Formatted Title
Shifting the Paradigm of Characterization and Remedy Decisions with Application of Machine-Learning Algorithms and Molecular Biological Tools
Background/Objectives
Biogeochemical processes, along with other physicochemical processes, govern rates of attenuation at contaminated sites. Characterization of these processes is critical to selecting, designing, and implementing efficient remedial strategies. Despite increased understanding of the influence of biological processes on the environmental fate of contaminants, the historical site assessment paradigm is focused on evaluations of physicochemical attenuation processes. During the past 20 years, molecular biological tools (MBTs), such as gene sequencing and quantitative polymerase chain reaction (qPCR), have been increasingly utilized to directly assess important biogeochemical processes at contaminated sites. However, the generation of high-dimensionality big (biogeochemical) data and the complexities of environmental microbiology can be challenges to effectively analyze and extract biogeochemical processes to support actionable remedy decisions and broader uptake within contaminated site management.
Approach/Activities
Application of machine-learning algorithms capable of addressing datasets with far more parameters (contaminants, geochemical parameters, MBTs) than samples typical of datasets for contaminated sites enables new approaches to synthesize data and support remedial decision making. Data science algorithms, such as the chemometrics/ecological algorithms sparse partial least squares (sPLS) and sparse redundancy analysis (sRDA), can reduce the multi-dimensionality of the data to identify natural, site-specific trends, such as biogeochemical gradients, key contaminant-degrading microorganisms, and potential limitations to contaminant biodegradation/biotransformations. A case study will be highlighted that focuses on the application of these models to a hydrocarbon-impacted source area and groundwater plume to support the conceptual site model and remedial strategy.
Results/Lessons Learned
Baseline samples collected from representative areas of a petroleum hydrocarbon-impacted groundwater plume (upgradient, source area, and downgradient) highlighted distinct differences between in-plume and upgradient microbial communities based on 16S rRNA gene sequencing and qPCR results. Application of unsupervised sPLS and sRDA algorithms focused identification to key microorganisms (genera level) involved in contaminant degradation and important biogeochemical processes. While these results largely aligned with more traditional evaluation approaches performed by subject matter experts, additional key microorganisms and subtle spatial biogeochemical were identified. A key benefit of this machine-learning approach is that it avoids common, deterministic rules-of-thumb (e.g., filtering sequencing data by percent abundance) and data reduction methods, in favor of identifying data-driven trends that better reflect actual site conditions. Further, it leverages small datasets with many analyzed parameters common to environmental datasets in an efficient data evaluation pipeline. Machine-learning applications to assess biogeochemical data, including MBTs, have the potential to shift the historical site assessment paradigm to further increase knowledge of field-scale microbiological processes, improve bioremediation approaches to be more precisely engineered, and ultimately achieve bioremediation goals and eventual site closure.