Our first step was to reach out to the researchers who had discovered this novel organism. However, we could not get a response, no matter how desperately we tried to contact them. The genomic data of the organism were also not available online because it had not been fully sequenced yet.

Due to our lack of resources, the testing phase in this cycle was geared toward testing the validity and feasibility of our idea.

DNRA being extremely understudied meant that we did not know which promoter gene was controlling its activity, and therefore did not know how it would respond to various environmental conditions. We learnt how to operate the NIH Genbank to search for other bacteria, at least the genes that were mentioned by name in the papers by the Chinese researchers. We then scoured the ATCC (American Type Culture Collection) and MTCC to locate what new bacteria were available to us.

Gene deletion has quite a few standard procedures, like homologous recombination and conditional knockout. Our professors advised us to identify the method that would best work for our skill level and laboratory equipment. We quickly learned that knockout might not be a feasible option for us, because it was unstable and might not work as expected when dealing with an entire metabolic pathway. We had to reassess.

Since the genomes of these bacterial species were publicly available data, we rejoiced. After an appropriate amount of rejoicing, we got down to business and began to look into how we could obtain them.

We first tried and failed to get in touch with the researchers who had published the paper. Next, we scoured MTCC and ATCC to no avail. One of our secondary PIs, Professor Umesh Varshney, advised us to locate local genetic banks or microbial culture collections to see if they housed the strains.

Since we could not test whether Bacillus subtilis contained DNRA genes in the wet lab just yet, we turned to publicly available data. Our dry lab team assessed GEM models, which are networks of all the metabolites, enzymes, and corresponding genes involved within a bacterial species. Since Bacillus subtilis had a pre-established GEM model, we ran experiments to check whether it could theoretically perform DNRA, and whether it was a good idea at all to even try and procure the strain. We did not obtain conclusive evidence through these trials.

Furthermore, we extensively used the BLAST tool from Genbank to compare the genes that carried out this aerobic DNRA pathway in the Bacillus and Neobacillus species with pre-established anaerobic DNRA carried out by other microbes. To our utter surprise, the proteins produced were identical.

The rise and fall of Bacillus:

While we were initially extremely excited about these novel bacteria and the fact that Bacillus subtilis was a regular in our professor’s laboratory, we quickly stumbled upon a major trade-off. Bacillus species are notorious for forming spores. Considering that our eventual plan was environmental deployment, this would present a major biosafety challenge. While dismantling the spore-forming properties of the bacteria via gene editing would be possible, it would create another layer of uncertainty that we were ill-equipped to deal with.

The Other Genes:

While we had mostly been focused on nirBD as the genes responsible for nitrite reduction to ammonium, the nrf operon emerged during literature review. This operon existed as two variations across bacterial species: nrfABCDEFG and nrfHAIJ. For greater detail, please refer to the Design page. Moving forward, we have considered all three genes in our experimentation and further DBTL cycles.

The breakthrough, and our solution to the aerobicity problem:

Our biggest hurdle so far has been the difficulty in obtaining these novel and exotic strains, something we describe in detail in our blog post. Researchers were unresponsive, and it seemed that the elusive property of these bacteria to perform DNRA aerobically would elude us forever. However, the tests we ran by BLASTing proteins and nucleotide sequences together gave us the connecting link needed to identify the research gap in this domain.

We digitally modified our operons and added restriction enzymes to the ends for blunt-end restriction enzyme digestion. We were to receive the operons in clonal vectors from TWIST Bioscience, then digest them and ligate them into pSEVA expression vectors. We built digital maps of all our plasmids. All of the genes and vectors have been uploaded as BioBricks.

When we cross-checked the amino acid sequences of our newly-built plasmids against the original protein sequences on Genbank, we found perfect matches. We were assured that our genes should be perfectly expressed. This enabled us to order the genes from TWIST Bioscience and prepare for the cycles ahead.

While this was a technical and highly specific issue, our operons were simply too large for TWIST Bioscience to synthesize in their entirety in the same plasmid. Therefore, we needed to split each of our operons so that they could fit within the synthesizable 5kbp limit.

Professors in the Microbiology and Cell Biology Department at our institute didn’t really work with P. putida KT2440. Therefore, we were advised to demonstrate our proof-of-concept in Escherichia coli DH5-Alpha, which would be easier to work with in our context.

We soon learned that P. putida KT2440 forms biofilms. Although biofilms are generally notorious for causing disease, P.putida KT2440 is actually a class of organisms that forms Plant Growth Promoting Biofilms (PGPB). These biofilms confer a multitude of benefits to the plant roots, as detailed in the project overview.

