Source-linked AI summary
Predicting drug response of tumors from integrated genomic profiles by deep neural networks
Yu-Chiao Chiu, Hung-I Harry Chen, Tinghe Zhang, Songyao Zhang, Aparna Gorthi, Li-Ju Wang, Yufei Huang, Yidong Chen
TL;DR
Tumor heterogeneity and intra-tumor sub-clones make accurate drug-response prediction and novel drug identification challenging. The study applies a model to tumor drug-response prediction, reporting IC50 prediction performance and identifying drug-response findings including CX-5461 potential.
Problem
Tumor heterogeneity and intra-tumor sub-clones make accurate drug-response prediction and identification of novel anti-cancer drugs challenging.
Method
The model is applied to predict tumor drug response and investigate the pharmacogenomics landscape of drugs.
Results
The model achieved an overall IC50 MSE of 1.53, with testing-sample MSE of 1.96, and identified CX-5461 potential in hematopoietic malignancies and LGG.
Takeaways & Limitations
The findings support the model’s applicability to tumor drug-response prediction and identify potential therapeutic options including CX-5461.
Takeaways & Limitations
The method suffers from complexity associated with high representational power, while predicted IC50 differences were modest.
Abstract
from arXiv · showhide
The study of high-throughput genomic profiles from a pharmacogenomics viewpoint has provided unprecedented insights into the oncogenic features modulating drug response. A recent screening of ~1,000 cancer cell lines to a collection of anti-cancer drugs illuminated the link between genotypes and vulnerability. However, due to essential differences between cell lines and tumors, the translation into predicting drug response in tumors remains challenging. Here we proposed a DNN model to predict drug response based on mutation and expression profiles of a cancer cell or a tumor. The model contains a mutation and an expression encoders pre-trained using a large pan-cancer dataset to abstract core representations of high-dimension data, followed by a drug response predictor network. Given a pair of mutation and expression profiles, the model predicts IC50 values of 265 drugs. We trained and tested the model on a dataset of 622 cancer cell lines and achieved an overall prediction performance of mean squared error at 1.96 (log-scale IC50 values). The performance was superior in prediction error or stability than two classical methods and four analog DNNs of our model. We then applied the model to predict drug response of 9,059 tumors of 33 cancer types. The model predicted both known, including EGFR inhibitors in non-small cell lung cancer and tamoxifen in ER+ breast cancer, and novel drug targets. The comprehensive analysis further revealed the molecular mechanisms underlying the resistance to a chemotherapeutic drug docetaxel in a pan-cancer setting and the anti-cancer potential of a novel agent, CX-5461, in treating gliomas and hematopoietic malignancies. Overall, our model and findings improve the prediction of drug response and the identification of novel therapeutic options.
Background
Tumor heterogeneity and differences between in vitro and in vivo systems leave tumor drug-response prediction unresolved. The study addresses this gap with a deep neural network integrating mutation and expression profiles.
- Tumor heterogeneity and intra-tumor sub-clones make accurate drug-response prediction difficult.
- Large-scale cell-line screening provides pharmacogenomic data because drug-response data from large patient cohorts are scarce.
- Differences between in vitro and in vivo biological systems have prevented realized translation from cell-line pharmacogenomics to tumor-response prediction.
- Deep learning has been applied to genomic and pharmacological tasks, including regulatory genomics, drug-target interaction prediction, drug screening, and drug synergy prediction.
- Data complexity and large training-data requirements have limited deep learning for integrating genomics data and comprehensively predicting drug response.
- The study develops a DNN that uses TCGA tumor data to learn mutation and expression representations, trains on GDSC cell-line pharmacogenomics, and predicts tumor drug response.
- The proposed model bridges cell-line pharmacogenomics knowledge and tumor genomic and transcriptomic abstraction to predict response to compound treatment.
Methods
The model combines mutation and expression encoders with a prediction network, using TCGA data for representation learning and CCLE data for IC50 prediction. The study evaluates 265-drug responses from matched genomic profiles and applies the trained model to tumors.
- Data: Expression data were collected for 935 cell lines, while mutation data came from CCLE and TCGA databases.The mutation dataset included 1,463 CCLE cells and 10,166 TCGA tumors.
- Data: The analysis retained four nonsynonymous mutation types and represented gene mutation states as binary mutation-or-wildtype values.Missense, nonsense, frameshift insertion, and frameshift deletion mutations were considered.
- Data: IC50 responses for 265 anti-cancer drugs were obtained from GDSC data, with IC50 measured in µM and represented on a log scale.Missing IC50 values were imputed using a weighted mean from five nearest drugs.
- Data: The analyzed datasets contained 622 cell lines with expression, mutation, and IC50 data and 9,059 tumors with expression and mutation profiles.
- Model: The model receives paired mutation and expression vectors, encodes them with Menc and Eenc, and predicts a D-length IC50 vector through P.Menc and Eenc are pretrained autoencoder encoders that reduce dimensionality and capture high-order features.
- Model: Menc and Eenc were pretrained separately on TCGA mutation and expression datasets, then linked to P and trained on CCLE data.
- Model: The complete model used a five-layer feedforward prediction network whose final layer generated IC50 values for d drugs.
