Engineering

Chapter I - Introduction

Our primary goal was to engineer novel toehold switches capable of detecting α-fetoprotein (AFP) mRNA within mammalian cells with expression of our protein of interest. We achieved this through a single, comprehensive Design-Build-Test-Learn (DBTL) cycle that took us from initial research to a final, curated list of candidates ready for synthesis.

Design: laying the foundation

The design phase of our cycle began with foundational learning. A thorough review of existing literature taught us the critical design principles of toehold switches, especially the nuances of adapting them for eukaryotic systems. We started our journey on classical paper by Alexander Green from 2014, and analyzed number of papers up to the current year.

This knowledge informed our core design strategy:

Target: α-fetoprotein (AFP) mRNA.
Chassis: mammalian HEK cells.
Architecture: we adopted an innovative structure, inspired by the 2019 iGEM Tel-Aviv team, placing the Kozak sequence after the toehold stem. This design promised a more stable off-state and simplified construction by removing the need for an N-terminal peptide linker.
Tool: we chose the Triggate web server, as it was specifically developed to support this novel architecture.

Build: executing the design

The build phase transitioned our design into practice. Initially, this involved using the Triggate server to generate toehold structures. Our first "test" of this process yielded promising results for simple queries, validating our initial approach.

However, as we moved to more complex and optimized queries, our process failed its test. The server consistently timed out, unable to complete the calculations. This was a critical roadblock. We were faced with a choice: abandon our innovative design or engineer a solution.

This challenge triggered a rapid, internal "learn-and-rebuild" loop within our main cycle. We learned by contacting Triggate's creator, Yehuda Almog, who confirmed the issue was a server-side timeout, not a fundamental flaw in the algorithm. This insight was pivotal. We then built a new, robust computational pipeline to overcome this limitation:

A "divide and conquer" script: I developed a Python script using Selenium to programmatically split our large AFP target into smaller, overlapping fragments and submit them as individual jobs.
An automated data pipeline: I created a Google Apps Script to automatically capture the hundreds of results and pipe them into Google Drive, followed by a Colaboratory script to parse and organize the data into a master Google Sheet.

This custom-built pipeline effectively solved the server-side timeout issue and allowed our team to generate the comprehensive dataset we needed.

Test: validating the pipeline

A "divide and conquer" script: Python + Selenium to split AFP target and submit jobs.
Automated pipeline: Google Apps Script → Drive → Colab → master Sheet.

This custom-built pipeline effectively solved the server-side timeout issue and allowed our team to generate the comprehensive dataset we needed.

Learn: analysis and selection

With a comprehensive dataset in hand, we entered the final, crucial Learn phase of our cycle. The goal was to derive actionable knowledge from the hundreds of candidates our pipeline had generated.

First, we learned from experts, consulting with several toehold switch researchers to establish a rigorous set of selection criteria. We consulted with several toehold switch researchers to refine our selection criteria.

Allen Liu – to consult the architecture of our toehold we contacted Dr. Liu. He kindly explained our doubts about integration of 5’UTR into our toehold and clarified that as the toehold is fully 5’UTR, additional stimulation is not necessary.
Antonis Guakonutis (iGEM Thessaly) – emphasized the importance of secondary structures of target mRNA and how to counteract them, prompting deeper analysis of predicted toehold structures.
Francois Rouleau – helped us understand how to travel the space of various toehold parameters and compare candidates across them.

Armed with this expert advice, we analyzed our data through a multi-step filtering process:

Functional consistency: robust predicted structures for both our reporter (GFP) and therapeutic (Gasdermin D) genes.
Specificity: BLAST against the human transcriptome to eliminate any candidates with potential off-target effects.
Biophysical ranking: prioritize a low stem Gibbs free energy (ΔG) to ensure a stable "off" state and minimize leakiness.

Chapter II - Cloning

We engineered two transcription units — TU1 (toehold–eGFP) for sensing and TU2 (RFP) as a transfection/control readout — and iterated through short DBTL cycles to de-risk cloning and screening. Below is a structured recap of what we planned — built — tested — learned across three iterations.

Iteration I
Design: plan with Asimov set

We initially planned to use the iGEM Parts Registry, basing our work on the Asimov Mammalian Parts Collection from the 2024 and 2025 Distribution Kits, generously shared by the iGEM JU 2024 team and Dr. Mateusz Wawro. Our goal was to assemble a transcription unit (TU1) for each toehold switch using Modular Cloning (MoClo): CMV promoter (BBa_J433000), the toehold switch sequence as the 5′UTR, an eGFP protein coding sequence lacking a start codon and Kozak sequence (included in the toehold switch), a synthetic 3′ UTR (BBa_J433018), and the rabbit β-globin poly(A) signal (BBa_J433023).

As a control for transfection efficiency and plasmid uptake, we also designed TU2, composed of the CMV promoter, strong 5′ UTR (BBa_J433003), a Kozak sequence with RFP CDS, the 3′UTR, and the rabbit β-globin poly(A) signal. Both TUs would later be merged into a Level-2 multigene construct for screening.

Iteration I
Build: transformations from kits

Dried DNA from the 2024 and 2025 kits was rehydrated in 10 µL of mQ water and 1–2 µL was used to transform 50 µL of E. coli DH5α and DH10β. Despite correct positive (pUC control) and negative (mQ) controls, only a few colonies appeared even after three attempts.

Iteration I
Test: sequence verification

Mini-prepped colonies were sequenced with primers flanking the inserts. Sequencing revealed that several plasmids did not match their reported identities, and even clones from the same transformation showed inconsistent results — including one that showed bGH (BBa_J433021), a part we had never attempted to extract — ruling out cross-contamination.

