🌿 ecology

Mining association rules for targeted spatiotemporal aquatic environmental DNA (eDNA) sampling

This paper demonstrates how unsupervised association rule mining, implemented in a new R package called RulesTools, can effectively uncover complex relationships between environmental covariates and brook trout eDNA concentrations or electrofishing outcomes from a small dataset, thereby offering a novel tool to guide targeted eDNA sampling and inform conservation decisions.

Original authors: Toth, N., Antonie, L., Hanner, R. H., Gillis, D. J., Phillips, J. D.

Published 2026-09-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Toth, N., Antonie, L., Hanner, R. H., Gillis, D. J., Phillips, J. D.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the quiet waters of a creek, life leaves behind a faint, invisible trail. When fish swim, they shed skin cells, mucus, and waste into the water, carrying their unique genetic code with them. Scientists call this environmental DNA, or eDNA. Instead of catching a fish to know it is there, researchers can simply filter a bucket of water and look for these genetic fragments. It is a gentle way to monitor wildlife, avoiding the stress and harm of traditional netting or trapping. However, finding this genetic material is not always straightforward. The DNA in water is fragile; it breaks down quickly due to sunlight, heat, and the activity of tiny organisms. Because of this, the amount of DNA found in a sample depends heavily on the conditions of the water itself, such as its temperature, acidity, and how fast it is flowing. Understanding exactly how these factors work together is crucial for conservationists who need to know where rare species live and how to protect them.

A team of researchers set out to map these hidden connections using a method usually reserved for analyzing shopping habits. They focused on a specific stretch of Hanlon Creek in Guelph, Canada, where native brook trout were known to live. In September 2019, they collected water samples and also used electric shocks to temporarily stun and catch fish, allowing them to count exactly how many trout were present at each spot. This dual approach gave them a complete picture: the genetic signal in the water and the physical reality of the fish population. Their goal was to see if they could predict the presence of trout or the concentration of their DNA just by looking at the mix of environmental conditions, like water temperature and pH levels.

To do this, the researchers turned to a technique called association rule mining. Imagine a computer scanning thousands of shopping receipts to find that people who buy bread often buy butter. The researchers applied this same logic to their environmental data. They took their measurements and sorted them into simple categories, such as "high" or "low" temperature, or "high" or "low" DNA concentration. They then asked the computer to find patterns where certain conditions consistently appeared together. From a dataset of only 126 observations, the computer initially found over 12,000 possible connections. This was far too many to make sense of, so the team used a strict filtering process to remove the weak or repetitive links, leaving behind a manageable list of 153 strong associations.

The results revealed a complex web of relationships that simple comparisons had missed. For instance, the study confirmed that higher water pH levels, which indicate more alkaline conditions, were strongly linked to finding higher concentrations of trout DNA. This aligns with what scientists already know: DNA tends to survive longer in less acidic water. However, the mining also uncovered more surprising and specific combinations. The researchers found that the type of sampling backpack used mattered; one brand seemed to consistently capture higher DNA levels than the other under certain conditions. They also discovered that the volume of water filtered played a role, with larger volumes generally leading to better detection, though not always.

Perhaps most interestingly, the study highlighted how different methods can tell different stories. In some cases, the water contained a strong signal of trout DNA, yet the electric fishing nets caught no fish at all. The data mining suggested that this might happen when the water is very cold or has low conductivity, which could mean the fish were hiding or that the DNA had drifted in from upstream. Conversely, there were situations where fish were caught, but the water showed very little DNA. These mismatches are not errors but clues. They suggest that the presence of DNA and the presence of fish are influenced by a unique combination of factors that change from moment to moment. The researchers noted that some of these patterns, such as the link between low conductivity and poor fishing results, might be due to the physics of how electric currents move through water, which affects the ability to stun fish.

The team also developed a new software tool to make this kind of analysis easier for others. They packaged their methods into a program that allows scientists to take their own environmental data, sort it into categories, and let the computer find the hidden rules that connect the dots. This tool does not just confirm what scientists already suspect; it acts as a guide, pointing out which combinations of temperature, pH, and flow rate are most likely to lead to a successful detection. It suggests that to get the best results, researchers should carefully record a wide range of environmental details, not just the water samples themselves.

While the study offers a powerful new way to look at ecological data, the authors are careful to note its limits. Because the dataset was relatively small, some of the strange patterns found might be coincidences rather than universal laws. The method is excellent for generating hypotheses and spotting complex interactions that traditional statistics might overlook, but it is not a final proof. The researchers emphasize that these findings are a starting point, a way to refine how we collect data and ask better questions. By understanding the specific conditions that help or hinder the detection of genetic material, conservationists can design better surveys, ensuring they find the species they are looking for without missing the subtle signs of life hidden in the water.

