Engineering

Result

High background on the screen: empty vectors and off-target bands, with Golden Gate efficiency low throughout.

Changed next

The whole build moved into the SubtiToolKit, and an intermediate cloning step went in ahead of Level 1.

1. DDesign

   Based on the characterization of the CcaS/CcaR optogenetic system reported in Castillo-Hair 2019 [1], we redesigned and generated the coding sequences of the green/red light-responsive two-component system (CcaSm3 and CcaR), together with the PCB production enzyme (ho1-pcyA fusion), a version that can increase the PCB level, and our target output and reporter coding sequences (LEA, ACCD, BoPEP4) by the modular cloning (MoClo) system [3].

   To assemble these functional transcription units into the MoClo 2.0 system provided from the distribution kit, we designated standard regulatory elements for each module. We designed custom PCR primers incorporating type II restriction enzyme recognition sites (BsaI for Level 1 assembly) flanked by specific standard overhangs following every Level 1 construct. This was designed to allow sequential assembly of individual promoters, RBSs, CDSs and terminators into single transcriptional units (Level 1), which could subsequently be directionally combined into multi-gene circuits at Level 2.

2. BBuild

   We performed PCR amplification on template DNA to add the designed restriction sites and overhangs to each individual part. The resulting raw PCR products were purified and subjected directly to Golden Gate digestion-ligation reactions with our Level 1 destination vectors (pJUMP-29) to construct the complete Level 1 single gene modules. The assembly mixtures were transformed into E. coli DH5α competent cells and plated on selective media.

3. TTest

   Transformant colonies were screened by colony PCR to check for successful Level 1 Golden Gate assembly, which yielded inconsistent and low-efficiency cloning. Amplified fragments were analyzed via agarose gel electrophoresis to assess insert size. A subset of colonies yielding bands near the expected sizes were supposed to be selected for plasmid isolation and subsequent restriction digest verification. All trials were unsuccessful and as a result, data is not shown.

4. LLearn

   Colony PCR screening and restriction digests revealed high background, including potentially empty vectors and off-target genes, making it difficult to distinguish from assembly failures and PCR artifacts. We learned that attempting Level 1 assembly straight from unverified PCR products introduced too much ambiguity, highlighting the need for an intermediate cloning step to generate clean, sequence-verified Level 0 entry parts.

Result

Colonies came up, and gave inconsistent or unconfirmed screens.

Changed next

Verify the donor before assembling with it.

1. DDesign

   Following the lessons learned in the DBTL 1 cycle regarding insufficient expression as shown in the gel electrophoresis results from the P_veg-RBS0 promoter-RBS combination, we redesigned the backbone of our PCB production module. Because of the problems with unverified PCR products identified in DBTL cycle 1, we realized that we must add an intermediate cloning step to isolate and sequence verify our Level 0 entry parts first.

   For our Level 1 vector, we utilized pSTK-1A-sfGFP from the SubtiToolKit (STK) library [4], as well as pSTK-1B-sfGFP, pSTK-1C-sfGFP and pSTK-1D-sfGFP for our PCB, CcaS and CcaR modules, respectively. We initially utilized the MoClo 2.0 standard in our first DBTL cycle, but we realized the resulting quality was quite poor as transformation didn’t reveal any colonies, and the Golden Gate assembly efficiency was low. As a result, we switched to the SubtiToolKit, a bioengineering kit for Bacillus subtilis.

   To convert this coding sequence from our original MoClo design into an STK-compatible Level 0 part, we designed a forward primer (Ho1 F) and a reverse primer (pcyA R) to PCR-amplify the entire Ho1:pcyA operon as a single amplicon directly from our existing MoClo construct. We designed a forward primer (ccaSm3 F) and a reverse primer (ccaSm3 R) to PCR-amplify the full ccaSm3 coding sequence; we created a forward primer (CcaR F) and reverse primer (CcaR R). We designed a forward primer and reverse primer for each CDS with the flank of the STK 4 bp overlap sequence and BsaI restriction enzyme recognition site built on the same STK system established in Module 1 and Module 2, with gene specific annealing regions at the 3′ ends; and we designed the according forward and/or reverse primer for each output module CDS (csn:ACCD, csn:LEA) following the same modular tail architecture as for other parts from MoClo.

2. BBuild

   We executed the Level 1 Golden Gate assembly following the SubtiToolKit workflow. The verified pGEM-Teasy-Ho1:pcyA Level 0 plasmid, along with STK library parts as well as MoClo 2.0 converted STK parts, were assembled together and inserted into the according pSTK-sfGFP vector. We utilized the type IIS restriction enzyme BsaI to digest all Level 1 fragments, followed by ligation with T4 DNA ligase. The resulting Golden Gate assembly reaction mix was transformed into E. coli DH5α competent cells and selected on ampicillin and chloramphenicol media based on the insert vector. Positive colonies were picked, screened and verified clones were extracted via miniprep. Then, either plasmid or liquid culture was sent for sequencing to confirm the successfulness of the independent modules in our optogenetic system.

3. TTest

   Colonies appeared and the screens were inconsistent.

4. LLearn

   With a raw PCR product as the donor, a wrong band can be a failed assembly or an artefact of the amplification. The next three cycles are all attempts to fix that one problem.

Result

Seven coding sequences cloned, screened and sequence-confirmed before any Golden Gate reaction.

Changed next

A marker clash nobody had checked for.

1. DDesign

   Based on the results from DBTL 2, we shifted our focus to the SubtiToolKit cloning workflow itself and left the genetic architecture of each module alone. Since the BsaI-only Level 1 entry step had introduced ambiguity in our earlier assembly, with colonies producing inconsistent or unconfirmed results, we revisited the cloning strategy at the PCR and vector-entry stage to improve reliability before moving into Level 1 Golden Gate assembly. The overall part composition for each module remained unchanged from DBTL 1 and 2; the redesign targeted the intermediate cloning steps used to generate clean, verified Level 0 parts.

2. BBuild

   To improve the reliability of our Level 0 entry step, we amplified each module’s coding sequence by PCR and cloned the amplicon into a pGEM-T Easy vector [5]. pGEM-T Easy provides a direct, rapid method for cloning PCR products through direct TA cloning: the vector is supplied pre-linearized with single 3′-thymidine (T) overhangs, which are complementary to the single 3′-adenosine (A) overhangs left on PCR products by Taq and other non-proofreading polymerases. This complementarity allows the amplicon to ligate directly into the vector without any additional end-polishing, restriction digestion, or blunting step, making it a fast and efficient intermediate cloning strategy compared to attempting Golden Gate entry directly from a raw PCR product.

   A plasmid backbone is also generally easier to work with and verify than a linear PCR product: pGEM-T Easy clones can be more reliably purified, screened by colony PCR, and sequence-confirmed before being subjected to the BpiI/BsaI digestion-ligation reactions required for Level 1 assembly, reducing the risk of carrying an uncharacterized PCR error into downstream Golden Gate steps. Positive pGEM-T Easy clones carrying each construct’s coding sequence were identified by colony PCR and sent for sequencing to confirm the insert prior to use in subsequent Golden Gate reactions.

