Tissue-guided LASSO for prediction of clinical drug response using preclinical samples

Abstract

1Prediction of clinical drug response (CDR) of cancer patients, based on their clinical and 2 molecular profiles obtained prior to administration of the drug, can play a significant role 3 in individualized medicine. Machine learning models have the potential to address this 4 issue, but training them requires data from a large number of patients treated with each 5 drug, limiting their feasibility. While large databases of drug response and molecular 6 profiles of preclinical in-vitro cancer cell lines (CCLs) exist for many drugs, it is unclear 7 whether preclinical samples can be used to predict CDR of real patients. 8 9We designed a systematic approach to evaluate how well different algorithms, trained on 10 gene expression and drug response of CCLs, can predict CDR of patients. Using data from 11 two large databases, we evaluated various linear and non-linear algorithms, some of 12 which utilized information on gene interactions. Then, we developed a new algorithm 13 called TG-LASSO that explicitly integrates information on samples' tissue of origin with 14 gene expression profiles to improve prediction performance. Our results showed that 15 regularized regression methods provide significantly accurate prediction. However, 16including the network information or common methods of including information on the 17 tissue of origin did not improve the results. On the other hand, TG-LASSO improved the 18 predictions and distinguished resistant and sensitive patients for 7 out of 13 drugs. 19Additionally, TG-LASSO identified genes associated with the drug response, including 20 known targets and pathways involved in the drugs' mechanism of action. Moreover, 21 utilizes a new approach for explicitly incorporating the tissue of origin of samples in the 43 prediction task. Our results show that TG-LASSO outperforms all other algorithms and can 44 accurately distinguish resistant and sensitive patients for the majority of the tested drugs. 45Follow-up analysis reveal that this method can also identify biomarkers of drug sensitivity 46 in each cancer type. 47

Artificial intelligence, machine learning, and drug repurposing in cancer

“…For drug repurposing, there is an added need for accurate tissue-specific drug efficacy predictions to study the efficacy of a drug in a relevant tissue-of-origin. Recent models, such as tissue-guided LASSO, make use of information on samples' tissue-of-origin to improve in vivo prediction performance [116]. It was shown that tissue-guided LASSO improves the clinical predictions and was able to distinguish resistant and sensitive patients for selected drugs.…”

A Deep Learning Framework for Prediction of Clinical Drug Response of Cancer Patients and Identification of Drug Sensitivity Biomarkers using Preclinical Samples

“…We used data corresponding to RECIST CDR of TCGA patients carefully collected in [10] and identified 14 drugs that satisfy two conditions: 1) there were at least 20 patients with known CDR values for each drug in TCGA database and 2) the ln(IC50) response of these drugs were measured in the GDSC database. Similar to previous studies [9, 14], we transformed the CDR of these tumours into a Boolean label in which “resistant” referred to patients with CDR of “stable disease” or “progressive disease” and “sensitive” referred to patients with CDR of “complete response” or “partial response”. These CDR values were used to evaluate the predicted drug response values using TINDL and other algorithms but were not used for training them.…”

“…On the other hand, large databases of molecular profiles of hundreds of in-vitro cancer cell lines (CCLs) and their response to hundreds of drugs [3–5] have enabled development of various ML algorithms for prediction of drug response [6–8]. Unfortunately, these models, even though accurate in predicting the drug response of held-out CCLs, usually do not generalize well to predicting the CDR of real tumours from cancer patients, and their prediction performance significantly deteriorates due to the major biological and statistical differences between CCLs and tumours [9].…”

“…In this study, our goal was to develop a deep learning computational pipeline, fully trained on gene expression profile and drug response of preclinical CCLs, to 1) predict the CDR of cancer patients and 2) to identify biomarkers of drug sensitivity for a variety of cancer drugs. Motivated by Huang et al [9], who showed that carefully incorporating information on the tissue (or cancer) type of the test samples can improve the predictive power of computational models, we developed a d eep learning pipeline with t issue-informed n ormalization (TINDL) to achieve these goals. Unlike methods mentioned above, TINDL requires normalization of only test samples, and as a result re-training of the model is not necessary for new test samples.…”

“…The first phase is responsible for prediction of CDR of cancer patients, while the second phase utilizes these predictions to identify a small number of genes that considerably contribute to the predictive ability of the model. Focusing on drugs shared between the Genomics of Drug Sensitivity in Cancer (GDSC) [3] and TCGA [2], we showed that TINDL can distinguish between the sensitive and resistant patients for 10 (out of 14) drugs, considerably improving the performance of other methods, including our previous work TG-LASSO [9]. TINDL utilizes a simple, yet effective, tissue-informed normalization to reduce the statistical discrepancies between the gene expression profile of the training and test samples.…”