Technical Summary: Mining Association Rules for Targeted Spatiotemporal Aquatic Environmental DNA (eDNA) Sampling

Problem Statement
Environmental DNA (eDNA) sampling offers a non-invasive alternative to traditional, destructive methods like electrofishing for assessing species occupancy. However, the efficacy of eDNA detection is influenced by a complex interplay of abiotic factors (e.g., temperature, pH, flow rate, conductivity). While individual environmental variables have been studied, the collective correlation of these variables with site electrofishing counts and eDNA concentrations remains unclear. Furthermore, standard machine learning approaches in eDNA research predominantly rely on supervised classification, which requires large, labeled datasets that are often unavailable due to non-standardized sampling protocols and missing metadata. Additionally, existing studies often lack adequate reporting of abiotic metadata, limiting the ability to model eDNA dynamics effectively.

Methodology
This study employs unsupervised association rule mining to discover hidden relationships among covariates within a specific case study of native brook trout (Salvelinus fontinalis) in Hanlon Creek, Guelph, Ontario. The dataset consisted of 126 transactions derived from a previously published study (Nolan et al., 2023) involving simultaneous eDNA collection and electrofishing.

Data Preprocessing and Discretization: Continuous variables were discretized into binary categories ("low" and "high") using biologically meaningful thresholds derived from literature (e.g., CCME water quality guidelines) and domain-specific knowledge. Variables included pH, dissolved oxygen, air and water temperatures, conductivity, filtered volume, sampler type (Halltech OSMOS vs. Smith-Root ANDe), and electrofishing catch. eDNA concentrations were discretized based on a limit of detection (LOD) of 13.3 copies/µL.
Association Rule Mining: The Apriori algorithm was utilized to mine strong association rules from frequent itemsets. The process involved:
- Setting minimum support and confidence thresholds at $1/n$ (where $n=126$ ) to ensure every rule was supported by at least one transaction.
- Generating rules in the form $X \implies Y$ (antecedent implies consequent).
- Applying a two-step pruning strategy: first, removing redundant rules (where a subset rule has equal or higher interestingness), and second, testing for statistical significance using Fisher's exact test ( $\alpha = 0.05$ ).
Evaluation Metrics: Rules were ranked using support, confidence, and lift. Lift was prioritized to quantify the dependency between antecedents and consequents, with values $>1$ indicating positive association.
Software Development: The workflow was encapsulated in a new R package, RulesTools, which extends the arules package with functions for discretization (dtize_col, dtize_df), rule comparison (compare_rules), and visualization (rule_euler, rule_heatmap).

Key Contributions

Application of Unsupervised Learning: This paper presents a proof-of-concept for applying unsupervised association rule mining to eDNA datasets, addressing the gap in literature regarding methods that do not require large labeled training sets.
Rule Discovery: From a small dataset of 126 observations, the mining process initially identified over 12,000 plausible association rules. A strict pruning strategy reduced this to 153 statistically significant and non-redundant associations linking abiotic covariates to eDNA concentrations (low/high) and electrofishing outcomes (presence/absence).
Tool Development: The release of the RulesTools R package facilitates the application of this methodology to other ecological datasets, offering specific tools for handling missing values, discretizing data, and visualizing rule overlaps and strengths.

Results
The analysis yielded distinct patterns across four target consequents:

High eDNA Concentrations: 41 significant rules were identified. These rules showed high confidence (median 1.000) and lift (median 1.680). Notable antecedents included the use of the OSMOS sampler and high filtered volumes, consistent with established literature. However, some rules suggested high eDNA in high-temperature conditions, which the authors note may be spurious due to the small sample size.
Low eDNA Concentrations: 35 significant rules were found. These rules were slightly weaker in confidence (median 0.750) but showed higher lift (median 1.853). Interestingly, rules for low concentrations were split between sampler types and showed complex interactions involving conductivity and dissolved oxygen.
Electrofishing Presence: 14 rules were generated, predominantly consisting of single-item antecedents. This simplicity was attributed to class imbalance, as brook trout were present in 80.9% of transactions.
Electrofishing Absence: 63 rules were identified. These rules exhibited very high lift (up to 5.250) but lower confidence (median 0.600), a statistical artifact of the minority class size (19% of transactions). Some rules suggested that high eDNA concentrations could co-occur with absent electrofishing catches, potentially indicating downstream transport of eDNA or false negatives in molecular assays.

Significance and Claims
The paper positions association rule mining as a valuable, exploratory tool for eDNA research that complements traditional statistical modeling. The authors claim that this method:

Reveals Complex Interactions: Unlike standard correlation analysis, which failed to find significant isolated factors in this dataset, rule mining uncovered context-dependent, multi-dimensional relationships among environmental variables.
Is Distribution-Agnostic: The method does not require data to follow specific parametric assumptions, making it suitable for the irregular and complex data typical of ecological studies.
Enhances Interpretability: The "if-then" structure of the rules makes findings accessible to non-technical stakeholders, potentially aiding conservation decision-making.

The authors maintain a modest tone regarding the findings, explicitly acknowledging limitations such as the subjectivity of threshold selection, the risk of overfitting due to the small dataset size (126 transactions), and the fact that association does not imply causation. They emphasize that the generated rules require expert validation and that the method is best used in conjunction with other statistical approaches rather than as a standalone predictive model. The study concludes that association rule mining is an emerging methodology capable of guiding metadata collection and informing future conservation strategies, provided that results are rigorously validated against domain knowledge.

More like this