Introduction
Fish in RAS is threatened by severe risks associated with the water quality, e ven though it is designed for optimizing growth conditions. Sudden mass mortality of fish is one of the major risks in RAS. In the past years, an increasing number of such incidents has been reported and most cases have been associated with H2S1 . To-date monitoring of H2S is not commonly implemented in RAS due to limitations in analytical methods to detect low concentrations relevant for fish health. This entails development of more precise analytical chemistry-based methods, or sophisticated methods for indirect detection of the hazardous gas. It has been shown across several fields that microbial communities and their composition can be an excellent proxy for predicting the state of a given environment2. D ata obtained by next generation sequencing (NGS) from environmental samples typically contains information on thousands of microbial taxa, which allows for the employment of supervised machine learning (SML) techniques. In the recently finished "MonMic" project, we have found that microbial communities are an excellent indicator for perturbations in RAS. Building upon that knowledge, we are exploring the potential of microbial communities as indicators for H2S in RAS.
Material and methods
Atlantic salmon (Salmo salar ) post-smolt were randomized and stock ed into control and exposure groups, and acclimated for two weeks in RAS prior to the H2S introduction. The control treatment comprised of one tank and one biofilter, while the exposure treatment included two tanks connected to one biofilter. The tanks and biofilters had a capacity of 0.8 m3 and 0.350 m3, respectively. The exposure group was stocked with two different densities of fish biomass, 10 and 30 kg/m3, and the control group was stocked with 30 kg/m3 . After the acclimation period, both exposure tanks were exposed to artificially added H2S for ten consecutive days. Sodium sulfide (Na2 S) was used to artificially introduce H2S. The starting concentration of H2 S was 1,25 µg/L. Doubling concentrations of H2S from previous exposure were used for subsequent exposures until a stress response in the fish was noticed. The final exposure of 160 µg/L was performed after nine days of the challenge. Different exposures were conducted after 24 hours, enabling fish recovery from the previous exposure. H2 S concentrations were monitored in real-time employing the SeaRAS AquaSense System (AQS).
Results
There was a difference in microbial communities at the start of the experiment between the two parallels (prior to H2 S exposure) (figure 1A). Communities of both treatments were also dynamic over time. Partial least square discriminant analysis (PLS-DA) and Random Forest (RF) SML have been used to attempt to classify H2 S exposed and non- exposed samples. Both approaches yielded excellent predictions, with PLS-DA classifying at 99%, while RF at 99.5% accuracy. Most prominent features contributing to the classification strength from both approaches were filtered out and classified (figure 1B).
Discussion
Although the control and exposure RAS were identical from design, and the same water was used to fill the systems, communities between them were different (figure 1A). In fact, this is not uncomon for identical RAS, as very subtle differences in operations, water chemistry, etc. can lead to such contrasts. There was also a time component influencing community dynamics in both treatments which seemingly hinders clear conclusions to what extent H2S has influenced community alterations in exposed RAS . Nevertheless, a larger time-wise spread of community dynamics in the exposure group compared to the control is an indication that these larger alterations of microbial community may be indeed influenced by the introduction of H2 S, rather than the standard community development. App lication of supervised machine learning indicates that there is a pattern in microbial communities refelcting the presence of H2S. Indeed, many of the prominant features are belonging to the taxa associated with sulfur oxidation, e.g., Marinicella , Pseudomonas , Hydrogenophaga , Ahrensia, Hoeflea, Phycisphaeraceae , or are associated with the metabolism byproducts of sulfur oxidation ( e.g., Colwellia, Marinomonas). It has to be mentioned, however, that the experimental desing was setup to simulate acute H2S conditions, with short-lasting peakseasing off quickly. We may speculate that with a setup simulating chronic exposure, community shifts and signatures would be even more pronounced.
Conclusion
1) Two identical RAS show spatial and temporal differences in microbial communities. 2 ) Microbial community based SML model exhibits a potential for H2S prediction in experimental RAS. 3) Features contributing to SML model distinction of H2 S exposed and non-exposed samples can be associated with sulfur metabolism.
References
1. Letelier-Gordo, C. O., Aalto, S. L., Suurnäkki, S., & Pedersen, P. B. (2020). Increased sulfate availability in saline water promotes hydrogen sulfide production in fish organic waste. Aquacultural Engineering , 89, 102062.
2. Frühe, L., Cordier, T., ... & Stoeck, T. (2021). Supervised machine learning is superior to indicator value inference in monitoring the environmental impacts of salmon aquaculture using eDNA metabarcodes. Molecular Ecology, 30(13), 2988-3006.