Overview

The goal of our project was to create a system of bacterial artificial intelligence. But to reach this goal, we had to determine smaller steps that would lead us to make the project feasible.

The first step was to characterize 11 RBS that we chose after the predictions characterization made by our Dry Lab, using the De Novo DNA software. This would help us ascertain whether all of them work and which ones are the most suitable for our system.

Determining the Weights

As mentioned in the Description page, the weights of our bacterial artificial system are determined by different Ribosome Binding Sites (RBS). Each RBS corresponds to an alternative translation rate, thus enabling the assignment of different weights to an input. This means that different sender subpopulations produce different amounts of acyl-homoserine synthase when encountering the input, namely aTc.

Our team performed Golden Gate Assembly (level 1 and level 2) in order to make the constructs with the 11 RBS variants (Table 1) (see see Design). Unfortunately, our first try did not yield functional constructs and they had to be reassembled through several engineering cycles as described in the Engineering Success tab.

Our first constructs, which included the overlapping version of the luxI gene and the mNeonGreen gene (design based on TeRe mechanism), did not produce any results as assessed both by induction experiments with multiple inducer concentrations (aTc) and in different bacterial strains (Figure 2). We also assessed the production of acyl-homoserine synthase and mNeonGreen with SDS-PAGE analysis and Coomassie staining (Figure 1) and in-gel fluorescence of cell lysates in two cell types BL21 (DE3) (level 2 construct) and DH5a-z1 (level 1 construct) after a 3h induction with aTc, centrifugation and sonication of the cells. The in-gel fluorescence did not produce any results except for the positive control (OL colE1 induced). The SDS-PAGE confirmed that our constructs did not work in BL21 cells (or the small quantity of protein was lost in the transfer step), but worked in a leaky way in DH5a-z1. This was the first sign that made us realize there was a problem with the chosen promoter for the TetR repressor, as DH5a-z1 cells have the repressor gene integrated in their chromosomal DNA.

Figure 1: SDS-PAGE: OL colEI for positive control for mNeonGreen. S4-DH5a-z1 cells produce LuxI but not mNeon Green both at induced and uninduced state. S4-BL21 cells do no produce LuxI or mNeonGreen with or without induction.

Measurements for mNeonGreen fluorescent protein were performed with excitation wavelength 476nm and emission wavelength 547nm, as recommended by the plate reader manufacturer for proteins whose emission and excitation wavelength overlap. According to the fluorescent protein database, mNeonGreen has an excitation wavelength at 506nm and an emission wavelength at 517nm. It is also important to note that, according to Tutol et al. (2019) [1], excessive presence of chloride ions can have a negative effect on mNeonGreen; that is why we tried changing the buffer where we resuspended the pellets before measuring them, in case this was affecting our measurements (Figure 3).

Figure 2: Testing senders pTU1-luxI/mNeonGreen with multiple aTc concentrations (2.16μΜ,0.4μΜ,0.09μΜ). This induction resulted in very low fluorescence signal, maximum 340 RFU (ratio) 20h post induction.

Figure 3: Testing the effect of chloride ions on mNeonGreen fluorescence; there was no significant effect when changing the solvent for pellet resuspension.

The resuspension of the cells to a chloride-free buffer did not change the negative results.

In a last attempt to pinpoint the problem with our constructs, we wanted once again to test if the luxI gene is expressed in the luxI/mNeonGreen construct. That is why we tested the effect of the supernatant of BL21 luxI/mNeonGreen S4 RBS cells after 3h induction with aTc on the receivers construct OL-colE1 (BL21). This experiment proved that the luxI gene is functional and can produce OC6 to activate the receivers and made us realize the problem is also located in the TeRe-based design with mNeonGreen (Figure 4). Although LuxI was expressed in BL21 cells, it is not detected in SDS-PAGE (Figure 1) and that can be attributed to very low production of this protein.Taking into account that the Coomassie staining happened after this polyacrylamide gel was used in transfer for Western blotting for some other samples, the remaining quantity was even smaller. That is maybe why the receivers were only slightly induced by the supernatant of BL21 senders. The same construct has been proven to be highly induced when using externally provided OC6 (see Receivers Results) and also compared to the final sender constructs (luxI/sfGFP, see Proof of Concept). For all these reasons we considered these constructs not functional and proceeded in creating new constructs.

Figure 4: We tested the effect of the supernatant of luxI/mNeonGreen S4 RBS cells after 3h induction with aTc on the receivers' construct OL-colE1 (BL21). This experiment proved that the luxI gene is functional and can produce OC6 to activate the receivers and made us realize the problem is the TeRe-inspired design with mNeonGreen.

