Context Document

Download .md file

PhaC Enzyme Engineering — LLM Context Document

Version: v2.2 Date: 2026.05.12 Engineer: JS Project goal: Engineer PhaC_Cn for increased PHB production.

Dataset scope note: This document is built around a single reference enzyme (C. necator PhaC1) and published mutagenesis studies on that enzyme and its close variants. It does NOT use diverse multi-species sequence alignments. See Section 3 for implications and compensating strategies.

0. IMPORTANT

0.1 General

• All sequences given or referred to here are amino acid sequences.
• We are using the wild-type enzyme PhaC from Cupriavidus necator (PhaC_Cn) as the starting point.
• PhaC is a polyhydroxyalkanoate synthase enzyme that polymerizes monomers into polyhydroxyalkanoate polymers.
• PhaC_Cn preferred product is poly-3-hydroxybutyrate (PHB), which uses monomer 3HB.

0.2 Notation

• PhaC_Cn is the wild-type sequence from Cupriavidus necator, also called Cn PhaC1.
• All mutations are notated in the form of AXB, where A is the single letter code for an amino acid in PhaC_Cn, X is the amino acid position index number in PhaC_Cn, and B is the single letter code for the amino acid substituted in this mutation. Example: A510T is the C. necator wild-type amino acid sequence for PhaC with a single amino acid substitution from alanine to threonine at position 510.
• C. necator is sometimes referred to as Ralstonia eutropha.
• PHB is poly-3-hydroxybutyrate, sometimes also called polyhydroxybutyrate or poly[(R)-3-hydroxybutyrate].
• Both the single letter codes and three letter codes for amino acids are used throughout this document.
• PhaC_Cs is the wild-type sequence from Chromobacterium sp. USM2 (Class I, 42% pairwise identity with PhaC_Cn).
• PhaC_Ps is the wild-type sequence from Pseudomonas sp. 61-3, also called Ps PhaC1 (Class II, 67% pairwise identity with PhaC_Cn).
• PhaC_Ac is the wild-type sequence from Aeromonas caviae (Class I, 37% pairwise identity with PhaC_Cn).

1. Enzyme Family Background

1.1 Classification

• There are four classes, but we are not considering Classes III and IV at this point.
• Class I and II share ~50% sequence identity; Class III/IV are more distantly related.
• This project focuses exclusively on Class I, using Cn PhaC1 as the sole reference

1.2 Reaction chemistry

• Catalyzes polymerization of (R)-3-hydroxyacyl-CoA thioesters into PHA
• Ping-pong (double displacement) mechanism:
  1. Acylation: acyl group transferred to catalytic Cys, CoA released
  2. Transacylation: acyl group transferred to growing polymer chain
• Lipase-like α/β hydrolase fold
• Catalytic triad: Cys – His – Asp
  • C. necator PhaC1 (Cn) reference numbering: C319, D480, H508
  • All residue positions in this document use Cn PhaC1 numbering unless noted

1.3 Substrate scope terminology

• We are primarily interested in scl, specifically in 3HB because my hypothesis is that increasing substrate specificity for 3HB will increase PHB production.

1.4 Why substrate specificity is structurally interesting

• Substrate-binding tunnel geometry determines acyl chain length tolerance
• Residues within ~5–10 Å of catalytic Cys are primary selectivity determinants
• mcl selectivity often results from removal of steric clash (smaller residues), not addition of new contacts — counterintuitive but well-supported
• Electrostatic environment affects CoA-thioester positioning
• Dimerization interface indirectly influences active site geometry (Class I/II enzymes)
• N-terminal domain is not well conserved; suggested to possibly be involved in substrate selection.

2. Structural Information

2.1 Experimental structures for reference

2.2 AlphaFold model for Cn PhaC1

• UniProt accession: P23608
• Overall pLDDT: 85.94
• Confidence notes: high confidence in core domain, low in N-terminal region residues 1–61
• Use AF model for: loop conformations, surface regions not in crystal structure
• Prefer crystal structure (5T6O) for: active site geometry, tunnel dimensions

2.3 Key structural regions (Cn PhaC1 numbering)

2.4 Substrate-binding tunnel residues

(Fill this table carefully — it is the core of your structural reasoning)

How to fill this table: Open 5T6O in PyMOL or ChimeraX. Select C319. Run: select tunnel_res, byres (all within 10 of resi 319). List those residues here with distances. This is worth spending 1–2 hours on — it will substantially improve LLM reasoning quality.

NOTE : fill section 2.5 when you have the time to go through PyMol. 2.3 and 2.4 were filled from literature.

2.5 Tunnel geometry notes

• Estimated tunnel constriction in WT Cn PhaC1: ~[X] Å (from structural analysis)
• Residues that form the constriction point: [list]
• Estimated minimum cavity volume for 3HHx-CoA accommodation: [X ų if known]
• [Add MD simulation or docking results here as they become available]

3. Dataset Scope, Limitations, and Compensating Strategies

This section is critical. Read before every LLM session.