- Evaluation: The complete network was trained with 80% training, 10% validation, and 10% testing samples, stopping after validation loss failed to decrease for three consecutive epochs.Performance was evaluated on the testing samples.
Results and Discussion
The model predicted tumor drug responses from genomic profiles, showing strong cell-line performance, tumor applicability, and pharmacogenomic patterns for established and potential therapies. Analyses identified drug-associated mutations, docetaxel resistance mechanisms, and tumor groups potentially sensitive to CX-5461.
- Model construction and evaluation in CCLE: Predicted IC50 values were highly consistent with true values in both magnitude and drug ranking, with Pearson correlations from 0.70 to 0.96 and Spearman correlations from 0.62 to 0.95.The corresponding P-values were all below 1.0×10^-29, and correlations were highly comparable between training/validation and testing samples.
- Model construction and evaluation in CCLE: The model outperformed linear regression, SVM, randomly initialized DNNs, and PCA-based DNNs in prediction error or stability.Linear regression and SVM had testing MSE values of 10.24 and 8.92; the model’s testing MSE variability was 0.13 versus 1.21 for random initialization and 0.17 for PCA.
- Tumor response prediction: The model reproduced known tumor-response patterns, including better tamoxifen response in ER-positive breast cancer and better afatinib and gefitinib response in EGFR-mutated NSCLC.Tamoxifen response differed between ER-positive and ER-negative patients at P = 2.3×10^-4; afatinib and gefitinib associations had P = 2.0×10^-7 and 6.6×10^-3.
- Pan-cancer pharmacogenomics: In pan-cancer analysis, 88.8% of significant mutation–drug pairs were associated with resistance and 11.2% with sensitivity in mutated samples.TP53, CSMD3, SYNE1, TTN, and RYR2 were the top drug response-modulating genes; TP53 was associated with 242 resistant and 9 sensitive drugs.
- CX-5461 pharmacogenomics: CX-5461 sensitivity was concentrated in hematopoietic malignancies and LGG, while TP53 mutations marked resistant tumors and IDH1 mutations preferentially marked sensitive LGG tumors.DLBC and another hematopoietic malignancy jointly accounted for 45.1% of predicted highly sensitive patients, LGG accounted for 48.4%, and TP53 mutation rates were 95.6% in resistant versus 13.2% in sensitive patients.
Conclusions
The study developed a DNN that transfers genomic representations learned from cancer cell lines to tumors for drug-response prediction. Applied to TCGA tumors, it identified established and novel resistance mechanisms and drug targets.
- The DNN extracts representative mutation and gene-expression features to transfer knowledge from cancer cell lines to tumors.
- The model showed reliable performance and outperformed four other methods.
- TCGA application identified both well-studied and novel resistance mechanisms and drug targets.
- The model can incorporate other omics data and study a wider range of drugs.
- The authors position the approach as supporting the realization of precision oncology.
Declarations
The declarations report data availability, absence of competing interests, funding sources, and author roles. The dataset supporting the conclusions is included within the article.
- Data availability: The dataset supporting the conclusions is included within the article.
- Competing interests: The authors declare that they have no competing interests.
- Funding: The research and publication costs were supported partially by NCI, NIH, CPRIT, and the San Antonio Life Science Institute.
- Funding: The funding sources had no role in study design, data collection, analysis, or interpretation.
- Author contributions: The authors conceived the study, designed the model, performed and interpreted analyses, and wrote and approved the paper.
Tables
The tables summarize model performance, mutation associations with drug response, and GO-term enrichment analyses for docetaxel and CX-5461.
- Model performance: Table 1 reports performance of the study’s DNNs and other models.
- Pan-cancer mutations: Table 3 identifies top gene mutations modulating pan-cancer drug response, including TP53, CSMD3, SYNE1, TTN, and RYR2.
- Docetaxel: Table 4 summarizes GO clusters enriched among genes associated with predicted docetaxel response, including mitotic cell-cycle functions.
- CX-5461: Table 5 summarizes GO clusters enriched among genes associated with predicted CX-5461 response, including extracellular matrix and cell-motion terms.
Figures
The figures present the model architecture and CCLE evaluation, then apply predicted drug responses to mutation associations and pan-cancer pharmacogenomic analyses. Together, they cover model construction, performance assessment, mutation–drug relationships, and resistant-versus-sensitive tumor profiles.
- Model construction: The model uses mutation and expression encoders linked to a prediction network and is trained on 622 CCLE samples to predict responses to 265 drugs.The architecture includes separate mutation and expression encoders, followed by a prediction network for IC50 values.
- Per-cancer TCGA associations: The TCGA analyses examine predicted IC50 differences between tumors with and without mutations and identify significant mutation–drug associations within cancer types.Significance was assessed with two-tailed t-tests, and the analysis highlights extreme adjusted P values and large absolute ΔIC50 values.
- Pan-cancer mutation associations: Across TCGA samples, the analysis focuses on 11 genes with mutation rates above 10% and shows examples of drugs modulated by TP53 and TTN mutations.The displayed thresholds include adjusted P < 1.0×10^-200 and ΔIC50 ≥ 0.7 or ΔIC50 < 0.