Engineering

Overview

We started with the aim of creating a high throughput workflow that can produce variational data and evaluate diametrically opposing characteristics. We achieved this goal through our several DBTL Cycles. Using STARs as a case study, we successfully integrated the target and switch into a single-plasmid system that can generate a library of different sense-antisense pairs. We also successfully designed a diametric selection marker that can screen cells for the mutually exclusive ON and OFF configurations of their RNA switch, thus capturing ON/OFF information.

RNA Switch Single-Plasmid System

Our objective was to engineer a one-plasmid system where both the RNA switch and the target are linked on a single strand. This would ensure: 1. 1:1 balance in a single cell for an easier control of expression 2. Physical linkage of the RNA switch and target sequences to enable future creation of variance needed for data generation. At this stage, we started by exploring the Toehold switch system and STAR system.

Developing the Toehold Switch Single-Plasmid System

Design

The single-plasmid system required the toehold switch fused to GFP (under the control of an aTc-inducible pTet promoter) to be joined with the target (under the control of an IPTG-inducible pLac promoter) on a single plasmid. Toehold switches were chosen for their high levels of orthogonality, modularity and specificity.

Build

The toehold switch and target were successfully integrated into a 300bp stretch of DNA on a single plasmid.

Test

The efficacy of the system was quantified by measuring the expression of downstream GFP when both the target and switch were expressed. Results were compared against GFP expression under the pTet promoter, without the switch. For further details on our Build and Test steps, Click here.

Learn

However, it was determined that the toehold single-plasmid system did not exhibit significant levels of GFP expression. We learnt that toehold switches might not be best suited to be integrated into a single-plasmid system, and that the change of promoters could have detrimentally affected system functionality. Hence, we decided to shift to using STAR RNA switches.

Developing the STAR Single-Plasmid System

Design

The single-plasmid system required the STAR antisense strand fused to GFP to be joined with the sense strand on a single plasmid. Learning from the previous DBTL cycle, we cloned the target and switch under the original constitutive promoters. STAR switches were chosen as they were well-characterized.

Build

The antisense and sense sequences were successfully integrated into a stretch of DNA on a single plasmid.

Test

The cells containing the single-plasmid system had a high GFP output, but the level of fluorescence was only around half that of the two-plasmid system. We learnt that the single-plasmid system was successful, but that it could be further optimised to reach higher levels of expression. For further details on our Build and Test steps, Click here.

Learn

It was determined that our system activated a significant level of GFP expression when both the sense and antisense strand were expressed. This demonstrates that the STAR single-plasmid system works successfully, as we have designed. We will focus on STAR in our next cycle which involves the creation of oligo library and developing the one plasmid system for Golden Gate assembly.

Oligo Library Generation and developing the Single-plasmid System for Golden Gate Assembly

In the next cycle, we aimed to engineer a one-plasmid system where an oligo pool of unique sense-antisense sequences could be inserted into the plasmid. This would enable the one-plasmid system to generate a library of different sense-antisense pairs, linking it to data generation.

Developing the STAR Oligo Single-Plasmid System

Next, we needed to see how the single-plasmid system fared when unique oligos were integrated into the plasmid.

Design

After functionality of the single-plasmid system was confirmed, we created a computational model to design our oligo pool through random generation. The STAR oligo single-plasmid system required the oligo (consisting of the sense and antisense pairs) to be joined with the backbone on a single plasmid via Golden Gate Assembly. This system would be able to generate strand-wise variation.

Build

The oligo and the backbone sequences were successfully integrated into a single plasmid by our novel Golden Gate Assembly protocol.

Test

The functionality of the system was tested by measuring the expression of downstream GFP when both the sense and antisense strand were expressed. Results were compared against GFP expression under the pTet promoter, without the antisense strand. For further details on our Build and Test steps, Click here.

Learn

It was determined that our system activated a significant level of GFP expression when both the switch and target were induced. This demonstrates that the unique oligo was successfully integrated into the single-plasmid system, as per our novel plasmid design and new Golden Gate protocol. It also showed that our system would be able to generate strand-wise variation.

Developing the Diametric Selection Marker

Our objective was to engineer a diametric marker that could directly screen for the ON/OFF characteristic in cells. This would enable us to capture ON/OFF information, despite the ON and OFF states being mutually exclusive.

Characterizing SacB

Firstly, we aim to characterize a negative selection marker.

Design

The SacB gene was chosen as it is a well characterized negative selection marker.

Build

The sacB gene was fused to a protein linker which was fused to the C-terminal of GFP, and the fusion protein was placed under the control of the IPTG inducible pLac promoter .

Test

SacB was characterized by measuring the growth of the cells and the expression of downstream GFP in differing conditions of 0%, 0.5%, 1%, 2.5%, 5% of sucrose and 0, 25, 50, 100μM of IPTG over 6 hours. For further details on our Build and Test steps, Click here.

Learn

The negative selection marker component of our diametric marker is meant to screen against leaky variants by killing them in culture. We now have found the ideal concentration of sucrose to use, 0.5%, which would allow us to tune the level of 'acceptable' leakiness.

Testing Killing Mechanism of sacB

Next, we tested the effect that intracellular sacB could have on a non-sacB producing population of cells.