3.1 What your dataset contains

• Reference enzyme: C. necator H16 PhaC1 (wild-type)
• Variants: Published point mutants from the mutagenesis literature; published point mutants predicted to impact substrate specificity from PhaC_Cs and PhaC_Ac mutagenesis data - these mutants have been formatted into PhaC_Cn sequence through alignment identification.
• Labels: Substrate incorporation data from those studies
• What it does NOT contain: Homologous PhaC sequences from other species, Class II sequences, or unlabeled natural variants

3.2 Implications and honest limitations

3.3 What this dataset IS good for

• The LLM can reason very effectively about:
  • Mechanistic hypotheses — why does mutation X change specificity, based on structure and chemistry?
  • Interpreting your experimental results — what does an unexpected outcome tell you about the mechanism?
  • Experimental design — which mutations to test next given what is known?
  • Identifying gaps — which positions have never been mutated but are structurally important?
  • Literature synthesis — connecting observations across papers into a coherent mechanistic model

3.4 Compensating strategies

To partially offset the lack of multi-species alignment data:

1. Lean heavily on structural reasoning (Section 2) — fill in the tunnel residue table as completely as possible; this replaces alignment signal as your primary source of positional hypotheses

2. Include Class II reference data explicitly — even if not in your training set, you can add a “comparative note” section describing which Cn PhaC1 positions correspond to Class II residues (from manual alignment of just Cn PhaC1 vs. Ps PhaC1). This gives the LLM evolutionary context without requiring a full MSA.

3. Weight negative results equally to positive — if you can find papers reporting mutations that failed to shift specificity, record them in Section 4.3. They are highly informative and rare in the literature.

4. Be explicit about data gaps in prompts — tell the LLM “position X has never been mutated in the literature” so it flags its reasoning as structural/hypothetical rather than evidence-based.

5. Use the LLM to propose positions to structurally analyze — ask it which tunnel residues it would prioritize examining in the crystal structure, then verify those manually before including them in subsequent prompts.

3.5 Class II reference comparison

(Manual alignment of just Cn PhaC1 vs. one Class II enzymes — fills in some evolutionary context without a full MSA)

How to fill this: Use a pairwise alignment tool (e.g. EMBOSS Needle at https://www.ebi.ac.uk/Tools/psa/emboss_needle/) with Cn PhaC1 (UniProt P23608) and Ps PhaC1 (UniProt Q9Z3Y1). This takes ~10 minutes and is worth doing. I used Benchling

4. Experimental Mutation Database

(This is the heart of your dataset — populate as completely as possible)

4.1 Key literature to mine for Cn PhaC1 mutations

• Tsuge et al. (2003) Macromolecules — F420 region, systematic Class I
• Amara et al. (2002) — systematic Class I mutagenesis panel
• Rehm et al. — early mechanistic mutagenesis
• Nomura et al. — broad-specificity engineering attempts
• Insomphun et al. — 3HHx incorporation focus
• Hiroe et al. — combinatorial mutagenesis
• [Add others as you find them — search PubMed: “PhaC mutagenesis” OR “polyhydroxyalkanoate synthase substrate specificity”]
• Taguchi et al (2002) https://academic.oup.com/jb/article-abstract/131/6/801/785238?redirectedFrom=fulltext
• https://www.frontiersin.org/journals/bioengineering-and-biotechnology/articles/10.3389/fbioe.2021.627082/full
• https://openaccess.wgtn.ac.nz/articles/thesis/Toward_Engineering_the_Substrate_Specificity_of_a_PHA_Synthase_PhaC_/17152079?file=31714520
• https://onlinelibrary.wiley.com/doi/abs/10.1002/mabi.200400075
• https://journals.asm.org/doi/full/10.1128/aem.00564-13
• https://www.sciencedirect.com/science/article/pii/S0021925820346068
• https://www.sciencedirect.com/science/article/abs/pii/S1369703X08000909

Mining tip: For each paper, extract: (1) every mutation tested, including ones with no effect — these are just as valuable, (2) exact assay conditions, (3) quantitative data where reported. Even a table footnote saying “A300G showed no change in specificity” belongs here.

4.2 Gain-of-function mutations (change in substrate specificity or activity)

4.3 Neutral mutations (no significant effect on specificity)

(Underrepresented in literature but critically important — record every instance you can find)

4.4 Deleterious mutations (loss of activity or expression)

4.5 Combinatorial / double mutants

4.6 Thermostability mutations

(Relevant when stacking specificity mutations)

4.7 Positions that have NOT been mutated in literature

(Fill as you read — these are candidate positions for novel exploration)

4.8 Data quality notes

• Substrate specificity data given qualitative only due to differences in experimental conditions
• In vitro CoA-release assays (DTNB) give intrinsic kinetic data but don’t fully reflect in vivo selectivity under substrate competition
• Some older studies used racemic substrates — stereospecificity may confound apparent chain-length specificity
• [Add specific notes about inconsistencies you notice across papers]

