We aim to develop novel AI tools and build robust deep learning models by combining chemical biological data and cancer cell multi-omics information that could accurately predict drug sensitivity. To this end, we designed a novel approach, termed DL4DR, which adopts Octave Residual Convolution neural networks (ORNNs) module to build models based on >140K chemical compounds for 1,056 cancer cell lines. In particular, we factorize the mixed feature maps by their frequencies and can therefore focus on feature maps with a lower spatial resolution, saving both memory and computational cost. This has been employed to build models for individual breast cancer cell lines as well as a universal model for all breast cancer cell lines. Our future goal is to build a foundation model for all cancers which can then be fine-tuned for specific tasks, for instance, herein for predictions of drug responses in all cancer types as well as for individual cancer patients. For more details, please refer to Frequently Asked Questions.

Input: Compound Names and SMILES