As described in the Engineering Success page, we then proceeded with creating a new construct which included the first 36 nucleotides of the luxI gene followed by the sfGFP gene in order to estimate the translation rate displayed by the different RBS variants taking into account that RBS sequences are context dependent. We tested this construct (level 2) for 11 different RBS in BL21(DE3) cells but there was no sign of induction (Figure 5). We present an example for this induction experiment for BCD2. We also tried to change the promoter which constitutively controlled the expression of the TetR repressor, but again we did not get any positive results (See Engineering Success) (Figure 6).

Figure 5: Induction experiment of BL21 luxI 36nt -sfGFP. Example for BCD2. The cells produced a stable amount of fluorescence expressed in RFU that did not increase significantly after induction.

Figure 6: Induction experiment of DH5a cells with pLacI promoter for the expression of TetR. The change of promoter to a previously characterized one for the same system did not yield the expected results.

So as described in the Engineering Success page, we proceeded with transforming the level 1 constructs in E.coli DH5a-z1, which have the tetR gene integrated in their genome under a promoter that was proved to work for our constructs. The DH5a-z1 cells gave us the opportunity to fully characterize the 11 RBS and the results are presented below.

RBS Characterisation

Abbreviation	Full Name	Part	Strenght
S4	Synthetic RBS Ath4	BBa_K4294004	Strong
S5	Synthetic RBS Ath5	BBa_K4294005	Strong
S10	Synthetic RBS Ath10	BBa_K4294010	Medium
S11	Synthetic RBS Ath11	BBa_K4294011	Medium
S12	Synthetic RBS Ath12	BBa_K4294012	Weak
S13	Synthetic RBS Ath13	BBa_K4294013	Weak
BCD1	BCD1	BBa_J428034	Weak
BCD2	BCD2	BBa_K4294102	Very Strong
BCD8	BCD8	BBa_K2680561	Weak
BCD12	BCD12	BBa_K2680529	Weak
BCD14	BCD14	BBa_K4294114	Weak

Table 1: All the Ribosome Binding Sites (RBS) characterized in our project classified by their strength as indicated by De Novo DNA (RBS calculator). The BCDs were also more accurately predicted by Mutalik, V. et al. [2] for our project but are not represented in this table

As described in the Protocols, we used the FlexStation3 (Molecular Devices) Plate Reader to conduct the fluorescence and the OD600nm measurements. The exact Protocol for the induction is described in the Protocols page. We present below the diagrams with the results for the luxI-sfGFP constructs for the 11 RBS in different time frames (0-240min) from the initiation of the induction and in different inducer concentrations (2-5,12*10^(-6)μΜ) (Table 2). Last but not least, we compare the RBS for the difference in sfGFP production with the luxI context dependence (Figures 4,5,6). All the measurements of fluorescence for sfGFP were performed with gain 300, excitation wavelength 485nm, emission wavelength 510nm. Regarding the OD600nm, all measurements were recorded without using the Pathcheck option to fulfil the needs of our project.

Table 2: Our results for S4,S5,S10,S11,S12,S13,BCD2,BCD8,BCD12,BCD14 in different concentrations at different time points, the concentration range used for aTc was 2μM to 5,12 *10^(-6) performing 5-fold dilutions, the maximum induction time was 6 hours, here we present data until 4 hours, further induction did not result in increase of sfGFP production. BCD1 is not represented here because it did not produce significant induction after repeating the same experiment two times.

Figure 7: Comparison of all Synthetic RBS as characterised by our measurements.

Figure 8: Comparison of BCDs as characterised by our measurements.

Figure 9: (a) Comparison of all 11 RBS as characterised by our measurements (b) Comparison of the weaker RBS that can not be best represented in figure 6(a).

From Table 2 and Figures 4,5,6 we can conclude that the characterised RBS can be sorted by strength as follows:

BCD2 > S11 > S13 > S10 > BCD12 > S4 > BCD14 > S12 >BCD8 > S5 > BCD1

These results differ significantly from our predictions. This can be attributed to the change of strain from our initial BL21 (DE3), for which the Dry Lab made the calculations, to DH5a-z1 in order for our system to work. It is also important to note that, as we learned during meetings that are recorded in our Integrated Human Practices page, both Dr. Skretas and Prof. Di Ventura mentioned that RBS are a good way to calculate weights but RBS calculators are not always reliable and should always be confirmed with experimental procedures. Τhat is why our team tested 11 RBS sequences instead of 3 that were needed for our project.

In this set of experiments we achieved the characterization of 11 RBS in the context dependence of luxI. This is a library that future teams may use in order to incorporate this gene to their contstracts and control the luxI translation in a way that fits the needs of their project.

