Text Mining-Based Suspect Screening for Aquatic Risk Assessment in the Big Data Era: Event-Driven Taxonomy Links Chemical Exposures and Hazards

Abstract

To improve the accuracy of mixture risk assessment, researchers are employing suspect analysis with expanded lists of contaminants in addition to conventional target lists. However, there are some inherent challenges for these instrument-based analyses, including subjective selection of suspect contaminants, no information for chemical bioactivity, requirements for costly verification, and limited regional coverage. As a supplementary approach, we propose a data-driven suspect screening and risk assessment method informed by mining big data from high-throughput screening bioassay platforms and the refereed literature. The Pearl River Delta (PRD) with main event drivers of arylhydrocarbon receptor (AhR) and oxidative stress (ARE) response was examined. Bioactivity concentrations were collected from the CompTox Chemicals Dashboard, which contained more than 900 000 substances. In addition, exposure metadata from 24 986 literature entries for the environmental occurrence and distribution of contaminants in the PRD over the past three decades were mined. Collectively, a regional distribution map of aquatic hazards induced by AhR- and ARE-active compounds was generated, indicating gradients of low to moderate risks. This study specifically reports a novel big data approach for addressing the increasingly common challenge of objectively selecting analytes during suspect screening, which was recently identified as an urgent research question to advance more sustainable environmental quality.

Prioritizing Organic Pollutants for Shale Gas Exploitation: Life Cycle Environmental Risk Assessments in China and the US

“…Bibliographic information, including the title and abstract, was indexed. Subsequently, a python program was developed and employed to facilitate the batch downloading of full-text literature . Documents were then converted from PDF to TXT format to serve as the raw corpus using python codes obtained from the PDFMiner library ().…”

“…Subsequently, a python program was developed and employed to facilitate the batch downloading of full-text literature. 28 Documents were then converted from PDF to TXT format to serve as the raw corpus using python codes obtained from the PDFMiner library (https://pypi.org/ project/pdfminer/). Later, an original lexicon was constructed by refining characteristic contaminants of hydraulic fracturing published by the USEPA (1,639 substances list), resulting in a final selection of 512 chemicals that contain environmental toxicity and biological activity data after referencing the Ecotoxicology knowledgebase (version 4) 28 and USEPA ToxCast invitrodb (v3.5) Chemical Lists (SI_2 data_01).…”

Deep Learning Bridged Bioactivity, Structure, and GC-HRMS-Readable Evidence to Decipher Nontarget Toxicants in Sediments

“…As shown in Figure , the EDT-based workflow consists of four phases including (i) experimental characterization of sediment extracts with AhR-CALUX and AREc32 HTS bioassays; (ii) evaluation of toxicity contribution from GC-MS quantified target contaminants by mixture toxicity modeling; (iii) construction of AhR- and ARE-ED related substance data sets through big data suspect screening; and (iv) DL-supported NTA using GC-qToF-MS with new evidence of HRMS-recognizable fragment ions (EDION). Experimental work included sediment collection, extraction, HTS bioassays, and chemical analyses (target and nontarget).…”

“…Workflow of the event-driven taxonomy (EDT)-based integrated testing strategy for mixture toxicity identification by integrative high-throughput screening (HTS), nontarget analysis (NTA) and deep learning (DL). (i) HTS bioassays with sediment extracts, (ii) target analysis of the sediment extracts using GC-MS, (iii) Data set construction using a previous big data suspect screening method, (iv) DL-supported NTA. PDMS: polydimethylsiloxane, ARE: antioxidant response element, AhR: aryl hydrogen receptor, EC 50 /AC 50 : effective/activity concentration triggering 50% of maximum activity, B: bioactivity, S: size and shape, M: molecular structure, MS: mass spectrometry; ED: event driver; EDFP: fingerprints related to a certain ED; EDION: fragment ions that are recognizable by MS and predicted from EDFP for a certain ED.…”

“…The objective of the present study was to develop a DL-assisted strategy that can bridge the data gap between toxicity profiles and chemical information through EDT and further integrate HTS bioactivity data to aid identification of nontarget toxicants in sediments. First, through our previously developed suspect screening approach, a pool of known substances eliciting certain bioactivities was retrieved (prior knowledge). , Second, through DL, these chemically diverse substances generating similar bioactivities were labeled according to the type of bioactivity (ED), and fingerprints of these substances corresponding to this bioactivity (defined as event driver fingerprint, EDFP) were predicted . Third, currently available MS databases was helpful to establish prediction models from fingerprints to HRMS-readable evidence (i.e., event driver ion, EDION) that can be used to characterize unknown pollutants causing the same bioactivity .…”