Predicting Drug Response in Cancer in Cancer Cell Lines using Deep Learning for precision treatments


Year:

Session type:

Theme:

Jyothi Deva datta reddy1,Dr.ram Reddy2,Samyukta Reddy3
1Cancer moonshot @ www.cancermoonshot.in,2Care cancer center,3CancerMoonshot

Abstract

Background

Precision oncology aims to improve cancer patient outcomes but  major challenge in cancer treatment is predicting the clinical response to anticancer drugs for each individual patient.

Since cancer, characterized by high inter-patient variance, the implementation of precision medicine approaches is dependent upon understanding the disease process at the molecular level. We are applying recent advances with Deep Learning Neural Networks (DLNN), suggests that DLNN could be trained on large data sets to efficiently predict therapeutic responses

Method

Data:

Compiled from CCLP & GDSC, 10001 Cell lines, 251 drugs We used GDSC 8 as our drug response data source for 139 therapeutic compounds, which provided IC-50 values for each compound, as well as information on tissue origin. Given their molecular profiling data, both large cell-line panels  (CCLE and GDSC) have been utilized in attempts to identify biomarkers for predicting drug  response of specific cancer cell lines 

Today’s complex “omic” data sets have been proven too multi-dimensional to be effectively managed by classical Machine Learning algorithms. To our knowledge, this is the first time that the DLNN framework is systematically applied to predict drug efficacy against cancer.

Technology – Deeplearning, Expert system, GPU’s

Results

According to the widely accepted AUC-based classification quality grading scale, classifiers that produce AUCs 0.90 - 1 are considered excellent, 0.80 - 0.90 are good, 0.70 - 0.80 are fair, 0.60 - 0.70 are poor classifiers, while classifiers with an AUCs below 0.6 are considered failed or random classifiers 34 .

Out of a total of 278 classification task corresponding to 139 drugs each with two responses (sensitivity and resistance), our pipeline produced 276 classifiers. Out of the 276 trained and tested DLNN classifiers, approximately 1% were excellent, 17% good, 54% fair, 24% poor

Conclusion

Recent advances with Deep Learning Neural Networks (DLNN),  this could be applied on large data sets to efficiently predict therapeutic responses