We built the operons, courtesy of our sponsor Snapgene, and digitally inserted them into pSEVA vectors 238, 258, 648, and 2313. Having cross-checked amino acid sequences, we placed the order and received our operons from TWIST Bioscience.

TWIST Bioscience was not able to synthesize one-half of our operons due to manufacturing issues, meaning we now had two operons to work with instead of three. Therefore, we compiled protocols to isolate, digest, and ligate the operons with the pSEVA vectors before transforming E.coli DH5-Alpha.

We also built nitrite and ammonium testing assays for confirming DNRA in the transformed bacterium, and biofilm assays for confirming basal level and post-transformation biofilm-forming capabilities of P. putida.

We isolated the plasmids from the shipping bacteria and digested them with our designed enzymes, then checked the DNA concentrations via a nanodrop machine. This process contained multiple DBTL cycles within itself, but we managed to obtain working concentrations for all DNA fragments multiple times. Find the details of how we troubleshooted issues during this phase in the Experimentation page.

We performed DNA ligation to connect the expression vector to the operons and transformed E. coli. Unfortunately, we did not see growth after our first transformation on Km agar plates. However, our dry lab modelled exceptional results, as detailed in the dry lab engineering webpage.

Our base-level tests for DNRA in P.putida started with optimizing the growth medium for P. putida KT2440 by testing different broth compositions from multiple sources. We then picked cultures with the highest growth and checked for base-levels of DNRA in P. putida, a test never before carried out with this specific bacterium, even though it does contain the nirBD genes that should cause production of nitrite reductase.

Plasmid kits can contaminate DNA samples with RNA and salts. To bypass this problem, we conducted manual plasmid isolations and compared the results as seen on the experimentation page.

Minimal growth media seemed to be the best for P. putida KT2440 growth, which was an interesting revelation considering literature consistently shows that a high C/N ratio is favored for DNRA. We discuss the implications on our experimentation page.

Once we had figured out our ideas, we sought out to look for studies and literature on eutrophication, especially in the Indian subcontinent.

We looked into the literature of biofilm formation. Already existing tools and models which we could use.

We looked for GEMs of the P. putida bacteria family which may perform DNRA all in hopes of better understanding the metabolic activities of the organism in context to DNRA specifically. This also included learning how to work with the cobra.py library which is used to work with GEMs.

Over time, P. putida Y-9 seemed to lead us to dead ends. With not many resources that would verify the claims of the original papers and no responses from the authors of the papers, we decided to shift our focus to a different microbe. We were not even able to find any genomic data of this strain, nor the amino acid sequence of the relevant proteins.

We kept testing and playing with GEMs of organisms that perform DNRA, varying the nitrate and ammonium content as input flux and inhibiting specific genes to see how it would affect bacterial growth and the output flux of ammonium.

While trying to model just the large-scale eutrophication of a lake over years, we realised that this would be far too complex a task for us to reasonably implement with our skills and the time we had, as we had to account for the complex interactions between agricultural soil, runoff water, groundwater, and rivers.

While moving through the project, this model becomes redundant, as by switching strains to B. salipalidus, these species can perform DNRA regardless of oxygen concentration.

Upon trying to model the soil conditions, we once again realised that the complexity of this endeavor was very high, so we decided to abandon this.

The BLAST comparisons of the Neobacillus species were widely successful, and we were able to successfully identify the exact subunits involved in the operons using the B. subtilis genome.

These subunits had not been characterized by the research group that identified the species, but we identified the orthologous genes with very high accuracy. Upon confirmation with literature, we were able to confirm that the binding sites and active regions matched exactly with the prediction.
We also identified the presence of certain genes that are crucial in biofilm formation.

We modeled the different structures of proteins arising from the various choices of restriction enzymes on AlphaFold, and upon analysis and comparison, we realised that for our choice of restriction sites, there was negligible change in protein structure and functionality.

From our soil modelling attempts, we learned how the complexity of accurately simulating soil dynamics with root interactions and leaching was beyond our current scope and resources.

The pathway modelling attempts made us realize the pursuit is futile, given we had little existing enzymatic data and lacked proper resources to conduct our own experiments in that regard.

Through genome comparison tests, we could identify orthologous genes, confirm the active sites with literature, and discover uncharacterized biofilm-related genes.

From protein modeling results, we learned that the terminal amino acid additions from the chosen restriction enzymes did not affect the protein structure significantly. Thus, we decided to go forth with the most economical and easily available restriction enzyme for the wet lab.

We began developing the model of the genetically modified bacterium, incorporating constant enzyme expression through a constitutive promoter. This helped us simplify the system and make it more tractable.

We started planning an improved version of the GenBank framework with a more user-friendly interface to assist future users, especially those without strong computational backgrounds.