We then proceeded to a second series of experiments for our constructs. This time we used the luxI only constructs. Since the luxI gene expression does not result in any visual output and since we did not have the resources to perform either affinity chromatography or Western blotting (even though we had included a 6*His-tag in our constructs), we chose an indirect way to quantify its production.

These experiments aimed to characterize the activation of the receivers caused by the different RBS. As mentioned in the Protocols, we used the supernatant of each one of the senders after 3h induction with multiple concentrations of aTc for the induction of the receivers (OL colE1 construct) and quantified the production of mNeonGreen produced when induced with different supernatants. The results are presented in Table 3.

Table 3: Our results for use of supernatants of S4,S5,S10,S11,S12,S13,BCD1,BCD2,BCD8,BCD12,BCD14 senders which were induced for 3h, the concentration range used for aTc was 2μM to 5,12 *10^(-6) performing 5-fold dilutions, the maximum induction time for receivers was 6 hours, here we present data until 4 hours, further induction did not result in increase of mNeonGreen production.

Figure 10: Comparison of all Synthetic RBS for the differences in activating the Receivers population.

Figure 11: Comparison of all BCDs for the differences in activating the Receivers population.

The results are not similar to the characterization of the RBS with sfGFP for many reasons. We tried to name some reasons why this difference is observed. First of all, these results concern the activation of the Receivers. The activation rate depends on the amount of OC6 produced by the Senders, which is not known, and it also depends on the concentration of Sender cells at the 3 hours time point, when the induction happened. Our team did not take into account the sender cells population density after the 3 hour induction, therefore the differences encountered between our characterization and the measurement of the luxI-sfGFP constructs can not be considered to be the same, because the results caused by the supernatants could be attributed to differences in the senders population density and not the RBS strength. Also, the induction time of the senders may not have been the appropriate one and more experiments are needed to optimize that. We also came to the conclusion that a different ratio of Senders:Receivers is needed. After sharing our results with the Dry lab, they informed us that the optimal ratio is 1:10 or 1:7. We incorporated this knowledge into the Proof of Concept experiments. Taking into account the above, we present these results to prove the capability of the senders to activate the receivers in a differentiated manner and to point out the importance of the appropriate balance between the two populations. We elaborate more on that in the Proof of Concept page.

Senders
Weights
RBS Characterisation
Receivers
Conclusions
References

As we described in our Design page, we wanted to test different constructs for the receivers’ circuit in order to better represent the binary form of the output of the Perceptron algorithm. When the intermediate molecule exceeds the threshold, then the output is produced. We designed our system that way to achieve a steep activation that better simulates the function of the algorithm (Output:1 or Output:0), and also achieve the reduction of leakiness.

The constructs we tried to test are presented in the table below:

Construct	Plasmid	BioBrick
Open Loop (OL)	pTU2A-RFP p15 or.	BBa_K4294801
Open Loop (OL)	pTU2A-RFP colE1 or.	BBa_K4294801
Positive Feedback (PF)	pTU2A-RFP colE1 or.	BBa_K4294802
Positive Feedback constitutive (PFc)	pTU2A-RFP colE1 or.	BBa_K4294803
Positive Feedback represor (PFr)	pTU2A-RFP colE1 or.	BBa_K4294804
Positive Feedback constitutiveand represor (PFc+r)	pTU2A-RFP colE1 or.	BBa_K4294805

We used different concentrations of externally provided OC6 to induce the Receivers separately from the Senders in order to characterise their activation.

After some preliminary experiments we decided to measure the Fluorescence without using the pathcheck option at the plate reader, as we realised that after some hours, the OD could not be measured with Pathcheck, due to cells' exponential growth. Also, our measurements were taken until the time point of 4h, because we observed that at approximately 3h fluorescence reaches its plateau and dying cells start to precipitate at the bottom of the plate, affecting the OD measurement.

Measurements for mNeonGreen Protein were performed with excitation wavelength 476nm and emission wavelength 547nm, as recommended by the plate reader manufacturer for proteins that the emission and excitation wavelength overlap, according to the fluorescent protein database mNeonGreen has an excitation wavelength at 506nm and an emission wavelength at 517nm.The protocol we followed for the induction with OC6 can be found in the protocols page.

We used different concentrations of OC6 for each construct according to the optimization experiments. We measured absorbance at 600 nm and fluorescence as described in our Protocols book. We then calculated the ratio Fluorescnce/OD600nm. The data demonstrating the ratio increasing are presented below in different concentrations and in the scale of time.

Figure 12: Induction of BL21 OL pTU2-RFP p15A or. with the following OC6 concentrations (μΜ): 100, 10, 1, 0.1, 0.01, 0.001, 0.0001. The uninduced cells are represented by the concentration value 0.0000001 for the diagram purposes.

It is observed that the fluorescence ratio increases as the concentration increases. But after 0.01 μΜ the ratio reaches its highest value.