3. TTest

   Colonies from the pGEM-T Easy transformations were screened by colony PCR using primers, and PCR products were run on an agarose gel alongside a DNA ladder to assess assembly success. A correct colony PCR band is one that runs at the expected size, calculated as the length between the T7 promoter and the SP6 promoter amplified by the screening primers; this confirms that a DNA fragment of the correct length was successfully ligated into the vector at the expected position. Colonies producing a single, clean band at this expected size were considered positive candidates for the correct insert. In contrast, colonies producing no band, a band at the size of the empty vector (i.e. no insert, indicating self-ligation of the pGEM-T Easy vector), or bands at incorrect or unexpected sizes were classified as negative and excluded from further analysis, since these patterns indicate either an empty vector or a truncated or mis-primed amplicon rather than the intended insert.

   Colonies showing the expected band size were selected and sent for sequencing to confirm that the cloned insert matched the intended coding sequence with no mutations, insertions, or deletions, prior to proceeding with Golden Gate digestion into the STK system.

4. LLearn

   The pGEM-T Easy TA cloning step proved to be a fast and reliable way to generate verified Level 0 entry material before doing Golden Gate assembly. Because colony PCR alone can only confirm that an insert of approximately the correct size is present, and cannot rule out point mutations or small indels introduced during PCR amplification, we relied on sequencing rather than enzyme digestion to confirm the fidelity of each cloned insert; we did not perform additional restriction digest verification at this stage, since correct band size on colony PCR combined with full-length sequencing was sufficient to confirm both insert identity and sequence accuracy before proceeding to BsaI-based STK assembly.

   This step did add an extra day to our workflow compared to attempting direct entry from the PCR product, since colonies had to be picked, screened, and miniprepped before sequencing. However, it reduced the ambiguity we had encountered in DBTL 2, where using PCR products directly made it harder to distinguish true assembly failures from PCR artifacts.

The six amplicons that went into pGEM-T Easy, as designed. Each carries the same tail at both ends: a spacer, a BpiI site that sets the Level 0 overhang, and a BsaI site that sets the Level 1 position. The sequencing reads that confirmed them are what let cycle 3 close.

Show details The seven sets and their expected sizes, two more screens, and what the sequencing added.

A plasmid is also easier to work with than a linear fragment. A pGEM-T Easy clone can be purified reliably, screened by colony PCR and sequence-confirmed before it ever meets the BpiI and BsaI digestion-ligation reactions, which keeps an uncharacterised PCR error from propagating into every downstream assembly. Three outcomes were scored negative and excluded: no band at all, a band at the size of the empty vector, meaning the vector self-ligated, and a band at any other size, meaning a truncated or mis-primed amplicon.

Set	Insert	Expected
A	pGEM-Teasy-Csn:Lea	811 bp
B	pGEM-Teasy-Csn:AmilCP	1,000 bp
C	pGEM-Teasy-Csn:Accd	1,372 bp
D	pGEM-Teasy-CcaSm3	1,642 bp
E	pGEM-Teasy-CcaR	937 bp
F	pGEM-Teasy-Ho1:pcyA	1,726 bp
G	pGEM-Teasy-PcpcG2-veg	412 bp

TA cloning proved a fast and reliable way to generate verified Level 0 material. Colony PCR alone can only say that an insert of about the right length is present; it cannot rule out a point mutation or a small indel introduced during amplification. Fidelity was confirmed by sequencing, and no restriction digest was run at this stage, because correct band size plus full-length sequencing already answers both questions the digest would have answered.

Result

The intermediate that fixed cycle 2 could not be selected against at Level 1. Rebuilt on the kit’s own entry vector, all three Level 0 screens came back clean.

Changed next

Repeated regulatory parts across the modules became a stated recombination worry.

1. DDesign

   After DBTL 3, we further optimized our SubtiToolKit cloning strategy (v2) in response to a compatibility issue identified with the pGEM-T Easy intermediate cloning step: the pGEM-T Easy vector uses the same antibiotic resistance gene (ampicillin) as our Level 1 destination vectors (1A, 1B, 1C, 1D), which complicated downstream selection and increased the risk of background contamination from carryover vector during Golden Gate assembly. To resolve this, we redesigned our cloning approach to construct each module’s coding sequence directly using STK Level 0 primers, skipping the pGEM-T Easy intermediate altogether. This step was only relevant to the CDS not cloned into pGEM-T Easy, as the ones from DBTL 3 were still usable in Level 1 Golden Gate assembly.

2. BBuild

   We amplified the coding sequence directly with primers carrying the STK Level 0 (BpiI) and Level 1 (BsaI) tail architecture, as described for Modules 1 to 4, and cloned the amplicon directly into the pSTK-0C CDS entry vector rather than through pGEM-T Easy. This removed the antibiotic resistance conflict, since pSTK-0 and our Level 1 destination vectors use distinguishable selection markers (chloramphenicol versus ampicillin), allowing a more obvious difference between correctly assembled Level 1 clones and the vector.

3. TTest

   Three Level 0 screens on 14 August, all clean at their expected sizes, and three clones from each went to liquid culture. The students’ document leaves this step to the gels themselves.

4. LLearn

   Although this approach resolved the antibiotic resistance conflict, it raised a new concern regarding potential recombination events in Bacillus subtilis. Because several of our modules share repeated regulatory sequences (e.g. P_veg, RBS0, B0015) across multiple constructs, we became concerned that homologous recombination between repeated elements, once all four modules were combined in B. subtilis, could destabilize parts of the assembled optogenetic circuit, potentially weakening the CcaS/CcaR signals. This concern motivated us to reconsider the repeated part usage across our modules going into the next design cycle.

Show details The two BoPEP4 screens from the same day, and the redesign this cycle handed to cycle 6.

Only the coding sequences that had not already been cloned in cycle 3 were affected; the ones already in pGEM-T Easy were still usable at Level 1.

Once all four modules sat in one cell, long stretches of identical sequence could give recombination something to act on, and the CcaS/CcaR signal path is exactly what would be damaged. The redesign written at the end of this cycle is one line: diversify the promoter, ribosome binding site and terminator across the modules so that no two of them share a long identical stretch. That is what cycle 6 was run on.

Result

Five constitutive cargoes built and screened, and the ACCD strain carried into the reactor.

Changed next

Building five modules on the same promoter, ribosome binding site and terminator is what made the recombination worry concrete.

1. DDesign

   To construct our circuit’s output module, we designed a library of candidate constructs capable of expressing and secreting functional biotic and abiotic plant protectants. Because the optimal protectant–secretion signal combination could not be determined prior, we adopted a parallel design strategy built on a unified expression format: the constitutive promoter P_veg, a composite ribosome binding site (EEBD-RBS0), a secretion signal peptide (SamyQ or Csn) fused upstream of the protectant coding sequence, a histag to view protein expression, and the B0015 terminator. All constructs were designed for assembly directly into the pSTK-EXP1-sfGFP destination shuttle vector, enabling immediate expression in E. coli as well as downstream transfer into Bacillus subtilis. To integrate parts that weren’t from the SubtiToolKit library (Csn, ACCD, Lea, BoPEP4), we designed custom primers following standard STK architecture: a 5′ spacer, a BpiI site, a BsaI site, and a gene specific annealing sequence.

