# TransCDR: a deep learning model for enhancing the generalizability of cancer drug response prediction through transfer learning and multimodal data fusion for drug representation

**Authors:** Xiaoqiong Xia, Chaoyu Zhu, Yuqi Shan, Fan Zhong, Lei Liu  
**arXiv:** [2311.12040](https://arxiv.org/abs/2311.12040)  
**Date:** 2023-11-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.12040v1/

## Key points

- Precision medicine relies on accurate drug response prediction for personalized treatment strategies
- Challenges in predicting cancer drug responses include limited data modalities, suboptimal fusion algorithms, and poor generalizability to novel drugs or cell lines
- TransCDR is a novel approach that uses transfer learning and self-attention mechanism to predict drug responses
- TransCDR excels in evaluating generalization of CDR prediction models to new compound scaffolds and cell line clusters
- Key factors influencing drug response prediction are Extended Connectivity Fingerprint and genetic mutations
- TransCDR outperforms state-of-the-art models and shows strong predictive capabilities on external testing sets like CCLE
- The model can be used to investigate biological mechanisms underlying drug response through Gene Set Enrichment Analysis
- Availability of source code and data on GitHub allows for further exploration and application of TransCDR

## Summary

, , , , 

In the field of precision medicine, accurate and robust drug response prediction is crucial for personalized treatment strategies. Various models have been developed to predict cancer drug responses (CDR) by leveraging drug and cancer cell line representations. However, challenges such as limited data modalities, suboptimal fusion algorithms, and poor generalizability to novel drugs or cell lines still exist. To address these issues, a novel approach called TransCDR has been introduced. TransCDR utilizes transfer learning to acquire drug representations and integrates multi-modality features of drugs and cell lines through a self-attention mechanism to predict IC50 values or sensitive states of drugs on cell lines. One key aspect that sets TransCDR apart is its ability to systematically evaluate the generalization of CDR prediction models to never-before-seen compound scaffolds and cell line clusters. In comparative evaluations against 8 state-of-the-art models, TransCDR demonstrates superior generalizability. Furthermore, it outperforms five variants that train drug encoders from scratch (such as RNN and AttentiveFP) across different scenarios. The most influential factors identified among multiple drug notations and omics profiles are Extended Connectivity Fingerprint and genetic mutations. The incorporation of an attention-based fusion module further enhances the predictive performance of TransCDR. Trained on the GDSC dataset, TransCDR exhibits strong predictive capabilities on external testing sets like CCLE and is also effective in predicting missing CDRs within GDSC. Additionally, the model is employed to investigate the biological mechanisms underlying drug response by categorizing 7,675 patients from TCGA into drug-sensitive or drug-resistant groups followed by Gene Set Enrichment Analysis. Overall, TransCDR emerges as a powerful tool with significant potential in advancing drug response prediction in precision medicine. The availability of source code and data on GitHub allows for further exploration and application of this innovative approach in the field.