Design

Since the mechanism of sacB is not well understood, we aimed to find out if sucrose converted to levan in one cell affects the survivability of other cells. Cells expressing RFP were taken from our lab.

Build

Cells expressing GFP-sacB were co-cultured with cells expressing RFP.

Test

Several cocultures were made, with different starting proportions of the 2 cell populations. All populations were exposed to either 100uM IPTG only, 1% sucrose only, both conditions or neither, and their growth tracked over 10 hours. For further details on our Build and Test steps, Click here.

Learn

It was determined that sucrose has a negative effect on cell growth, independent of sacB expression. We learnt that not only does sacB not negatively affect the growth of surrounding cells, it provides a growth advantage to non-sacB producing cells, which assists in our use of the marker to distinguish leaky and non-leaky RNA switches.

Testing SacB Functionality

We aimed to test our screening workflow independently of the RNA switch, using the negative selection marker alone.

Design

Three different plasmids were designed to give the permanently off sacB, permanently on sacB and inducible sacB characteristic. The sacB gene was chosen as it is a well characterized negative selection marker.

Build

The GFP-sacB inducible plasmid required sacB gene fused to a protein linker which was fused to the C-terminal of GFP, and the fusion protein was placed under the control of the IPTG inducible pLac promoter.

Test

A mixture of cells containing the three plasmids were grown under differing conditions, in order to test if sacB would be able to negatively select for the appropriate plasmid. For further details on our Build and Test steps, Click here.

Learn

It was determined that the cells containing SacB were killed the presence of sucrose. This demonstrates that the SacB negative selection works successfully, as we have designed.

Testing LacI Mutagenesis

We aimed to test our screening workflow independently of the RNA switch, using the pLac promoter as an example.

Design

Our workflow for screening ON/OFF strength of the RNA switch should in theory be applicable to any switch that controls protein expression. Hence we focused our efforts on the pLac promoter, a well characterized and commonly used promoter in E.coli.

Build

Error Prone PCR was used to mutate the LacI gene and simulate a variety of switches with different ON/OFF ratios and leakiness

Test

The effect of mutagenesis was tested by measuring the expression of GFP under the control of IPTG-inducible pLac promoter in the cells with mutated LacI. Results were compared against GFP expression of cells containing normal LacI. For further details on our Build and Test steps, Click here.

Learn

Different variants were shown to have different responses to IPTG, ranging from increased expression to increased leakiness or insensitivity to IPTG. Thus, we established that error-prone PCR could produce a variety of phenotypes to check our screening against. This also demonstrated that we could generate pair-wise variation.

Deep learning modelling

Deep learning model has 2 directions. We looked into predictive and generative models.

The first step was to look at Wasserstein Generative Adversarial Networks as it was suggested from human practices.

Design

The generator's primary goal is to produce RNA switch sequences corresponding to a given target sequence. Our initial approach is to explore WGANs (Wasserstein Generative Adversarial Networks) and Transformers as the starting point for our model building process.

Build

The input dataset, sourced from 'A Deep Learning Approach to Programmable RNA Switches,' was processed, including filtering out null entries. A GAN was created with a generator and discriminator.

Test

WGAN implemented with gradient penalty failed to compile. The technical issue prevented us from going forward with this model architecture.

Learn

Due to time and knowledge constraints, we shifted to a simpler Transformer architecture for RNA switch sequence generation. The Transformer model successfully generated accurate sequences, achieving impressive error rates of 0.06 for characters and 0.05 for words.

Concurrently, we looked into creating a Generalised Predictive Model for RNA interactions

Design

In order to capture the RNA-RNA interactions directly, we took inspiration from Siamese-Bert, which creates identical copies of a language model that analyses each sequence. While its used for sequence similarity tasks, We adapted it for interaction modelling.

Build

A large dataset of 9 million RNA-RNA interaction was gathered from the RNAInter database(Kang et al., 2021) and a Siamese Language architecture was designed to analyse the sequences to predict binding activity. 80% of the data set is used for training and 20% is used for testing.

Test

We tested using the data and using traditional metrics for regression tasks such as, Pearson R and Mean Squared Error.

Learn

The model was a success at predicting binding! It achieved a Pearson R value of 0.937 showing strong correlation and a Mean Squared Error of approximately 0.01.

Next on the predictive side, the model was finetuned specific RNA technologies

Design

Given our generalised model worked, we then aimed to test it on specific downstream tasks. We used the Toehold On/Off ratio data.

Build

The model architecture remained the same but was used as a starting point for fine-tuning.

Test

The model was evaluated using the evaluation metrics used to assess the original papers. For the Toehold efficiency model, Pearson R and Mean Squared Error were the metrics.

Learn

Our results improved with a pearson r of 0.56 and a mean squared error of 0.056. From this, we validated our hypothesis that language data allowed for finer prediction of RNA properties.

Reference

Kang, J., Tang, Q., He, J., Le, L., Yang, N., Yu, S., Wang, M., Zhang, Y., Lin, J. S., Cui, T., Hu, Y., Tan, P., Cheng, J., Zheng, H., Wang, D., Su, X., Chen, W., & Huang, Y. (2021). RNAInter v4.0: RNA interactome repository with redefined confidence scoring system and improved accessibility. Nucleic Acids Research, 50(D1), D326–D332. https://doi.org/10.1093/nar/gkab997