2. BBuild

   We physically assembled all candidate Level 0 parts using BpiI-mediated Golden Gate cloning into pSTK-0-sfGFP, replacing the sfGFP dropout cassette. The resulting Level 0 entry plasmids were transformed into chemically competent E. coli DH5α, isolated, and sequence-verified. In the second assembly step, the verified Level 0 protectant cassettes were combined with pSTK parts (P_veg, EEBD-RBS0, and B0015) and the pSTK-EXP1-sfGFP destination shuttle vector in a BsaI Golden Gate reaction. Ligated expression vectors were transformed into E. coli DH5α, screen-confirmed by miniprep digestion, and prepared for downstream expression and functional testing.

3. TTest

   Transformant colonies across all five output candidates (BoPEP4, Csn:Lea, Csn:ACCD, SamyQ:LEA, SamyQ:ACCD) were initially screened via colony PCR and analyzed via agarose gel electrophoresis to confirm complete construct insertion into pSTK-EXP1-sfGFP. Liquid was then cultured and sent for sequencing to confirm the existence of the correct inserts. Positive clones producing single, clean bands at expected amplicon sizes were then electroporated into Bacillus subtilis. Transformant colonies were screened by colony PCR to confirm the presence of the full-length output expression cassettes. Electrophoresis of PCR products on agarose gels yielded sharp, distinct bands at the expected sizes, matching our prior sequence-verification data in B. subtilis 168. Verified single colonies were inoculated into liquid media, which were then scaled up for downstream functional testing in our bioreactor.

   ACCD expression of the constitutive circuit was measured over the span of 44 hours. Liquid was then extracted every two hours from the shell site and lumen site of the bioreactor for a western blot in order to optimize and quantify protein expression.

4. LLearn

   Testing of our ACCD in our bioreactor provided critical insight into both expression levels and circuit stability. Using the Csn:ACCD construct allowed us to directly monitor protectant production and evaluate secretion efficiency; however, constitutive runs through our bioreactor and a western blot revealed that while ACCD expression was not clearly detectable, the overall protein yield was relatively low and did not reach optimal physiological levels. This highlighted the need for future optimization of expression elements (such as promoter strength or codon optimization) to boost output efficiency for demanding protectants.

   Furthermore, from an assembly perspective, although this cycle is mostly focused on the constitutive expression of ACCD, we were also working on our inducible system at the same time. Constructing multiple modules containing identical regulatory elements (P_veg, EEBD-RBS0, and B0015) highlighted an important concern. Reusing identical promoter and terminator sequences across all four modules of our full circuit created substantial sequence redundancy. When introduced into Bacillus subtilis, long stretches of sequence identity carry a high risk of triggering homologous recombination, which could destabilize or excise parts of our optogenetic circuit.

Show details The five-cargo library, the LEA arm that failed its digest, and what a colony PCR in B. subtilis can and cannot say.

Every cargo built behind P_cpcG2-veg in module 4 was also built this way, which is the second reason the series exists. All of them went into pSTK-EXP1-sfGFP, so the same plasmid could be screened for expression in E. coli and then carried into B. subtilis unchanged. Four parts the library needed were not in the kit at all, Csn, ACCD, LEA and BoPEP4, and each got a custom primer on the standard four-element tail.

All five output candidates were screened the same way. Clones giving a single clean band at the expected size were electroporated into B. subtilis and screened by colony PCR again, which returned sharp bands at the expected sizes in the new host. Those single colonies went into liquid culture and then up to reactor scale. The protein this strain makes, and where it ends up, is the Results.

The five constitutive constructs

The five constructs built in pSTK-EXP1-sfGFP, and where each one got to.

#	Architecture	Purpose	Where it got to
1	Pveg → EEBD → RBS0 → SamyQ-TAT-HA-AmilCP → B0015	Visible reporter	Built.
2	Pveg → EEBD → RBS0 → Csn:ACCD → B0015	The cargo carried forward	J9 sequenced clean, 27 July.
3	Pveg → EEBD → RBS0 → spacer → GanA-Gala-6×His → B0015	Galactosidase arm	All twelve L lanes failed, 30 July.
4	Pveg → EEBD → RBS0 → Csn:Lea → B0015	LEA protein	Screened at 1,372 bp, 25 July.
5	Pveg → EEBD → RBS0 → SamyQ-Bopep4-v2t1 → B0015	The designed peptide	Level 0 at 536 bp, 14 August.

Result

The rebuilt module sequenced positive on three clones, and three Level 2 clones were checked at all three internal junctions on 19 September.

Changed next

The light plate run that would test the switch. It has since run, and Results carries it.

1. DDesign

   To address the recombination risk identified in DBTL 4, we developed a third version of our SubtiToolKit design (v3) aimed at reducing sequence redundancy across the four modules of our optogenetic circuit. Rather than reusing identical regulatory parts (promoter, RBS, terminator) across multiple modules, we diversified these elements in Module 1 to minimize the length and number of repeated sequences that could serve as substrates for homologous recombination in B. subtilis.

2. BBuild

   We modified the affected constructs by substituting the promoter, RBS, and/or terminator with alternative STK parts for Module 1, distinct from those used elsewhere in the circuit, while preserving its coding sequence. This was intended to maintain equivalent transcriptional and translational output while breaking up long stretches of sequence identity shared between modules.

3. TTest

   Eleven of twelve colonies carried the expected 1,833 bp band on 30 July; A1, A2 and A3 sequenced positive the next day. Every colony of the galactosidase arm on the same gel failed. The students’ document leaves this step to the gels themselves.

4. LLearn

   This last DBTL cycle of our optogenetic light circuit allowed us to finally move on to Level 2 assembly, followed by LPA testing to verify the production of protectants by light signals.

Show details What was varied and what was not, the rebuilt module, the two Level 2 devices, and three contested records.

The team’s own cloning-progress table is the clearest statement of what was varied, with the three regulatory slots pulled apart. Only the promoter, the ribosome binding site and the terminator ever moved; no coding sequence changed in this cycle.

Module	Round	Promoter	RBS	Terminator	What it turned on
1 · PCB production	first	Pveg	RBS0	B0015	No successful clone on the first way of cloning it.
	second	PtagC	RBS	L3S2P21	The second round gave a successful clone. Promoter strength stays pending because the first version was never obtained.
2 · light sensing	first	PlepA	RBS0	B0015	Promoter strength. Both versions were built, and both exist.
	second	PliaG	RBS0	B0015	PliaG is the one sequencing confirmed first.
3 · transcriptional output	only	Pveg	RBS0	B0015	No change at any point.
4 · output	only	PcpcG2	RBS0	cold D, then back to B0015	The terminator swap was made against recombination and then undone.

Module 4: Original terminator cold D (Failure) to B0015. Because Professor Huang thinks recombination in Bacillus is not the issue, we changed our terminator back to our original plan, B0015.

So the cold double terminator survives as an artefact of a redesign we undid. The other half of the motive stands on its own: module 1 sequenced clean on P_veg and RBS0 and gave no expression signal, and the rebuild is the part of this cycle we kept.

Level 2, and a stoichiometry change

With module 1 rebuilt, the four modules went into Level 2. Three consecutive Golden Gate rounds returned nothing with no design change between them. On 28 August the insert-to-vector ratio was swept at 1:5 and 1:7 and both returned positives; on 1 September two clones were confirmed across three internal junctions at once. Module 2 was built on two promoters and both survived, so the Level 2 assembly was run twice: ACDI carries the P_lepA photoreceptor and ABDI the P_liaG one. Until a graded induction curve exists there is no way to choose between them, so both were carried.

ABDI is the weaker of the two: clone V7 has been checked one junction at a time on separate gels and never all three at once. No Level 2 construct has been sequenced end to end, and the build record’s Level 2 sequencing cell is empty for both devices.

Result

94 mL/min held, the run stopped early on a leak, and 2 mL of shell-side medium plated on LB grew nothing.

Changed next

Biocompatible tubing, a magnetic stirrer, and the sampling port moved to the reservoir outlet.

1. DDesign

   Initially, the objective of our design was to establish a functional continuous flow system utilizing a PES hollow fiber membrane. This preliminary build provided initial insights into structural limitations and potential hardware optimizations. The design of our Bioreactor comprises a PES Hollow Fiber Membrane, medium reservoir, Peristaltic pump, and luer valve, offering a simple, continuous flow-through mechanism.

2. BBuild

   Within this first cycle, we utilized a PES hollow fiber protein membrane (lumen volume: ~200 mL) connected to a 500 mL LB medium reservoir containing 300 mL of LB medium, inoculated with 20 mL of a multi-overnight B. subtilis liquid culture.

3. TTest

   Within early phase testing, the peristaltic pump achieved the target flow rate of 94 mL/min; however, the trial was prematurely terminated due to localized medium leakage within the peristaltic pump. We then evaluated microbial viability, yielding final OD600 readings of 0.304 (in-tube) and 0.48 (medium reservoir).

4. LLearn

   The divergent OD600 values between the tube and reservoir indicated poor microbial viability and pointed to two key hardware limitations for the next DBTL cycle. Exhibiting these factors, we hypothesize that the lack of active mixing led to nutrient stratification and cell sedimentation within the reservoir, limiting growth.

Show details The containment plate, and the three changes this cycle asked for.

To confirm the microbial presence and viability, 2 mL of the medium was extracted from the Shell side of the Hollow fiber membrane and was plated on LB agar to examine colony formation. From this, we observed the absence of colony formation, providing functional proof that our bioreactor is bio-safe for public agricultural usage.

Furthermore, to overcome the observed lack of microbial growth, we plan to incorporate an active magnetic stirring mechanism, inspired by the Biology laboratory Hot plate magnetic stirrer, to prevent sedimentation and ensure a homogeneous nutrient distribution. In addition, to improve the data collection accuracy and minimize contamination risks, we plan to relocate our initial sampling ports directly to the medium reservoir’s outlet within our next DBTL cycle.

In order to address the mechanical leakage, we identified the need for biocompatible tubing (BPT) to ensure higher durability and more stable sealing between transition points of the pump.

What that plate can carry is narrower than the sentence above it. No replicate count was stated for it, and its incubation time and temperature were not recorded. The reactor record now counts two platings in all, the second run beside a lumen-side plate that grew, which is what makes an empty plate mean anything; it states no replicate count either. The barrier is also the only containment stage in the loop, and the 0.2 µm pore that holds cells back passes vesicles and cell debris.

The plate is a standing procedure now, no longer a result. The user manual asks for the shell side to be plated on the first run after every membrane change, and after any run in which the pressure trend changes, and for the harvested fluid to be treated as live culture until that plate has grown nothing. It says in as many words that the membrane separates by size and not by molecule: everything under the pore crosses, the product with it. Containment is the state of the module, the tubing, the connectors and the pump head around it, and it is re-established at each module change; June proved it once, for one module.

Result

The reactor plateaued at OD600 0.8 to 0.9 after about two hours, against the flask’s 0.6 to 0.8 after four and a half.

Changed next

Run it again at 22 °C, on higher-durability BPT tubing.

1. DDesign

   Improving on the previous design, a 200rpm magnetic stir was added to the medium reservoir to prevent sedimentation and ensure a homogeneous nutrient distribution. In addition, the sampling sector was placed closer to the medium reservoir to acquire a more accurate measurement as the reservoir holds much more relative volume compared to the hollow fiber membrane, giving us a more averaged prediction of the total OD of the entire bioreactor module.

2. BBuild

   The same loop with the 200 rpm magnetic stirrer under the reservoir and the sampler moved to its outlet; the drawing below is the team’s own illustration of the prototype.

3. TTest

   As before, the peristaltic pump maintained the target flow rate of 94 mL/min for the entirety of testing. However, instead of experimenting with room temperature, the bioreactor(experiment) and flask(control) were maintained at 37 degrees Celsius. In contrast to our last attempt, this time the experiment ran successfully, helping us secure the full growth curve of our bacteria, with the bioreactor reaching the plateau phase (OD600 ~0.8-0.9) after 2 hours, and the flask reaching the plateau phase (OD600 ~0.6-0.8) after 4.5 hours.

4. LLearn

   The successful generation of the OD600 growth curve confirmed and justified that adding the continuous magnetic stirring mechanism is crucial to preventing sedimentation and maintaining optimal system efficiency.

Result

Steady growth at 22 °C through the photometer, with the leak and degradation problems gone.

Changed next

Instrument the loop: pressure, flow, and a valve that holds when the power is cut.

1. DDesign

   The tubing was upgraded to biocompatible BPT tubing. The system was evaluated at a reduced operating temperature of 22 °C, placing our experimental setup within our plant growth incubator. Furthermore, this test would be utilizing our self-made photometer design.

2. BBuild

   The loop went into the incubator with the printed photometer tower plumbed into the return leg, downstream of the pump, where a longer plumb gives the pump’s sawtooth room to smooth out.

3. TTest

   The run itself is the 400-hour photometer record in the photometer track.

4. LLearn

   Our B. sub is able to grow steadily while passing through our photometer, with our new tubing, at 22 degrees. It was dropped into the reservoir from a large height, creating slight bubbling that helped dissolve oxygen passively. The newly used BPT tubing resolved all leak and degradation issues.

Result

Pressure, flow and a servo pinch valve went on. The magnetic pump and its reservoir were rejected, and neither probe could be got into a sealed vessel cleanly.

Changed next

Run every subsystem at once, on cells.

1. DDesign

   Two questions settled before anything reached the bench. Scale: a hollow-fibre membrane has to run at a wall shear rate of 2,000 to 6,000 s⁻¹, which fixes the cross-flow once the geometry is chosen. And the pinch valve actuator, scored against four criteria agreed before any option was looked at.

2. BBuild

   A pressure transducer on the retentate side on 12 July, a servo replacing the stepper valve, the full rig assembled over four days in July, then a magnetic pump, a cooling reservoir and a cap bored for dissolved-oxygen and pH probes in August.

3. TTest

   Transmembrane pressure was on a screen the same evening the sensor went on, with a heavy pump sawtooth on the trace. The magnetic pump topped out below 200 mL/min against a recommended 263 mL/min and could not lift liquid out of the serum bottle.

4. LLearn

   Autonomy is blocked on sterility, not on electronics. Neither probe body survives autoclaving and the reservoir closes with a rubber septum, and no amount of firmware addresses that.

   The answer in the end was to stop autoclaving the loop. The assembly guide sterilises in place: a litre of 3.5 % hydrogen peroxide circulated for fifteen minutes, drained, then a litre of distilled water at a time until no peroxide is left and the rinse comes back neutral, which in our hands is about four rinses. The photometer is blanked after that and never before. It keeps the probes and the glass out of an autoclave, and it moves the problem without solving it: this is chemical disinfection of an assembled loop, and the manual still treats everything the loop has held as biological waste at the end of a run.

Show details The scale arithmetic, the valve criteria, what each prototype settled, and a documentation failure that changed a process.

The scale question, settled without a test.

	Strategy A, scale up	Strategy B, scale down
Module	63 mm by 300 mm, about 1,300 fibres	10 mm by 600 mm, 11 fibres
Membrane area	1.3 m²	0.02 m²
Cross-flow at 2,000 s−1	16.25 L/min	129 mL/min
Cross-flow at 6,000 s−1	48.74 L/min	388 mL/min
Pump available	1.5 L/min	within reach

No test was run because none was needed: the arithmetic settles it by a factor of ten to thirty, and we scaled down. The valve went the same way. Four criteria were agreed before any option was scored, so they could not be bent toward the answer: does it hold when power is cut, how much holding force against permeate pressure, can we build it from parts already in the lab, and how much firmware does it need. A stepper releases when it loses power and the loop is full of live engineered culture, so the first criterion outranked the other three and decided it alone. We replaced a valve we had already built with a servo.

What each prototype in this cycle changed, and what it settled.

#	Date	What changed	What it settled
III	4 Jul	Module frozen at the small geometry	Scale settled on arithmetic; specification re-checked against invoices
IV	12 Jul	Pressure transducer on the retentate side, 0 to 1 MPa, read through a divider	Transmembrane pressure on a screen the same evening, with a pump sawtooth on it
V	Jul	Servo pinch valve replacing the stepper	Holding on power loss
VI	16 to 20 Jul	Full rig assembled over four days; module characterised on pure water across eight flow settings, twice	The rig ran; the record of building it did not
VII	5 to 23 Aug	Magnetic pump, cooling reservoir, cap bored for dissolved-oxygen and pH probes	One accepted, two rejected; autonomy blocked on sterility

Freezing the specification meant checking every line against invoices instead of against memory, and two of our own recorded claims did not survive. The old specification said the membrane had a 50 kDa molecular weight cut-off and blocked outer-membrane vesicles. The module we actually own is 0.2 µm microfiltration, which passes small vesicles and cell debris along with the roughly 36 kDa ACC deaminase that is meant to pass. The barrier is cellular, not molecular, and we withdrew both claims in writing.

One failure here is a documentation failure and it changed a process. Prototype VI was photographed nineteen times across four days, with gaps, duplicates and no captions, and we cannot reconstruct the assembly order from them. A future team rebuilding this rig would get the bill of materials from Hardware and would not get the order things go together in. The next build is photographed against an agreed list.

The transducers this cycle fitted are the ones the next question needed, and the answer took until October. Transmembrane pressure across a multi-day run was the record’s highest-priority open parameter, because a cake of retained cells and debris on the lumen wall is the most likely limit on service life. The run of 14 to 16 October turned the measurement into a trend: the pressure the module needs beyond what clean water needs at the same flow rose from 2.0 to 3.8 kPa across 42 hours, with two excursions where the permeate line was opened. Nothing had to be cleaned, and one run sets no replacement interval.

Result

OD600 0.83 to 2.45 AU across 53 h 22 min and 6,387 logged samples, with transmembrane pressure flat at −0.0005 bar after 1,132 L through the loop.

Changed next

Both integration cycles were built: v2 put five servo valves and the safety interlocks on, v3 added pH control, the green strip and one operator interface. Still open after them are the probes inside the sterile boundary, and a fouling challenge the October run only begins.

1. DDesign

   This cycle tests integration, not the optogenetic circuit. ACC deaminase is placed under a constitutive promoter so the strain expresses continuously, which removes light control as a variable. Culture circulates through the lumen side, protectant crosses into the shell side, and cells are retained. Everything else in the loop exists to keep that separation running unattended: a peristaltic pump circulates the culture, a medium reservoir feeds it, the in-line photometer reads OD600 without opening the system, and pressure and flow sensors confirm perfusion holds at the set rate.

2. BBuild

   The rig went into the 37 °C incubator with the constitutive strain in a 500 mL reservoir, and every subsystem was switched on at the same time and logged to one file, on a 30 s median window so one row of the log covers the same interval on all three signals.

3. TTest

   Engineered B. subtilis was run in the system and ACC deaminase was secreted across the membrane.

   The blot behind that sentence is in the cloning track, Figure 16: lumen-side cell lysate and medium against shell-side medium, sampled every two hours from 14 to 44 hours, with the ACC deaminase band at about 42 kDa. At that expression level a shell-side band is faint at best.

4. LLearn

   ACC deaminase was successfully secreted to the shell side but expression is weak.

   The blot says it crosses. It does not say how fast, and no shell-side concentration against time has been measured, so the protectant transfer rate is open, and it, not growth, may be what sets throughput. No medium feed was fitted, either, so what was taken off the shell side was never replaced.

Result

Three measured spacings that became the fixed geometry of every later build.

Changed next

The bench layout became a printed instrument in the loop.

1. DDesign

   We used magnetic helping hands, the clip tool that holds parts locked in place and is normally used for soldering, to hold our amber 600 nm LED in line with a purchased paralleling lens. That gives parallel light, which a 2 mm slit then narrows, so only a small part of the beam carries on through the beam splitter to the two lux sensors.

2. BBuild

   We locked the LED, slit, beam splitter and both lux sensors in position relative to each other and left only the paralleling lens free to move. Fixing everything else made the lens position the single variable.

3. TTest

   With the other components locked down, we shifted the paralleling lens along the axis until the light path steadily illuminated both sensors. Doing that meant nudging the slit and the sensors slightly as well, to bring the path into line. Then we measured the distance between every component and turned the layout into a drawing our drylab hardware team could develop into a CAD model with everything anchored in place.

4. LLearn

   We measured the tip of the LED to the focal lens at 5.5 cm, the lens to the center of the beam splitter at 10 cm, and the center of the beam splitter to each sensor at 2 cm.

Result

Unusable. The readings stepped with bubbles and the baseline drifted under loose mounts.

Changed next

Stand the light path up, and tighten the component tolerances.

1. DDesign

   The light path stays flat and horizontal, since the bench arrangement was already on one plane. And every component stays independently adjustable, because we did not yet trust the cycle 1 spacings to be final.

2. BBuild

   The whole optical train mounts on a single printed rail. The flow cuvette is plumbed into the recirculating loop with tubing at both ends, so the instrument reads the culture as it passes through.

3. TTest

   Data fluctuation was way too drastic due to bubbling within the flow cuvette as well as high tolerance for lenses and beam splitters. The draft marks the run itself: actual data not found.

4. LLearn

   Bubbles in the flow cuvette were sitting directly in the beam, and with a horizontal path a bubble had nowhere to go. The independently adjustable mounts were also too loose. Adjustability helped during alignment and hurt everywhere else.

Result

Standing the path up moved the bubbles, and the disagreement with the benchtop instrument turned out to be a procedure, not an optic.

Changed next

Blank both instruments on the same liquid. The protocol that came out of it, and stayed, is to blank on the run’s own growth medium.

1. DDesign

   V2 stands the light path up. The LED fires upward through the flow cuvette, so a bubble in the cell rises along the beam axis and leaves the light instead of settling in it. Buoyancy solves the problem; flow does not come into it.

2. BBuild

   A wide square base plate was printed with a mounting adapter for the rail, so the assembly stands rigidly on the bench with the optical axis vertical. The component seats were tightened so the lens and splitter hold their measured positions.

3. TTest

   Two samples were run and compared against the Bio-Drop. The reference channel held at 92.5 lux across both readings. That tells us the source was stable between measurements, so the drop on the sample channel came from the sample and not from the LED.

4. LLearn

   We had been blanking ReLEAF against growth medium. The Bio-Drop is calibrated on distilled water. The roughly 20% disagreement we had been attributing to our optics was largely a procedural artifact.

   The protocol that came out of it is in the user manual, and it goes the other way: blank on the growth medium the run will use, and blank the reference spectrophotometer on the same medium. What matters is not which liquid, it is that both instruments see the same one; the manual names blanking them differently as the commonest cause of disagreement between the two. Changing the sensor settings, the estimator or the path-length factor voids the blank, and an old blank is flagged for the operator, who decides whether to take a new one.

Result

The two titanium dioxide concentrations separate cleanly, the difference from the Bio-Drop falls from about 20 % to 8 %, and the reference channel moves 6.4 % between blank and sample.

Changed next

Enclose the light path, drive the LED directly, and put real culture through it.

1. DDesign

   Tilting the whole instrument means no surface inside the cuvette is horizontal, so a bubble has a continuous upward path along the wall and nothing flat to rest against. The tilt went into a wedge under the body, not into the optics.

2. BBuild

   The body sits on a printed wedge with structural support underneath so it does not flex under its own weight. A dual luer valve with syringes is fitted to the loop, used to bleed gas out before the instrument is started.

3. TTest

   The sample channel dropped from 137.63 lux on the blank to 120.44 lux at 1 mL and 111.62 lux at 2 mL, giving readings of 0.3712 and 0.5785 au. The reference channel sat at 17.23 lux on the blank and 18.33 lux on both samples.

4. LLearn

   The reference channel is the bigger finding. It moved 6.4% between the blank and the samples, and the reference beam is split off before the sample, so nothing in the cuvette should be able to change it. A reference channel we cannot trust is a reference channel we cannot correct with.

   Either the source drifted or the reference path is picking up light it should not see. Room light is the likeliest version of the second: at this build the path is open to the bench and neither sensor is shielded from it, which is what puts covers on V4.

Result

An unbroken run past 400 hours on wild-type B. subtilis 168 at 22 °C, with nothing withdrawn from the loop.

Changed next

Calibrate the instrument against a reference, on a standard that does not grow.

1. DDesign

   Component positions are fixed at their default alignment; every mount surface that touches glass is lined with TPU; and snap-in covers run the full length of the light path from LED to sample sensor, so no room light reaches either detector. A photometer that responds to the lights being switched on is not measuring the sample.

2. BBuild

   The instrument runs on an Arduino UNO Q. Both lux sensors report over the same bus. The interface is a live web dashboard served on the local network, so the run can be watched from any machine without touching the rig.

   That build is the Pro, and it is the one the assembly guide describes: a 600 nm LED without an internal resistor, a plano-convex paralleling lens, a 90:10 splitter cube, the 2 mm flow cuvette, and two GY‑302 sensors that share one bus because the board sets each one’s address, the sample at 0x5C and the reference at 0x23. Eight parts print in black PLA, and the seats that hold glass are painted in TPU so the lens, the cube and the cuvette bed into TPU instead of bare PLA.

3. TTest

   A continuous run of wild-type B. subtilis 168 at 22 °C, drawn from the bioreactor loop and passed through the flow cell in line, so no sampling event interrupted the culture and no material was removed to measure it. The run passed 400 hours.

4. LLearn

   Noise is tight below about 0.4 OD and ragged above about 0.8, so the useful precision is not flat across the range. The rejection rate moved from 42.9% to 14.3% to 0% within one run, and the rejection rule is not yet written down anywhere.

   It is written down now, and it has moved on since the run. The specification in the user manual takes four dark readings, turns the source on for 60 ms and takes sixteen lit readings in each window; a reading counts as rejected when it falls more than 5 % below the upper envelope of its burst, and above 50 % rejected the whole measurement is discarded and retaken at once. A second gate sits after it, which refuses a reading whose dispersion or whose jump from the last accepted value is too large and reports the median of the last three accepted values instead. The run’s own three rates belong to the older fourteen-reading burst, where 42.9 % and 14.3 % are exactly six and two readings out of fourteen, so they are a count and not a drift over the run.

Result

Raw output agrees with the benchtop instrument to a median paired difference of 3.8 %, and both track the mass standard at R² above 0.999.

Changed next

Repeat on cells, extend the range, and confirm against the iGEM silicon dioxide standard.

1. DDesign

   Using culture to calibrate would lead to issues in growth time, settling, or cell lysis, leading us to come up with TiO₂ suspension as a stable scattering target. The iGEM standard is done with SiO₂, but TiO₂ performed nearly exactly the same suspended in 1% SDS solution and was much cheaper.

2. BBuild

   The calibration loop consisted of a magnetic stirrer, a peristaltic pump, our photometer, and a 500 mL reservoir, whose cap had holes for the inlet, the outlet, a sampling port, and an air hole that prevented a vacuum forming during sampling.

3. TTest

   Plotted against mass added, both instruments track the standard at R² > 0.999 (ReLEAF 0.9999, Bio-Drop 0.9998), and plotted against each other the raw points fall along the line of identity, agreeing to a median paired difference of 3.8%.

4. LLearn

   We originally expected the ReLEAF readings would need a calibration transform; instead the two runs showed that no calibration is needed. Fitting our readings to the Bio-Drop would tie them to the reference instrument’s own measurement error. A calibration step could make the readings worse.

   The decision survived into the shipped instrument, and the manual says what it costs: no calibration curve against a reference spectrophotometer is applied, so the optical density the reactor logs is raw output, linear across the working range and not traceable to any reference instrument. It also names where that stops holding. At high cell density the source and the sensor are not linear together, so the top of the range has to be checked with a dilution series and never assumed.

Result

Optical density, pH, pressure and flow read live, with the pinch and Luer valves commanded from the same interface.

Changed next

A live record only reports the present, so the next cycle gave the software a model of a healthy run.

1. DDesign

   A bioreactor standing beside a rice field has nobody on hand to draw a sample, so every channel an operator would read by hand had to come off the vessel on its own.

2. BBuild

   Sensors and valves wired to one board that logs every reading, with the interface showing the four channels live and sending pump and valve commands back over the same link.

3. TTest

   We ran B. subtilis six times in the bioreactor, the longest continuously for 434 hours, and kept every log.

4. LLearn

   Readings alone answer only what the culture is doing now. Telling drift from ordinary variation needs something to measure it against.

Result

Replayed over the long run, the alarm flagged the photometer at 216.3 h, 26.7 h before the optical density channel failed at 243.0 h.

Changed next

An alarm still reports what has happened, so the next cycle asked the software to look forward.

1. DDesign

   The long run drifted twice for different reasons, so the alarm scores two axes separately: the reference beam, which never passes through the culture, and the growth rate.

2. BBuild

   We trained an autoencoder on the healthy first 120 hours of the 434-hour run and replayed the rest of the log through it.

3. TTest

   Between 220 and 240 h the reference beam stepped down while the optical density still looked clean: the sensor axis alarmed on every window and the process axis stayed quiet.

4. LLearn

   The separation works, and it worked on a finished log. The models are replayed after each run, and a person still makes every decision.

Result

The Kalman filter beat persistence at every horizon, cutting the two-hour forecast error by 41%.

Changed next

Every candidate was fitted and scored on the same run, so the next cycle registered a prediction before the run that tested it.

1. DDesign

   The grower needs to know how much biomass there will be in a few hours. That means a forecast, and a baseline honest enough to beat: persistence, which predicts no change.

2. BBuild

   Kalman filter, local linear, global linear and exponential forecasts, each made every hour from 12 h and using only the data logged up to that point.

3. TTest

   Over 102 hours of perfusion, optical density rising from 0.84 to 3.77, the two-hour error fell from 0.067 to 0.040 OD600 against persistence.

4. LLearn

   Global linear weighs early data as heavily as recent, so it points the wrong way when growth changes speed, and exponential overshoots. Weighing each reading against the noise kept the filter on real changes.

Result

The rate fitted to the October run landed inside the interval registered before it.

Changed next

The loop is still open: the models replay logged data and a person makes every decision.

1. DDesign

   A forecast scored on the run it was fitted to proves little, so we registered the prediction and its pass criteria before any run of the new strain.

2. BBuild

   The prior is a logistic fit from the light plate, registered on 1 October 2026: 0.2986 h⁻¹, 90% interval 0.2705 to 0.3300.

3. TTest

   The first such run broke down after ten hours, so the forecast was scored on the 14 to 16 October run: 0.286 ± 0.010 h⁻¹.

4. LLearn

   The rate crossed from a plate well to the bioreactor; the ceiling did not, and the secondary endpoints stay unscored for want of a cross-calibration between the two scales.

   Green light, dissolved oxygen and the field sensors are tested separately and are not in the loop. The Ask tab is switched off here, since the wiki allows no outside request.

Result

The concept holds. Nothing about it is quantitative: output was uncontrolled and external light was not blocked.

Changed next

Measure every LED, sort them into matched groups, and shield the cultures.

1. DDesign

   Put a green and a red LED in each column so we can switch between them quickly and exercise the on and off control of protectant production directly. Nothing is calibrated at this stage, because the question is whether the concept works at all before we spend effort making it precise.

2. BBuild

   Dual LED per column, allowing for quick adjustment from Green to Red light output. Wire management prints, included from the start, since even at this size the harness gets unmanageable fast. The array was driven straight from a breadboard.

3. LLearn

   Output consistency between LEDs was not controlled. External light was not blocked.

   Internal-resistance variance is a common problem in cheap single-bulb LEDs: nominally identical parts draw different currents and emit different intensities at the same drive level. As long as that is true, an intensity tier is not a real quantity, and the wavelength comparison is not fair either, since the green and red channels would be running at unknown relative brightness. Ambient light adds an unknown amount of illumination to every tube, and it lands hardest on the low-intensity condition, which is exactly where the separation has to be cleanest.

Result

Green ran from 2.01 to 2.54 kLux and red from 360 to 438 lux, spreads of 26 % and 22 %. Sorted into fours, each group holds to about 1 to 2 %.

Changed next

Set the tiers by pulse-width modulation, and enclose the electronics.

1. DDesign

   The approach is to measure the light output of every LED individually and use that to identify units with similar internal resistance. From the matched units we form three groups of four per wavelength, one group per intensity tier, arranged as columns. Because the LEDs within a column are matched, the column can be recalibrated as a unit through straightforward mathematical adjustment, so that every column at a given tier matches every other: that is what turns low, medium and high from labels into reproducible quantities. Coverage was added to block external light, so that the intensity a culture receives is the intensity we set.

2. BBuild

   Every LED was driven at 5 V and measured individually on a lux meter using a printed sensor fixture, so each reading is taken at the same distance and orientation. Each LED carries a tag made of its row and its column, rows A to D running downward and columns 1 to 5 across, so any physical LED can be traced back to its own measurement. The array was then rewired with the matched units grouped into columns.

3. TTest

   All twenty green LEDs were read in kLux and all twenty red in lux, at 5 V, one at a time, and the grids are in the details below, laid out by measurement tag and not by the array’s own 4 × 6 grid. Sorting those into groups of four by output gives the matched sets the tiers are built from, and the six channels were read again after rewiring.

4. LLearn

   Sorting by measured output tightens each group considerably: the green low group holds 2.33 to 2.34 and the green mid group 2.35 to 2.37, well inside 1%, and the red groups hold to about 2% each. But the separation between tiers is smaller than we need. Green low to green high is about 4% and red low to red high about 10%. Those gaps are the parts’ own spread after sorting; they are not three distinct light levels. The tiers therefore have to come from the PWM duty cycle rather than from LED selection: matching makes the four LEDs in a column agree with each other, and software sets how bright that column runs.

   Matching is a build step now, not a one-off: buy twenty LEDs of each colour, read every one at 5 V through a 10 mm aperture taped over the meter, label each with its reading and keep the closest twelve. The guide asks the four LEDs of a group to read close together and records what our own build managed, 0.4 % to 1.8 %; the spare eight of each colour are what makes that possible.

Result

An instrument that can be handed to someone else: enclosed electronics, code-set intensity, and a holder that fixes the distance between LED and culture. The overnight run on a shaker snapped a post off the bottom plate, pulled wires loose, and leaked light between the wells.

Changed next

Repeat that overnight run on the repaired base, measure photon flux, not lux; check uniformity between wells; run the circuit under it.

1. DDesign

   An outer frame print houses the LED wiring and the perf board, so no electronics sit loose on the bench. Light intensity is varied in code, so a run is defined by a set of values, with nothing for anyone to remember to adjust. Slider prints carry the LED holder and the tube holder, so the stack comes apart for rewiring. A tall print holds the test tubes at a fixed position over the array, so the distance between LED and culture is the same every run.

   Six parts are printed, all in black PLA Matte on the same profile, and the tube holder is printed at four wall loops so that no light crosses between wells. A screwdriver is the only tool the assembly itself needs; the soldering iron is for the four M3 heat-set inserts that take the perf board.

2. BBuild

   Control runs from an Arduino Uno R4 WiFi. PWM signals from the Arduino drive three MOSFETs, one per intensity tier, and software sets the duty cycle for each tier. The selected intensities are held constant for the duration of a test, so each culture receives a fixed, well-defined dose; three calibrated tiers across two wavelengths give six defined light conditions per run.

   The switching stage moved with the build. In cycle 2 the three MOSFETs stood on a breadboard on the bench beside the LED holder; here they sit on the perf board underneath it, with the housing over them. Each tier is one IRLZ44N on its own Arduino pin, and the run sketch writes one value per pin and never changes it again, so the dose a tube receives is set by three numbers in the source.

   The three tiers as the run sketch sets them, rounded. Each tier is a green column and a red column of four tubes.

   Tier	Pin	Written	Duty
   Low	D9	5	2 %
   Medium	D10	31	12 %
   High	D11	255	100 %

3. TTest

   During our overnight test run, the connection point on the bottom plate broke off. The LPA was on a shaker for the run; the frequency was most likely too high for the printed post. Several wires came undone and had to be resoldered. The white tube holder print also allowed light to leak into neighbouring holes, which compromised our data, as the test tubes were receiving light from more than one source.

