Skip to content

Implementation of transfer learning approaches for predictive modeling of anticancer drug sensitivity.

Notifications You must be signed in to change notification settings

dhruba018/Transfer_Learning_Precision_Medicine

Repository files navigation

Transfer Learning for Precision Medicine

Reference: Application of transfer learning for cancer drug sensitivity prediction.

Predictive modeling of drug sensitivity is an important step in precision therapy design and often there is a shortage of suitable data for modeling. Hence we attempt to use data from multiple sources for modeling purposes and the recent advent of large-scale pharmacogenomic studies offers a convenient getaway from this conundrum.

Description

Data

We use in vitro transcriptomic and drug sensitivity information from the administration of various anticancer drugs on multiple Cancer cell lines of different subtypes. This data were obtained from two renowned large-scale pharmacogenomic studies -

There is a considerable overlap between the two studies and we have explored the existing inconsistencies between datasets in our previous work in Briefings in Bioinformatics: Evaluating the consistency of large-scale pharmacogenomic studies.

Approaches

We have combined datasets from CCLE and GDSC through Transfer Learning since the samples from two different sources cannot be used together directly. To eliminate the distribution shift present in the two sets, we have implemented two different TL approaches -

Latent Variable based Cost Optimization

We use the notion of a Latent variable space to model the underlying similarities between the genomic and sensitivity datasets and attempt to minimize the discepancies through cost optimization.

lvco_eqn

where zp and zs represents the primary (target) and secondary (source) sets, and w is the underlying latent variable.
We have implemented three different optimization based approaches -

  • Latent regression prediction
  • Latent-latent prediction
  • Combined latent prediction
Domain Transfer via Nonlinear Mapping

We implement a one-to-one feature mapping between the samples in the primary and secondary datasets using the Polynomial regression mapping to transfer the primary data to the secondary space and perform prediction using the larger datasets available in the secondary space. Note that, this approach assumes the existence of a set of matched samples between the two sets.

dtnm_eqn

where zp, k and zs, k represents the k-th feature (gene/drug) in the primary and secondary sets, and γ is the polynomial mapping coefficient.

The details of these approaches are described in our 2018 paper: Application of transfer learning for cancer drug sensitivity prediction. Below provides an overview of the TL scenarios involved in this implementation.

TL_summary

Note

We also implemented a Dimensionality Reduction based Transfer Learning approach using the Principal Component Analysis. The details of this work can be found in this paper: Dimensionality Reduction based Transfer Learning applied to Pharmacogenomics Databases.

File description

This repository contains the necessary code to reproduce the results described in the paper and the corresponding source for the data used in the simulation experiments.

  • Data: Contains data for modeling. The processed data used in our experiments can be found here
  • MappingTransLearn.m: Main function for Domain Transfer with Nonlinear Mapping approach
  • LatentPredTransLearn.m: Main function for Latent Variable based Cost Optimization approaches - includes all three approaches
  • TransLearnConcise_v2.m: Main experiment code
  • DimRedTransLearn.m: Main file for Dimensionality Reduction based Transfer Learning

How to Cite

If you use either the Domain Transfer TL approach or Latent Variable Cost Optimization TL approach for your research/application, please cite the following paper -

Dhruba, S., Rahman, R. et al. Application of transfer learning for cancer drug sensitivity prediction. BMC Bioinformatics 19, 497 (2018). DOI: https://doi.org/10.1186/s12859-018-2465-y

If you use our work on the exploration of inconsistencies in the large pharmacogenomic studies, please cite the following paper -

Rahman, R., Dhruba, S. R. et al., Evaluating the consistency of large-scale pharmacogenomic studies, Briefings in Bioinformatics, 20 (5), 1734 – 1753 (2019). DOI: https://doi.org/10.1093/bib/bby046

If you use the Dimensionality Reduction based TL approach for your research/application, please cite the following paper -

Dhruba, S.R., Rahman, R. et al., Dimensionality reduction based transfer learning applied to pharmacogenomics databases, In 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology society (EMBC), Honolulu, HI, 1246 - 1249 (2018).
DOI: https://doi.org/10.1109/EMBC.2018.8512457

About

Implementation of transfer learning approaches for predictive modeling of anticancer drug sensitivity.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages