In the process of our project development, we performed many engineering cycles to improve our design and experimental implementation. Here we present our engineering progress in three sections: engineering of two-component system, engineering of logic gate, and engineering of software. For each engineering process, we record it following the four general stages (design, build, test, and learn) to demonstrate our improvement.

Signal amplification

Cycle 1: Weak fluorescence signal

Design

At first, we used EGFP as the reporter gene. It was placed under the promoter of the two-component system, so that if the two-component system worked, we could see the green fluorescence.

Build

The cloning of EGFP under the promoter of the Nisin two-component system was successful. Then it was transformed into the E. coli BL21 strain. Nisin was added to induce the expression of EGFP.

Test

Unfortunately, results showed that the signal was too weak to be seen even under the fluorescence microscopy.

Learn

We speculated that the promoter of the two-component system was so weak that it was almost impossible to observe the expression of downstream EGFP by naked eyes. Nevertheless, our goal was that the signal produced by the engineered E. coli could be readily observed at home. Therefore, we designed three ways to amplify our signal.

Cycle 2: Three ways of signal amplification

Design

The first way of signal amplification was to use amplifiers. We intended to empower the polymerase system without laying too much metabolic pressure on our engineered E. coli. Therefore, we chose to introduce the σ of T3 polymerase and the core of the T7 polymerase. This combination ensured higher level of transcription while limiting the toxic effect to a manageable extent.

The second way of signal amplification was to enhance the promoter strength of the two-component system. We designed a software to predict the strength of mutant promoters. Related details are described on Software page.

The third way of signal amplification was to use a sensitive and readily-discerned reporter gene. We chose lacZ in place of EGFP, considering that a small amount of lacZ expression can lead to evident change of color.

Build

To implement the first plan, we successfully introduced the σ of T3 polymerase and the core of the T7 polymerase into the E. coli BL21 strain.

To implement the second plan, we successfully designed a software to predict the strength of mutant promoters.

To implement the third plan, we successfully replaced EGFP with lacZ as the reporter gene.

Test

To test our first solution, we demonstrated the efficiency of the amplifier by fluorescence spectrophotometry.

To test our second solution, we selected the top10 strongest promoters according to our software prediction and tested their promoter intensity respectively. We demonstrated that our software possessed a 90% accuracy in predicting promoter strength.

To test our third solution, the induction result of the two-component system after replacing EGFP with lacZ could be easily detected by observing the blue color.

Learn

Firstly, we demonstrated that the amplifier could fit in our gene circuit to amplify signals efficiently.

Secondly, we learned that we could use the mutant promoters with more enhanced strength to improve our product.

Thirdly, we found out that LacZ is suitable for signal representation for household devices. It can magnify faint signals, and the depth of color can also represent the strength of signals.

Cycle3: lacZ knock out

Design

Since we intended to select lacZ as a reporter gene downstream of the two-component system, the presence of lacZ in the genome of the strain BL21(DE3) would interfere with our design. Therefore, we attempted to knock out the lacZ gene in the BL21(DE3).

Build

We applied the CRISPR-Cas system to perform genomic knockouts. Specifically, we chose the CRISPR/AID system, a tool capable of converting cytosines to thymine (C-T) at 18-21 bp upstream of the NG/NGG-PAM sequence, to complete the mutation. We found six CAA/ CAG/ CGA codons in the lacZ gene of the BL21(DE3) genome, and four of these codon classes have NGG sequences at 18-21 bp upstream, which can be used as sites for designing gRNAs. We cloned these four gRNAs into the scaffold of gRNA in the gRNA plasmid by Hifi. Sequencing results verified that the cloning of gRNA-2 and gRNA-4 were successfully constructed.

Test

We co-transformed the successfully constructed gRNA plasmid with dCas9 plasmid into BL21(DE3) and coated the plate on petri dish containing X-gal for screening. The BL21(DE3) that was not transformed was also coated as a positive control.

Learn

After one day of incubation, the colonies growing on the petri dish were all blue, similar to the plates of the positive control, indicating that our LacZ knockdown failed. The reason may be that the gRNA we used was not efficient and caused off-target. We should try more gRNAs for knockdown at the same time.

Response Time Reduction

Cycle1: Constitute Promoter

Design

Initially, nisK and nisR were placed under the T7 promoter and induced by IPTG. However, it took time for nisK and nisR to be expressed and embedded in the membrane, only after which the induction of the two-component system was possible. Consequently, there would be a lag between the adding of Nisin and the representation of signals. Therefore, we decided to replace the T7 promoter with a constitute promoter, rendering advanced expression of nisK and nisR.

Build

We successfully replaced the T7 promoter with a constitute promoter.

Test

The constitute promoter allowed for advanced expression of nisK and nisR, so that the induction of the two-component system could take place immediately after adding Nisin.

Learn

Replacing inducible promoter with constitute promoter can shorten the response time of Nisin induction, thus reducing the response time of the whole detection process of sperm quality examination.

Cycle2: Simultaneous induction

Design

Initially, we added IPTG first to induce the expression of nisK and nisR, then Nisin to induce the two-component system pathway. However, it took a long time to separate the induction into two stages. Therefore, we intended to add IPTG and Nisin together so that the experiment period could be shortened.

Build

We added IPTG and Nisin together to induce the expression of nisK/nisR and the two-component system pathway simultaneously.

Test

The result of simultaneous induction was basically the same with separate induction, while the former took shorter time.

Learn

The induction of the expression of on-membrane receptors of the two-component system and the induction of the two-component system pathway can be conducted simultaneously, reducing the experiment time without influencing the effect.

Affi-PmrB with more amino acids retained

Design

Our previous modification of PmrB was to accurately replace its extracellular domain with the double-α-helix affibody recognizing Fc fragment. However, though Affi-PmrB/A can activate the expression of downstream EGFP under IgG induction, the result showed by microplate reader did not reveal an expected kinetic feature of Affi-PmrB’s function as a receptor. The result of molecular docking showed that the domain of Affi-PmrB interacting with Fc fragment overlapped with its transmembrane domain by 1-2 turns of α-helix, so we inferred that this might be the reason why Fc did not have a proper interaction with our Affi-PmrB.

Build

To solve this problem, we designed a new Affi-PmrB variant with more amino acids from PmrB retained. The amino acid 34-37 and 62-64 of PmrB, which exactly flank the transmembrane domain at the outer-membrane compartment, were retained to separate the recognition domain and the transmembrane domain.

Test

The prediction made by AlphaFold2 showed that the new variant also had a structure similar to PmrB. Meanwhile, the recognition and the transmembrane domain separated well according to result of molecular docking.The new Affi-PmrB variant with more amino acids was constructed into Affi-PmrB/A system and induced with IgG. Then we applied fluorescence spectrophotometry to measure the fluorescence intensity by microplate reader. Our results confirmed that the fluorescence intensity was much higher in the experimental group with antibody induction.This affi-pmrB variant seems improve the affinity between receptor and antibody.

Learn

Through this cycle, we have a deeper understanding of the relationship between the structure and function of affi-pmrB receptor.

Two in one plasmid construction

Cycle 1

Design

At first, we planned to construct two plasmids, one containing serine integrase and the other containing verification sequences, using pET28a as the backbone with different antibiotic resistance.

Build

By reducing the concentration of two antibiotics to half of its working concentration, we obtained enough bacteria containing the two plasmids from co-transformation.

Test

We induced serine integrase at different temperatures for different time. Sequencing results indicated that the sequence between attP and attB had not been inverted in both induced and uninduced bacteria solution. In contrast, colony PCR using a pair of primers which can only anneal to inverted sequence suggested that all the bacteria contained inverted sequence. We speculated that the reasons behind such inconsistency might be:

• Leakage from the lacZ operator resulted in inverted sequence in uninduced bacteria.
• The expression level of serine integrase might be too low for inverted sequence to be detected by Sanger sequencing.

To find out the differences between the induced and uninduced bacteria, we conducted qPCR using the primers which can only anneal to inverted sequence. We found that for induced bacteria, less Ct was needed (about 3 Ct) for reaching the threshold, indicating that our induction was successful.

Learn

We learned that a promoter with less leakage would be more suitable for our system. Moreover, we should use a more sensitive method (for example, qPCR) to detect the effect of serine integrase.

Cycle 2

Design

We changed the backbone for serine integrase to pBad24 to reduce leakage. Then we measured the inversion rate of integrase by qPCR and fluorescence spectrophotometry after co-transformation.

Build

We cloned two kinds of serine integrase in pBad24, one was the same as the one in cycle 1 and the other was a gift from Fang Ba. We conducted experiments using the orignal intergase first.

Test

Uncertain about whether the tiny difference in fluoresence intensity could be detected, we conducted qPCR first. However, no stable results that matched our expectation could be obtained. Gel electrophoresis of qPCR products showed nothing, and SDS-PAGE of bacteria solution indicated no difference between induced and uninduced ones. After consulting with our advisor Xiaofei Ge, we proposed three possible reasons to explain our failure:

• The P_BAD promoter on the backbone may be malfunctioning.
• The induction condition may not be the optimal one.
• Co-transformation might not work because pBad24 and pET28a have the same origin of replication.

We planned to verify the P_BAD promoter, explore the optimal induction condition, and try to place all the gene elements on one single plasmid to avoid co-transformation in the next cycle.

Learn

We learned that plasmids with the same origin of replication could hardly exist in the same bacterial cell. We also learned more experience of troubleshooting.

Cycle 3

Design

We combined the two plasmids into one single plasmid and measured the effect of integrase using qPCR and fluorescence spectrophotometry. To verify the P_BAD promoter intensity, we placed mCherry in the MCS of an empty pBad24 to characterize the promoter intensity. We also constructed another plasmid containing mCherry-OR sequence to see if OR sequence would influence the expression of integrase. We intended to measure mCherry fluorescence of the bacteria containing one of these two plasmids with or without arabinose induction.