4. LLearn

   All four screws on the bottom plate are now used, not two, which spreads the load across four points instead of concentrating it on a pair, and every soldered connection was reinforced. Neither change has been through a second overnight run on the shaker, so the repair is untested. The white tube holder was replaced with a black one, and its wall thickness increased from two loops to four.

   Both repairs are the build now: the assembly guide screws the base plate to the housing from underneath with four #10‑32 screws through the middle of each edge, and prints the tube holder in black at four wall loops. Three bench tests go with it, one after the LEDs are soldered, one after the MOSFETs are wired, and the second one again once the stack is screwed together, so a column that stops answering its MOSFET is caught before a culture is in the tube. None of that is the overnight run, which is still the test this cycle owes.

Result

Three boxes tried and dropped: a 21.1 × 14.1 × 11.4 cm opaque tub, a 7.5 cm bought box that leaked, and a 6.5 × 6.5 × 10 cm clear box too narrow to grow in. A 1000 µL tip box replaced all three.

Changed next

Keep a clear tank and find the darkness somewhere else.

1. DDesign

   The vessel has to hold enough ½ MS medium to last a run, keep light off the roots, leave board area for as many seed positions as a run needs, and come back from the autoclave intact so it can be used again.

2. BBuild

   Prototype 1 was a large opaque tub, 21.1 × 14.1 × 11.4 cm, filled with about 500 mL of ½ MS medium under a foam board cut from sheet and taped black, carrying five agar-containing seed holders. Prototype 2 was a small bought box, 7.5 × 7.5 × 5.3 cm, with holders for the lids and labels on the box lid. Prototype 3 was a tall clear box, 6.5 × 6.5 × 10 cm, with a square black-taped foam board.

3. TTest

   The first box wasted most of what it offered: the board had room for more seed holders than the five we used, the foam was too fragile to reuse, and cutting and taping it was slow and hard to scale up. The depth of the box made the root system difficult to observe, and the black wall made it difficult to tell when the medium went turbid. Prototype 2 leaked from two sites when the box was moved, the joint and the bottom, and black tape over the hole did not stop it; the lid fell into the box unless it was placed at a specific angle. Prototype 3 held opacity and depth together, and then the foam board fitted so tightly that it was hard to reach when the medium was changed, and the box was too narrow for the plants to grow and limited how many we could run in it.

4. LLearn

   A 1000 µL tip box answered all of it at once. It is 12.8 × 10.1 × 9.5 cm, which does not take up much space and still gives the roots room; it is transparent, so the root system and the turbidity of the medium can both be read off the wall; it survives cleaning and autoclaving, so it is reused; and one board in it carries anything from 5 to 13 holes. Contamination was less frequent than in the two boxes before it.

   What a clear wall gives up is the darkness the roots need, and that debt is paid in cycle 4 from outside the tank.

Result

Foam boards for every box, and three plates cut, taped and drilled at 300 mL, 1.5 L and 2 L. The level was set by hand before each run and topped up during it, and the foam went soft and broke in the water.

Changed next

Print the board, and put the flotation inside it so there is no level left to set.

1. DDesign

   Labs that grow Arabidopsis in liquid medium improvise the holder: a piece of foam cut to size, a mesh taped over a container, or a fixed plate sitting above a reservoir. Conn and colleagues set out the version we started from, seeds germinated on agar in punched tube lids and then moved onto a board. We did the same, and sized a plate for each of the three box volumes we had on the bench.

2. BBuild

   Foam cut to the box, taped black, and drilled on a grid of seed positions; the agar-containing seed holders dropped into the holes. Three plates, for the 300 mL, 1.5 L and 2 L boxes, each hand-labelled in red tape. The medium sat underneath at a level set by eye before the run.

3. TTest

   The level turned out to be a narrow window. It has to sit just under the board and as high as it can be got: too low and the roots lose contact, too high and the board floods. And it does not stay where it is put. The plants drink and the medium evaporates, so a level that was right at the start of a run is wrong by the end of it.

   The foam failed separately. It became soft after it stayed in the water for a long time, and a softened board breaks in the middle of changing the medium.

4. LLearn

   Two faults with one answer each. Somebody has to watch the reservoir and top it up, and every top-up is a chance for conditions to drift between replicates, which is the variable the board was there to remove in the first place. And a board cut, taped and drilled for one box is not a part another lab can make the same way twice. So print the board, and stop setting a level at all.

Result

A printed plate that floats on its own frame. The calculation did not get the freeboard right; washers did.

Changed next

Run the displacement calculation again on every design change until the poles are not doing anything.

1. DDesign

   Move the flotation into the frame: sealed air chambers that are part of the structure, with nothing to glue on and nothing to come loose in the medium. Sizing them is a displacement problem. The plate has to displace its own mass, plus the seedlings, plus whatever water the printed part takes on, while sitting at the freeboard that puts the roots in the medium and the crowns above it.

2. BBuild

   Printed in PLA on a standard FDM machine, pontoons hollow at 15 % gyroid infill with a teardrop cross-section that prints without support. A pole at each corner took washers, so the freeboard could be trimmed after printing without reprinting the whole part. A handle went on each side of the board, so the plate could be lifted out to change the medium without anyone touching the seedlings.

3. TTest

   It floats, and the corner poles get used. One worked case from the drafts has the plate at roughly 180 g, where the pontoons came up short by about 10.7 cm³ of volume each. The printed plate was never weighed, so the case cannot be checked against the part that exists.

4. LLearn

   The poles are the safety margin on a calculation we did not trust yet. They are also four parts, four washers and a decision per run. Every change to wall thickness or footprint moves the mass and the displacement at the same time, so the calculation has to be rerun on every design change, not once at the start.

Result

Sixteen printed parts, 42.27 cm³ of sealed air, 3.5 g of added load per millimetre of draft, and the tip box sitting inside a 14.7 × 11.5 × 6.2 cm light-blocking box.

Changed next

Weigh an assembled plate and read its freeboard off the load line, which is the one measurement that would close this track.

1. DDesign

   Smaller footprint, clip-in handles so the plate lifts out of the reservoir without anyone touching the seedlings, and thirteen seed positions in rows of four, five and four. The buoyancy is computed from the STLs, not estimated: the two chambers hold 42.27 cm³ of sealed air, and down the pontoons’ straight sides the hull gains 3.51 cm³ per millimetre of draft.

   

2. BBuild

   Sixteen parts in PLA, no supports, no adhesives and no fasteners. As drawn, the waterline sits 6.0 mm under the slab, a freeboard of 9.2 mm, where the assembled plate displaces 54.2 cm³.

   The tip box then goes inside a black light-blocking box of 14.7 × 11.5 × 6.2 cm with a notch cut in one side. Keeping the tank dark holds back algal growth, and the clear inner box still lifts out when we want to look at the root system.

3. TTest

   The plate carried the salinity and heat assays on Plants: four boxes side by side on four identical plates, photographed from above through the run. No run on this plate has had its freeboard measured.

4. LLearn

   Every 3.5 g added sinks it about 1 mm, so the sensitivity is known and the load line is drawn; what is missing is one weighing. Printed solid, an upper bound, the sixteen parts come to 92.6 g and would float with 1.9 mm of freeboard, which is the number that retires the 180 g worked case still in the drafts.