Towards natural mimetics of metformin and rapamycin

Abstract

Aging is now at the forefront of major challenges faced globally, creating an immediate need for safe, widescale interventions to reduce the burden of chronic disease and extend human healthspan. Metformin and rapamycin are two FDA-approved mTOR inhibitors proposed for this purpose, exhibiting significant anti-cancer and anti-aging properties beyond their current clinical applications. However, each faces issues with approval for off-label, prophylactic use due to adverse effects. Here, we initiate an effort to identify nutraceuticals-safer, naturally-occurring compounds-that mimic the anti-aging effects of metformin and rapamycin without adverse effects. We applied several bioinformatic approaches and deep learning methods to the Library of Integrated Network-based Cellular Signatures (LINCS) dataset to map the gene- and pathway-level signatures of metformin and rapamycin and screen for matches among over 800 natural compounds. We then predicted the safety of each compound with an ensemble of deep neural network classifiers. The analysis revealed many novel candidate metformin and rapamycin mimetics, including allantoin and ginsenoside (metformin), epigallocatechin gallate and isoliquiritigenin (rapamycin), and withaferin A (both). Four relatively unexplored compounds also scored well with rapamycin. This work revealed promising candidates for future experimental validation while demonstrating the applications of powerful screening methods for this and similar endeavors.

Keywords: compound screening; deep learning; geroprotector; metformin; natural; nutraceutical; rapamycin.

Figure 1

Workflow diagram showing multi‐level analysis for screening and ranking nutraceuticals that mimic metformin and rapamycin in transcriptional and pathway activation response. A subset of 871 LINCS compounds were selected from the UNPD and KEGG BRITE databases. Perturbations with those compounds in cancer cell lines were compared with perturbations with metformin and rapamycin to estimate similarity at the gene and pathway level and deep learning techniques were employed to recognize the transcriptional signature of metformin and rapamycin and screen for matches amongst the LINCS compounds.

Figure 2

DL‐based similarity to metformin (A) and rapamycin (B). Significance of natural compound was determined as the ‐log10(p‐value) and odds ratio for compound according to Fisher's exact test performed on the DNN output for each perturbed sample. Only compounds with ‐log10(p‐value)>4 and odds ratio > 1 are shown.

Figure 3

Gene‐ and pathway‐level similarity to metformin (A) and rapamycin (B). Significance of natural compound was determined as the ‐log10(p‐value) of the most significant perturbation of compound according to Fisher's exact test. Percentage of common pathways designates the amount of pathways that have the same direction of the change as Metformin. Only compounds with ‐log10(p‐value)>4 and over 50% of common pathways are shown.