In Figure 13 we present the results of the induction of the same construct, but using the other pTU2-RFP origin plasmid (coLE1).

Figure 13: Induction of BL21 OpLo pTU2-RFP colE1 or. with the following OC6 concentrations (μΜ): 100, 10, 1, 0.1, 0.01, 0.001, 0.0001. The uninduced cells are represented by the concentration value 0.0000001 for the diagram purposes.

Then, in order to achieve steeper activation, we tested the PF construct, but as it can be seen at Figure 14, the results showed no significant activation of the receivers.

Figure 14: Induction of BL21 PF pTU2-RFP colE1 or. with the following OC6 concentrations (μΜ): 0.062, 0.031, 0.0155, 0.003875, 0.0019375, 0.00096875, 0.000484375, 0.000242188, 0.000121094. The uninduced cells are represented by the concentration value 0.00001 for the diagram purposes.

Figure 15: Induction of BL21 PFc pTU2-RFP colE1 or. with the following OC6 concentrations (μΜ): 10, 10, 1, 0.1, 0.01, 0.001, 0.0001. The uninduced cells are represented by the concentration value 0.000001 for the diagram purposes.

Once again PFc did not present higher and steeper activation than the OL.

Figure 16: Induction of BL21 PFr pTU2-RFP colE1 or. with the following OC6 concentrations (μΜ):100, 20, 4, 0.16, 0.0064, 0.00128, 0.000256, 0.00001024. The uninduced cells are represented by the concentration value 0.000001 for the diagram purposes.

PFr presented a steeper activation in comparison with OL, as it can be seen in Figure 12. By designing and testing all the different constructs we managed to accomplish the goal of a steeper activation.

Finally, the last construct PFc+r did not work either, as presented in Figure 17.

Figure 17: Induction of BL21 PFc+r pTU2-RFP colE1 or. with the following OC6 concentrations (μΜ): 10, 2, 0.4, 0.08, 0.016, 0.0032, 0.00064, 0.000128, 0.0000256, 0.00000512. The uninduced cells are represented by the concentration value 0.0000001 for the diagram purposes.

Conclusions

“Perspectives” aims to create a functional bacterial artificial intelligence tool. A tool that leads the receivers populations making a decision according to their surroundings. In order to make a functional communication system, there were several obstacles that had to be surpassed. First we had to choose a way that the weights - analogues, would be represented, our team chose for this cause RBS sequences in order for “Perspectives” to be a versatile, “plug in” system, for future use. Choosing the right RBS sequences was challenging, because although there are helpful online tools that determine an RBS' strength, they do not always reflect the experimental results. That is why our team chose to experimentally characterize 11 RBS variants, instead of just the three we needed. The RBS were chosen after using RBS calculators. We were called to do a lot of troubleshooting in order to make our system work, but we finally achieved a very thorough characterization of 11 RBS variants, which consist of different Senders populations that can be used to fulfil the needs of future implementations. Our characterization also describes the behaviour of different RBS variants in the context of the luxI gene.

Our future goals regarding the senders include introducing different inputs of different natures such as chemicals, light or temperature. We also aim to include negative weights to our project by including inputs that lead to the expression of a quorum quenching enzyme in order to increase safety and also be able to introduce factors that “inform” our system that there is a factor that contradicts the activation of the receivers.

The second component of “Perspectives”, the receivers, had also to be carefully designed and also had some challenging aspects. As our project is inspired by computer science, in bioinformatics language, we wanted the receivers population to be able to take a binary decision, a decision in the form of “0” or “1”. This means that the receivers population as mentioned in our results but also in the Design page had to be steeply activated, that is why we tested multiple constructs, some simple and some more complex. From the constructs we tested as indicated by the figures above, the most functional but without steep activation function were the open loops (ColE1) that we also used for the interaction between the senders and the receivers. But the one that seems extremely promising is the PFR construct that is presented with a very steep slope.

Our future goals include testing the system as a whole with the PFR construct to optimise its function.

References

1. Tutol, J. N., Kam, H. C., & Dodani, S. C. (2019). Identification of mNeonGreen as a pH-Dependent, Turn-On Fluorescent Protein Sensor for Chloride. Chembiochem : a European journal of chemical biology, 20(14), 1759–1765. https://doi.org/10.1002/cbic.201900147 2. Mutalik, V. K., Guimaraes, J. C., Cambray, G., Lam, C., Christoffersen, M. J., Mai, Q. A., Tran, A. B., Paull, M., Keasling, J. D., Arkin, A. P., & Endy, D. (2013). Precise and reliable gene expression via standard transcription and translation initiation elements. Nature methods, 10(4), 354–360. https://doi.org/10.1038/nmeth.2404