allowed it to notably reduce cases of negative transfer and sucessfully transfer knowledge across biological models.
Model Representation:
TUGDA framework for multi-task learning and domain adaptation in cancer drug response prediction. The layer L receives input data from different biological models and maps them to a latent space Z. Then, the multi-task layer S uses these latent features to make predictions, as well as compute task-uncertainties U t for regularizing the amount of transfer from tasks/domains in A to the latent features in Z by employing an autoencoder regularization. Using adversarial learning, the discriminator D receives the extracted features from Z and regularizes L to learn domain-invariant features. L, S, A and D consist of a single fully connected layer.
We provide two notebooks as examples of how the training and testing is perfomed using TUGDA's framework. Both notebooks are self-contained (e.g., install required libs and load the necessary data).
Step zero: bash setup_repo.sh
-
For MTL settings, please refer to notebooks. In this notebook you can reproduce TUGDA's result for Figure 2.
-
For Domain Adaptation settings, please refer to notebooks. In this notebook, you can reproduce TUGDA's result for the domain adaptation from cell-lines to PDX (Figure 3).
In this repository we used data from the publicly available GDSC and PDX Novartis datasets.
Peres da Silva, R., Suphavilai, C. & Nagarajan, N. TUGDA: task uncertainty guided domain adaptation for robust generalization of cancer drug response prediction from in vitro to in vivo settings. Bioinformatics 37, i76–i83 (2021). OUP Bioinformatics
For additional information, help and bug reports please email Rafael Peres da Silva ([email protected])