A Regression-Based Approach for Accurate Source Tracking Using Microbial Communities

Introduction

In the past decade, significant advancements have been made in the field of microbial source tracking (MST), enabling the identification of the origins of microbial contamination. MST techniques have found diverse applications, ranging from assessing water contamination to determining source contributions in legal arena [1]. These techniques are also valuable for quantifying the distribution of source contribution, and they have been widely employed in areas such as microbial forensics, microbiome-based healthcare, and environmental monitoring. By analyzing multiple microbial samples and decomposing the microbial community, it becomes possible to establish links between the sources/individuals and the evidence/sink, while also estimating the respective proportions of the contributors. Throughout the text, we will use the terms of source and evidence.

Microbial communities exhibit substantial variation from sample to sample. It is often feasible to track the source of a sample via its unique microbial composition. However, in practical scenarios, the evidence samples typically comprise a mixture of multiple sources, and their composition can fluctuate under varying environmental conditions. This poses challenges in efficiently tracing the sources. A recent study by Gloor, et al. [2] highlighted the discordance among three mouse metagenomics studies, as they reported distinct lists of statistically significant taxa with minimal overlap [2].

A few methods have been developed for microbial source tracking. SourceTracker, a tool for microbial community analysis, estimates microbial proportions using Bayesian models, providing valuable insights into complex ecosystems [3]. This method allows uncertainty in the source and is able to track unknown sources. However, its accuracy relies on well-characterized source environments in the reference database, and it assumes well-mixed communities, limiting applicability in heterogeneous environments.

Additionally, the computational time of SourceTracker can be a consideration, especially for large datasets. FEAST (Fast Expectation-Maximization for microbial source tracking), is a method adept at estimating the contribution of numerous potential sources in microbial communities, especially when dealing with unknown or poorly understood sources within the target microbial community [4]. While it excels in source identification and quantification, it is important to note some potential disadvantages. FEAST may face challenges in scenarios where the underlying assumption of convex combinations is violated, such as in cases with significant variability between certain source environments. Additionally, the method’s applicability could be influenced by technical confounders in specific datasets.

The RAD method based on Aitchison distance emerges as a new tool in microbial source tracking, particularly in forensic contexts [5]. Its notable advantages include accurate detection of sources with non- zero mixture components, with infrequent selection of non-significant sources in replications. Integrating RAD into parameter estimation significantly enhances the precision of predicted source proportions for Bayesian SourceTracker, leading to reduced absolute and relative error rates. Additionally, its unique feature of not assigning proportions to an unknown source, while advantageous in certain scenarios, may limit its ability to account for certain complexities in microbial communities. Nonetheless, in the realm of forensic investigations, RAD proves invaluable for its ability to filter sources and estimate microbial proportions, enhancing trace evidence analysis and aiding in crime scene investigation.

In this paper, we introduce a novel regression-based method called REST (REgression-based Source Tracking) that has wide-ranging applications in forensic studies and environmental monitoring. The primary goal of our method is to accurately identify sources from a pool of sources using microbial materials collected at a crime scene. Additionally, we apply REST to trace the compositional pattern of water samples in a real data analysis. However, it is important to note that our method is not limited to these specific applications and can be extended to various scenarios, including general source tracking such as detecting fecal contamination in environmental samples.

By employing the REST method, we aim to enhance the accuracy and reliability of source tracking in forensic and environmental studies. The ability to identify individuals and trace the origins of microbial compositions in various samples opens up possibilities for improving criminal investigations, understanding environmental contamination, and addressing public health concerns.

Methods

REST operates by estimating the probability of sources’ presence in a collected sample or determining the percentage contributions of different sources in a given evidence sample. The method begins by selecting important contributors as a preliminary step. It then identifies which sources are responsible for the microbial composition observed in the evidence sample. In cases where there are potential “unknown” sources, REST constructs a profile for these sources and subsequently conducts an agreement check to evaluate their plausibility. Finally, REST estimates the proportions of each source contributing to the evidence sample.

The REST framework consists of four steps (Figure 1). In step 1, Relative Aitchison Difference source selection (RAD) [5] is employed to identify sources that contribute to the microbial evidence. The identified important contributors from step 1 are utilized in step 2 to construct linear regression models, and the residuals from the models are transformed to create a profile for potential “unknown” source. In step 3, an agreement analysis is conducted to examine the status of the “unknown” source, i.e., whether there exists unknown source or not. Finally, the proportions of the selected sources which are included in the regression model, and the potential unknown source if it exists based on step 3 are estimated. The details of the method can be seen in the Appendix. We compare the proposed method REST with the existing methods, FEAST, SourceTracker, and RAD using both synthetic data and real data analyses.

Click to enlarge

Discussion and Conclusion

The use of human and environmental microbiome data in source tracking has seen significant development. However, challenges remain, including storage constraints, contamination risks, and the lack of robust datasets. Urgently needed are accurate statistical methods with strong power for forensic, healthcare, and environmental applications. To achieve widespread use, we must establish appropriate sample storage, reliable datasets, and resilient statistical approaches.

In this paper, we propose the REST method for accurately identifying and estimating sample probabilities within a pool of sources using microbial materials from a specific location. Our approach involves contributor selection, unknown source profiling, an agreement check, and proportion estimation. The contributor selection step narrows down the range, while the agreement check confirms the presence of a missing source. Our method adeptly identifies contributors, excluding unrelated sources and reducing noise. Recognizing the possibility of missing information, we construct a profile for potential missing data, using the ICC for the agreement check. A small ICC value indicates no “unknown” source, streamlining proportion estimates by reducing noise.

REST has demonstrated superior performance compared to existing microbial source methods in both simulation studies and real data analysis. Its applications span various fields, including forensic studies, contamination detection, and human identification. One limitation of the proposed method is the requirement for two or more replications of sequencing evidence for an agreement check. In practical terms, collecting two replicates is reasonable and affordable. In cases where evidence collection occurs only once, especially when the touch is imprecise, technical repeats can mitigate noise. If no replicates are available, we recommend always including the “unknown” source for proportion estimation, ensuring minimal noise in scenarios with missing evidence. In addition to accuracy, computational time is a crucial consideration. REST falls between FEAST and SourceTracker in terms of computational efficiency.

To enhance the accuracy of the REST method, we should consider the characteristics of the data. Microbiome data are typically high-dimensional and sparse, and in this case, they are compositional data constrained to sum to a constant. Analyzing compositional data is best approached through a log-ratio transformation [10, 11, 12]. We experimented with log-transformation on the original data, incorporating a small pseudo-count for zero counts, and an isometric log- ratio transformation. The latter, an orthonormal isometry, addresses challenges associated with centered log-ratio transformation in constructing a potential “unknown” source. Data contamination poses another consideration. The presence of DNA contamination is common and can impact accuracy during analysis. For future work, integrating microbiome and metabolomics data holds promise for improving accuracy in linking objects or locations with sink/ evidence samples. Conflicts of Interest: None declared.

Acknowledgements: We heartedly thank Dr. Sadowsky’s lab for the fecal data. CODE AVAILABLE: https://github.com/anlingUA/REST Funding: This work has been partially supported by the United States Department of Agriculture (ARZT- 1361620- H22-149) to L.A.