Artificial intelligence uncovers carcinogenic human metabolites

Abstract

The genome of a eukaryotic cell is often vulnerable to both intrinsic and extrinsic threats due to its constant exposure to a myriad of heterogeneous chemical compounds. Despite the availability of innate DNA damage repair pathways, some genomic lesions trigger cells for malignant transformation. Accurate prediction of carcinogens is an ever-challenging task due to the limited information about bonafide (non)carcinogens. This, in turn, constrains the generalisability of such models. We developed a novel ensemble classifier (Metabokiller) that accurately recognizes carcinogens by quantitatively assessing their chemical composition as well as potential to induce proliferation, oxidative stress, genotoxicity, alterations in epigenetic signatures, and activation of anti-apoptotic pathways, therefore obviates the need for bonafide (non)carcinogens for training model. Concomitant with the carcinogenicity prediction, it also reveals the contribution of the aforementioned biochemical processes in carcinogenicity, thereby making the proposed approach highly interpretable. Metabokiller outwits existing best practice methods for the carcinogenicity prediction task. We used Metabokiller to decode the cellular endogenous metabolic threats by screening a large pool of human metabolites and identified putative metabolites that could potentially trigger malignancy in normal cells. To cross-validate our predictions, we performed an array of functional assays and genome-wide transcriptome analysis on two Metabokiller-flagged, and previously uncharacterized human metabolites by using Saccharomyces cerevisiae as a model organism and observed larger synergy with the prediction probabilities. Finally, the carcinogenicity potential of these metabolites was evaluated using a malignancy transformation assay on human cells.