Abstract 2090: Machine learning-derived gene signature predicts strong sensitivity of several solid tumors to the alkylating agent LP-184

Abstract

Abstract A novel clinical agent, LP-184 is being developed in conjunction with a dedicated machine learning-guided response signature, to allow optimal benefit of LP-184 through genomics-guided therapy. To define correlates of tumor genomics with sensitivity to LP-184, we used RADR™ (Response Algorithm for Drug Positioning and Rescue), a proprietary artificial intelligence (AI)-driven platform, and CellMinerCDB (cross database)™, a systems biology platform integrating molecular and pharmacological datasets on cancer cell lines. The input for our correlational analyses, include LP-184 IC50 data on NCI-60 cell line panel representing drug response, and multi-omics data on these cell lines. CellMinerCDB™ based analysis of LP-184 using a Lasso regression model generated a 38 gene response signature that included expression of APP, NEK6, EGFR, SQSTM1, SLC25A42, SLC16A10, POLD1, SMARCC1, POLG2, CHEK1. In building the model, the signature demonstrates a sensitivity R2 = 0.98 between observed and predicted IC50 values. Additional omics data including methylation and protein levels validate the relevance of several signature genes: APP, NEK6, EGFR, SQSTM1. Concurrently, a RADR™ based analysis of LP-184 using an Artificial Neural Network (ANN) classifier model generated a 10 gene response signature with PTGR1 and PTPN14 as the top weighted genes by relative importance. To establish a signature with highest possible sensitivity, we integrated these 2 additional genes in the 38 gene signature, creating a modified 40 gene signature now having an R2 = 0.99. We next used laboratory experimental studies to measure the performance of this signature. Using an independent set of 18 cell lines not used in signature development, we obtained IC50 values and categorized cell lines as sensitive or resistant. Our in silico results matched in vitro experimental results for 13/18 cell lines (72 percent accuracy). We further applied this 40 gene signature to interrogate 1036 cell lines representing a spectrum of tumor types in the Cancer Cell Line Encyclopedia (CCLE) and obtain a signature-derived IC50. The predicted IC50 ranges from 2.7nM to 114.6uM. The least responsive tumors represent hematological cancers, reproducing the observations made in the NCI-60 panel. 92 of 116 (79%) hematologic cancers in the CCLE database showed a predicted IC50 above 1µM with a median of 3.4uM. The top 279 cell line records with a predicted IC50 below 100nM represented solid tumors with NSCLC, renal, CNS and head & neck cancers as the most sensitive. In conclusion our results demonstrate that the development of LP-184 guided by tumor gene expression patterns modeled using a combination of algorithms and signatures provides a valuable component to the armamentarium of drugs in diverse solid tumors. Citation Format: Umesh Kathad, Aditya Kulkarni, Jean Philippe Richard, Terri Lehman, Rama Modali, Kishor Bhatia, Panna Sharma, Fathi Elloumi, Yves Pommier, William C. Reinhold. Machine learning-derived gene signature predicts strong sensitivity of several solid tumors to the alkylating agent LP-184 [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 2090.