Prediction of Human Population Responses to Toxic Compounds by a Collaborative Competition

Abstract

When it becomes completely possible for one to computationally forecast the impacts of harmful substances on humans, it would be easier to attempt addressing the shortcomings of existing safety testing for chemicals. In this paper, we relay the outcomes of a community-facing DREAM contest to prognosticate the harmful nature of environment-based compounds, considering their likelihood to have disadvantageous health-related effects on the human populace. Our research quantified the cytotoxicity levels in 156 compounds across 884 lymphoblastic lines of cells. For the cell lines, the transcriptional data and genotype are obtainable as components of the initiative known as the Tox21 1000 Genomes Project. In order to accurately determine the interpersonal variations between toxic responses and genomic profiles, algorithms were created by participants in the DREAM challenge. They also used this means to predict the inter-individual disparities of cytotoxicity-related data at the populace level from the organizational characteristics of the considered environmental compounds. A sum of 179 predictions was submitted and then evaluated at odds with experiment-derived data set to the blinded lot of participants. The cytotoxicity forecasts performed better than random, showcasing modest interrelations and consistency with a complexity of trait genomic prognostics. Contrastingly, the response of population-level predictions to a variety of compounds proved higher. The outcomes spotlight the likeliness of forecasting health-associated risks with regards to unidentified compounds, despite the reality that one’s risk with estimation accuracy persists as suboptimal. Most of the computational means through which chemical toxicity can be predicted are more often than not based on non-mechanistic cheminformatics-inspired solutions. They are typically also reliant on descriptions in QSAR arsenals and usually related to chemical structures rather vaguely. The majority of these computational methods for determining toxicness also employ black-box math algorithms. Be that as it may, while such machine learning models might possess much lower capacities for generalization and interpretability, they often achieve high accuracy levels when it comes to predicting a variety of toxicity results. And this is reflected unambiguously by the outcomes of the Tox21 competition. There is a huge capitalization on the ability of present-day Artificial Intelligence (AI) to determine the benchmark data of Tox21 with the aid of a series of 2D-rendered chemical drawings, using no chemical descriptors whatsoever. In particular, we processed some unimportant 2D-based molecules sketches within a controlled convolutional neural 2D network—also represented as 2DConvNet). We also demonstrated that today’s image recognition tech culminates in prediction correctness which can be compared to cutting-edge cheminformatics contraptions. Moreover, the 2DConvNet’s image-based model was evaluated comparatively dwelling on a set of external compounds from the stables of the Prestwick chemical library. They led to an experimental recognition of substantial and initially undocumented antiandrogen tendencies for diverse drugs in the generic and well-established categories.