Build

We successfully constructed the plasmid containing both integrase and verification sequence of intergase (pLogic 1), and the two plasmids for promoter verification.

Test

Fluorescence spectrophotometry results indicated that the P_BAD promoter worked well with minor leakage, and OR sequence had little influence on the expression of downstream reporter gene. Characterization of integrase using pLogic1 was also successful by fluorescence spectrophotometry, but not qPCR.

Learn

We learned that combining two plasmids into one can be an efficient way to avoid co-transformation. Moreover, we were still striving to improve our protocols for qPCR.

Replacement of cro

Design

In our previous design, we used the cro sequence from HUST 2016 iGEM project (BBa_K2036027), which has been validated to work with the cI repressor to interoperate. However, in the course of our experiments, the derepression effect of Cro was always unsatisfactory. Consequently, we considered replacing the cro sequence with another homologous one.

Build

In a paper on bacterial memory elements^[1], we found another homologous but slightly different cro sequence. We synthesized the sequence and replaced the original cro sequence with the new one in the plasmid by HiFi cloning.

Test

We used the same protocol as the previous experiments to verify the functionality of Cro again.

Learn

The results demonstrated that the new Cro was more suitable for our system. So we chose this version of Cro in the final design of pLogic2.

Cycle 1: Random Forest → RNN

Design

We first attempted to apply the traditional machine learning model to predict the promoter strength of a given sequence. To be more specific, we applied the random forest model.

Build

We built the random forest model for our data set based on the skicit-learn library.

Test

However, no sound result was obtained because we could not determine which features were important when manually selecting features using the traditional machine learning model.

Learn

We realized that the traditional learning model was not feasible. Therefore, we hoped to try the deep learning method.

Cycle 2: RNN → LSTM

Design

Since our goal was to extract information from the promoter sequence, we considered employing the RNN model with sequential characteristics, which may improve the original effect.

Build

We built the RNN model for our data set based on Pytorch.

Test

After testing, the results outperformed the original one, but still failed to meet our expectation.

Learn

We speculated that this was likely due to the gradual decrease in the weight of bases far from the transcription start point during the parameter update process.

Cycle 3: LSTM → Transformer

Design

In order to obtain better results, we improved the RNN model to the LSTM model. The LSTM model could capture the information of the base in the promoter sequence far from the starting point, which is more conducive to information extraction.

Build

We built the LSTM model for our data set based on Pytorch.

Test

After testing, the results outperformed the RNN model, but we desired for better results.

Learn

Essentially, the LSTM model was still a variant of the RNN model. However, due to the sequential nature of its parameter update process, its parallelism was poor.

Cycle 4: Transformer → Transformer_Encoder

Design

We searched for methods with higher parallelism to break through the resctrictions. Therefore, we turned our attention to Transformer, a model with higher amount of parameters and higher representation power.

Build

We built the Transformer model for our data set based on Pytorch.

Test

After testing, the results outperformed the LSTM model, but the running speed was much slower.

Learn

The Transformer model was originally suitable for machine translation tasks, but in practice it was not fully applicable to our biological problem. The Decoder structure in Transformer may not be necessary in our problem, which may affect the training speed.

Cycle 5: Transformer_Encoder → Transformer_Encoder+Dropout

Design

In order to make the Transformer model more suitable for this biological problem, we chose to use only the Encoder structure in the Transformer, which was good enough to extract features.

Build

We built the model of Transformer_Encoder for our data set based on Pytorch.

Test

After testing, the results outperformed the precious model, but the phenomenon of overfitting always occurred.

Learn

We introduced the Dropout mechanism to solve the problem of overfitting and received wonderful results.

For more detailed explanation of the abovementioned model, please visit our Modeling page.

Cycle: Pressure pump drive → Finger-pressing pump drive

Design

At first, we used traditional microfluidic technology, which drove the liquid flow by external pressure pump. The gradient generator composed of serpentine channels was located on both sides of the main channel. And this design made the chemokine solution flow in from one side and flow out from the other side.

But this microfluidic chip need external devices to provide power for the flow of liquid, which would bother our customers if they need to connect external devices for the microfluidic chip at home. In order to solve the problem, liquids with different concentrations of chemokine will be driven by a finger-pressing pump to form a concentration gradient. The chemokine loading wells in the gradient generator became much larger to fit in the size of human fingers, and only one side of gradient generator was retained to make the chip small enough for household use.

Build

After determining the structure of the mcrofluidic chip, the channel was built using PolyJet 3D printing technology via a Stratasys Connex3 Objet260 3D printer.

Test

The chip of new design can still form concentration gradient, and the liquid flow out of two finger-pressing pumps can fill in the overall channels of our chip.

Learn

We introduced the finger-pressing Pump to make our products more convenient for customers.

[1] Kotula, J. W., Kerns, S. J., Shaket, L. A., Siraj, L., Collins, J. J., Way, J. C., & Silver, P. A. (2014). Programmable bacteria detect and record an environmental signal in the mammalian gut. Proceedings of the National Academy of Sciences of the United States of America, 111(13), 4838–4843. https://doi.org/10.1073/pnas.1321321111