For the ML model, we began identifying publicly available soil datasets that included variables such as pH, temperature, rainfall, and nitrogen content to serve as training data.

In preparation for scaling, we conceptualized how the predictive model could be integrated with real-world soil data to generate fertilizer optimization recommendations.

Initial testing focused on simulating the metabolic pathway of the modified bacterium to verify that constant enzyme expression provided stable conversion rate predictions under different conditions.

We benchmarked the GenBank system to identify the most critical UI components needing improvement. We decided to focus on the BLAST framework.

Preliminary ML runs were planned to evaluate the model’s ability to correlate key soil parameters and predict nitrate-to-ammonium ratios.

The framework for product scaling was tested conceptually by assessing if model outputs could meaningfully inform bacterial biomass and fertilizer input ratios for different soil types.

From the theory on genetically modified bacterium circuits, we learnt that simplifying enzyme expression to a constant rate made pathway simulations much more manageable while maintaining biological relevance.

From evaluating the GenBank framework, we confirmed that improving its interface could greatly enhance usability for biologists without computational training.

Upon inputs from our senior Armaan Khetarpaul, we decided to use XGBoost models instead of neural networks, since they are less prone to overfitting.

From the product scaling concept, we learnt that the combination of pathway modelling and soil condition predictions could directly guide practical applications, improving both efficiency and sustainability.

We used an XGBoost Regressor with hyperparameters tuned through a previous search, with our chosen configuration emphasizing balanced learning with moderate regularization. We applied a smearing estimate (a standard bias correction technique) when converting predictions back to the original scale. To generate the synthetic data, we used our ODE model and literature to predict by what percentage the rate of DNRA will increase in the soil.

We built the model on COPASI software, drawing inspiration from the previous iGEM team Cattlelyst from Wageningen 2021, to construct the ODEs and obtain some of the rate constants. We obtained the other values from previous characterisations of the enzymes from the literature. We later expanded the model to include nitrate leaching and N uptake by plants at a constant rate.

We created a preliminary framework in Python, learnt to bypass the use of Pio-python.

We evaluated the model on the test set using standard regression metrics on the original ammonium scale. The results obtained were pretty good considering the small data size, explaining almost 60% of the variation in ammonium levels, meaning that the model is picking up on many important patterns.

After going through the literature and simulating conditions on our ODE model, and running the ML model on some different conditions, we finally settled on a percentage increase value obtained by comparing the maximum conversion rates in our model and some rates from the literature and the previous iGEM team Wageningen.

We had several testing phases by changing constant values and observing conversion rates, and noticed that there were two issues initially: unrealistically high cell growth and accumulation of ammonium in the cellular component. We fixed these issues by capping the cell growth at a fixed value and by restricting ammonium import beyond a certain concentration of cellular ammonium.

We also varied the enzyme concentration across various values and observed the cellular nitrate values, which were all within permissible values and would not cause toxicity. While varying the rate of enzyme concentration, we also observed a wide range of conversion efficiencies from about 10% to 98%, and at very high concentrations (>M) of enzyme, the model broke down.

Upon including the leaching and uptake, and varying the values and observing efficiency,y and by comparing with real-world values, we decided that the most sensible approximation for the rate would be 10^-11mol/L. We also switched around enzyme concentrations after this inclusion and observed that in some ranges the change in efficiency is drastic, but in others there was barely any change.

For testing, we presented our solutions to said-tool users for their feedback and insights on how to improve on it. We optimized the code in JavaScript and built a GitHub website to host our tool.

We realised that the dataset being quite small, with only 902 samples, likely limits the model’s performance. More data would help it learn more nuanced patterns and improve accuracy further. We believe that future iGEM teams should focus on expanding the dataset with more diverse soil samples and conditions. This would allow for training more robust models that can better capture the complexity of soil ammonium dynamics.

From the huge amount of testing done, we figured out the limits of the model and what are the realistic ranges of enzyme concentration, cell volume, uptake, and leaching value, which are vital for being able to predict optimal conditions for given conditions of soil and plant type. This testing helped us understand the potential utility of our model in designing Wet Lab experiments, by constraining certain factors such as the ratio of nitrate: nitrite reductase to ensure no nitrite build-up. It also helps us better characterize the functionality of the modified bacterium so that we at least know a ballpark estimate of what the expected results should be.

We hope the tool we developed can be of use to people in the field in some way. We were able to learn a lot of new things about how things actually work behind the simple webpage that we see. But most importantly, the development of GenBankUI was a fun way to interact with our juniors, who helped us with the development of this tool! We could familiarize them with the world of synbio and iGEM.