briankimstudio/awesome-drug-response-prediction

Resources for drug response prediction resesarch

Awesome Drug Response Prediction (ADRP)

The ADRP is a curated list of resources for drug response prediction research maintained by Brian.

Goal

To identify effective anti-cancer drugs for patients by examining each patient's unique gene level information using machine learning algorithm.

Challenges

Currently, there are three major challenges regarding training data: 1) Scarcity, 2) Low ratio of samples to features (LRSF), and 3) Heterogeneity

1. Scarcity: Volume of data

• Lack of cancer clinical data

Approaches

• Merge multiple data for training : Multi-omics data analysis

  Regarding when to merge multi-omics data for training, there are two different ways.

  1. Early integration : Concatenate omics data, then learn features via autoencoders
     • Disdavangages
       1. Each omics' unique distribution is disregarded
       2. Must be normalized appropriately
       3. Dimention is increased
  2. Late integration : Learn features, then concatnate
     • Avoid disadvantages of early integration method
     • Ex) MOLI (Sharifi-Noghabi et al., 2019) : Somatic mutation, CNA, gene expression data were used for training.

• Transfer learning

• Domain adaptation

Source: wikipedia

2. Characteristics of data

• High dimensionality
• Low ratio of samples to features: The number of features often exceeds 10,000 while the number of samples is around 800. Thus, it is not ideal data set for Deep Neural Network (DNN).
• Imbalanced classes: The number of samples in each class are extremly imbalanced. In general, the number of sensitive (1) is smaller than the number of resistant (0).

Approaches

• Multi-task learning (MTL)
• Feature engineering
• Autoencoder: Feature extraction

3. Heterogeneity: Difference between in vitro and in vivo.

• Whereas training data are from cell lines or animal models(PDX), test data are from patiens. And there is difference between these two types of data.
• According to research articles, the difference is caused by batch effect.

Approaches

• Domain adaptation (DA)

Methods

Measurement

• The half-maximal inhibitory concentration (IC₅₀): Amount of drug inhibiting biological component by 50%

Type of data

Typically, data frame consists of rows representing samples and columns representing features and label.

Features

• Gene expression
• Mutations
• Copy number aberrations (CNA)
• Epigenetic data
• ...

Label (class)

• Drug sensitivity (IC₅₀) : Sensitive (1) or Resistant (0)

Public data for training

Database for preclinical data

Database	Samples	Features	Label
Cancer Cell Line Encyclopedia (CCLE)	Tumor cell lines	GE, Mutation, CNA(?)	IC50
Genomics of Drug Sensitivity in Cancer (GDSC)	Tumor cell lines	GE, Mutation, CNA	IC50
Patient-Derived Xenografts (PDX)	Animal model	GE, Mutation, CNA	RECIST

Clinical data

Database	Samples	Features	Label
The Cancer Genome Atlas (TCGA)	Patients	?	?
2014	Patients	?	?

Anti-cancer drugs

• Docetaxel
• Cisplatin
• Gemcitabine
• Paclitaxel
• Erlotinib
• Cetuximab
• ...

Common machine learning library

• PyTorch
• Keras / Tensorflow
• scikit-learn

Machine learning algorithms

Drug response prediction is a binary classification problem (sensitive or resistant). Therefore, any binary classifier can be used to estimate drug response.

• K-Nearest Neighbors (kNN)
• Support Vector Machine (SVM)
• Ridge Logistic Regression
• Random Forests
• Deep Neural Network (DNN)

References

Survey

• Adam, G., Rampášek, L., Safikhani, Z., Smirnov, P., Haibe-Kains, B., & Goldenberg, A. (2020). Machine learning approaches to drug response prediction: Challenges and recent progress. Npj Precision Oncology, 4(1), 1–10. https://doi.org/10.1038/s41698-020-0122-1
• Chen, J., & Zhang, L. (2021). A survey and systematic assessment of computational methods for drug response prediction. Briefings in Bioinformatics, 22(1), 232–246. https://doi.org/10.1093/bib/bbz164
• Firoozbakht, F., Yousefi, B., & Schwikowski, B. (2022). An overview of machine learning methods for monotherapy drug response prediction. Briefings in Bioinformatics, 23(1), bbab408. https://doi.org/10.1093/bib/bbab408

Algorithms

Deep learning

• Sakellaropoulos, T., Vougas, K., Narang, S., Koinis, F., Kotsinas, A., Polyzos, A., Moss, T. J., Piha-Paul, S., Zhou, H., Kardala, E., Damianidou, E., Alexopoulos, L. G., Aifantis, I., Townsend, P. A., Panayiotidis, M. I., Sfikakis, P., Bartek, J., Fitzgerald, R. C., Thanos, D., … Gorgoulis, V. G. (2019). A Deep Learning Framework for Predicting Response to Therapy in Cancer. Cell Reports, 29(11), 3367-3373.e4. https://doi.org/10.1016/j.celrep.2019.11.017
• Sharifi-Noghabi, H., Zolotareva, O., Collins, C. C., & Ester, M. (2019). MOLI: Multi-omics late integration with deep neural networks for drug response prediction. Bioinformatics, 35(14), i501–i509. https://doi.org/10.1093/bioinformatics/btz318

Datasets

• Ding, Z., Zu, S., & Gu, J. (2016). Evaluating the molecule-based prediction of clinical drug responses in cancer. Bioinformatics (Oxford, England), 32(19), 2891–2895. https://doi.org/10.1093/bioinformatics/btw344
• Gao, H., Korn, J. M., Ferretti, S., Monahan, J. E., Wang, Y., Singh, M., Zhang, C., Schnell, C., Yang, G., & Zhang, Y. (2015). High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nature Medicine, 21(11), 1318–1325.
• Iorio, F., Knijnenburg, T. A., Vis, D. J., Bignell, G. R., Menden, M. P., Schubert, M., Aben, N., Gonçalves, E., Barthorpe, S., & Lightfoot, H. (2016). A landscape of pharmacogenomic interactions in cancer. Cell, 166(3), 740–754.