Iteration I
Learn: confirm & pivot

After consulting iGEM’s representative Cristiane Toledo and Asimov’s representative Traci Haddock, we confirmed our protocol was correct. Additionally, we learned that other teams (e.g., Barnacure WCIS iGEM 2025) had similar issues. Due to time constraints, we abandoned the Distribution Kit strategy and opted to synthesise redesigned parts directly, which also simplified our cloning workflow.

Iteration II
Design: GG-ready, no scars

We ordered sequences designed identical to the Asimov collection, but already assembled, with a Golden Gate Assembly cloning site allowing insertion of 5′UTR and CDS. Because the toehold switch and protein coding sequences cannot tolerate Golden Gate scars (a 4-nt scar would cause a frameshift), we ordered them as continuous fragments without scar. We also ordered an already assembled transcription unit 2 with overhangs necessary for assembly; however, the vendor could not synthesise the full transcription unit due to heavy secondary structure, so we split it into three fragments with 20-nt overlaps for Gibson Assembly.

Iteration II
Build: assembly attempts

Ordered constructs were transformed and plasmid stocks prepared. We attempted to first assemble TU2 as a linear construct and then use it in Golden Gate Assembly with TU1 and a Level 2 backbone, although it was unsuccessful. Then we prepared a linear vector to first clone TU2 into a plasmid and then proceed with Golden Gate Assembly, but it also failed.

Iteration II
Test: no correct clones

Unfortunately, we were unable to obtain any positive colonies, even after several repetitions and modifying the protocol.

Iteration II
Learn: narrow scope

With deadlines approaching, we chose to proceed using only TU1 and interpret results conservatively. Winning an Ansa award for iGEM teams later enabled us to order highly structured sequences like TU2 for future work.

Iteration III
Design: lacZ for blue–white

To simplify screening, we inserted a lacZ cassette into the replaceable region of the TU1 plasmid, allowing blue–white screening during insertion of toehold switches. We did not do it initially, because it would make blue–white screening of colonies during combining TU1 and TU2 impossible.

Iteration III
Build: Golden Gate insertions

We cloned the lacZ insert into TU1 and used BsaI Golden Gate Assembly to introduce different toehold switches with eGFP.

Iteration III
Test: PCR & prep

Correct insertions were confirmed by colony PCR, followed by midipreps on 100 mL of overnight culture to obtain endotoxin-free plasmids for human cell transfection.

Iteration III
Learn: ready for transfection

This approach proved efficient and fast. Sanger sequencing verified the constructs, and they were ready for experiments to assess toehold switch activity.

Chapter III - Choosing our bests

To design novel toehold switches that could serve as targeted therapy for hepatocellular carcinoma, we refined our validation and selection strategy. After several DBTL loops we converged on a procedure that fits our scope, resources, and timelines.

Iteration I
Design: evolution of complexity

The Design phase of our cycle began with the simple idea of checking our toehold switches designs in two steps. First, we wanted to do the primary screening process on a simple and cheap organism, then we wanted to test only those promising ones on a second, more complex object (just like it's usually done during drug development research — first tests on mice and the most promising on people).

Iteration I
Build: choice of screening hosts

It was obvious that the second step would be tested on the target cells — human hepatocellular cancer (cell line: HepG2).

We spent more time on choosing our first screening host. A brief idea was tests on commonly used bacteria, but due to the different structure of prokaryotic translation initiation sites we quickly abandoned this solution. Then we turned to the most simple and commonly used eukaryotic organisms — yeasts. It was supposed to be a time- and budget-friendly preselection, and only five most promising toehold designs would be checked on the more expensive in culture mammalian cells.

Iteration I
Test: talk with expert

We consulted our idea with Magda Konarska, who worked on the yeast genome. We learned from her that expression between yeast and mammalian cells can differ due to other components and complexes involved in the mRNA-binding process in the cytosol. That could possibly lead to the choice of wrong toehold designs.

Iteration I
Learn: overcomplicated?!

After learning that we could have selected wrong toehold designs by mistake while performing prescreening on yeast, we discussed with Magda Konarska different screening methods. The conclusion of our talks was quite surprising: our selection process was overcomplicated for the scale of our project. An easier, less time-consuming, and affordable solution was to perform all selection on the target cells.

Iteration II
Design: what about controls?

When we decided on tests on the HepG2 cell line, we considered different cell lines for negative control. Our toehold designs are capable of detecting AFP mRNA within mammalian cells; after detection the toehold becomes active, so we needed a cell line where the toehold would stay in the inactive form. One of the most popularly used human cell lines without α-fetoprotein (AFP) expression were HEK293T.

We felt conflicted about comparing tests on different cell lines and came up with the idea of developing a HEK293 cell line with expression of α-fetoprotein (AFP) and using it as a positive control instead of HepG2.

Iteration II
Build: talk with experts

We discussed the development of the new HEK293 cell line with Pawł Sikorski. He told us that while it is possible to make cells express α-fetoprotein, establishing a stable cell line with α-fetoprotein expression will be extremely time-consuming.

We also consulted with Maciej Cieśla on the idea of using lentiviruses or plasmids to make new HEK293 cells with α-fetoprotein expression. He revealed that unfortunately this method would be quite resources-consuming. Conclusions from both experts were that it would be better to use already well-established cell lines. They also reminded us that the human body is built out of different cell types and suggested that it might be acceptable to test our toehold switches on different cell lines.

Iteration II
Test: literature search

We decided to check in literature whether people test subjects of their research on different cell lines and compare the results. We found that it is a frequent methodology used by scientists.

Iteration II
Learn: back to original

We learned that our original idea — to test our toehold designs on different cell lines and collect results on flow cytometry — was well thought-out. We also considered more tests on other cell lines in the future.