5. Your Starting Enzyme (Wild-Type Cn PhaC1)

5.1 Identity

• Organism: Cupriavidus necator H16
• UniProt accession: P23608
• PhaC class: I
• Gene: phaC1 (in pha operon: phaCAB)
• Full sequence length: 589 aa

5.2 Known properties of WT Cn PhaC1

• Native substrate preference: scl — 3HB (major), 3HV (minor), very minimal mcl - 3HO, 3HDD
• Specific activity: [X nmol/min/mg — fill from literature or your own data]
• Thermostability: [Tm or optimal temperature]
• Expression in E. coli: [your experience — yield, solubility]
• Dimerization: active as dimer; monomer is inactive
• Known issues: [e.g. requires careful lysis conditions, prone to aggregation at high concentration]

5.3 Full WT sequence

MATGKGAAASTQEGKSQPFKVTPGPFDPATWLEWSRQWQGTEGNGHAAASGIPGLDALAGVKIAPAQLGDIQQRYMKDFSALWQAMAEGKAEATGPLHDRRFAGDAWRTNLPYRFAAAFYLLNARALTELADAVEADAKTRQRIRFAISQWVDAMSPANFLATNPEAQRLLIESGGESLRAGVRNMMEDLTRGKISQTDESAFEVGRNVAVTEGAVVFENEYFQLLQYKPLTDKVHARPLLMVPPCINKYYILDLQPESSLVRHVVEQGHTVFLVSWRNPDASMAGSTWDDYIEHAAIRAIEVARDISGQDKINVLGFCVGGTIVSTALAVLAARGEHPAASVTLLTTLLDFADTGILDVFVDEGHVQLREATLGGGAGAPCALLRGLELANTFSFLRPNDLVWNYVVDNYLKGNTPVPFDLLFWNGDATNLPGPWYCWYLRHTYLQNELKVPGKLTVCGVPVDLASIDVPTYIYGSREDHIVPWTAAYASTALLANKLRFVLGASGHIAGVINPPAKNKRSHWTNDALPESPQQWLAGAIEHHGSWWPDWTAWLAGQAGAKRAAPANYGNARYRAIEPAPGRYVKAKA

5.4 Substrate-binding pocket region

(Residues 300–340, centered on C319 — paste into prompts as needed)

Residues 300–340:
           AIEVARDISGQDKINVLGFCVGGTIVSTALAVLAARGEHPA
Position:  300                319                  340
                              ^C319 (catalytic)

5.5 Catalytic and key residue positions (for quick reference)

6. Engineering Target

6.1 Primary goal

[State precisely, e.g.:]

Increase substrate specificity towards 3HB specifically or scl generally with combinatorial mutations in the N-terminal and elsewhere, with the hypothesis that this will increase overall PHB production.

6.2 Secondary goals

• Avoid total loss of activity

6.3 Acceptable tradeoffs

• Up to 30% reduction in activity acceptable

6.4 Hard constraints — DO NOT VIOLATE

• Do NOT mutate catalytic triad: C319, D480, H508
• Avoid dimer interface residues: 70-88

6.5 What has already been tested

(Update after every experiment round — prevents redundant suggestions)

7. Production and Assay Context

7.1 Expression system

• Host: Cell-free expression, E. coli BL21(DE3) lysate
• Vector: pTwistChlor-HighCopy
• Expression conditions: unpurified cell-free reaction
• Typical soluble yield:
• Purification: none

7.2 Eventual In vivo PHA production conditions

• Host: Cyanobacterium aponium sp. UTEX 3222
• Co-expressed pathway genes: native PHA biosynthetic genes
• Carbon source(s): atmospheric carbon dioxide
• Growth conditions: TBD
• PHA content range (WT): unknown

7.3 In vitro activity assay (if used)

• Assay type: visual inspection for insoluble PHB granules
• Substrate(s): 3HB-CoA
• Buffer conditions: HEPES-KOH pH 7.5

7.4 PHA analysis

• Extraction method: TBD
• Monomer analysis: GC-MS for identification]
• Quantification standard: LC-MS
• Throughput: TBD

8. Reasoning Guidelines for LLM

8.1 Dataset context — tell the LLM explicitly at session start

Always include this statement at the top of each session prompt:

“My dataset consists only of C. necator PhaC1 (wild-type) and published point mutants of this single enzyme. I do not have a multi-species alignment. All positional reasoning should be grounded in (a) the experimental mutation database in Section 4, and (b) structural analysis of PDB 5T6O / AlphaFold model P23608. Do not infer specificity determinants from phylogenetic patterns — that data is not available.”

8.2 Prioritization criteria (in order, adjusted for this dataset)

1. Direct experimental evidence — mutations in Section 4 with measured outcomes
2. Structural/mechanistic reasoning — based on 5T6O crystal structure and tunnel geometry (Section 2)
3. Analogy to Class II — using the pairwise comparison in Section 3.5, noting explicitly when this is being used
4. Chemical intuition — physicochemical rationale for a substitution, flagged as [SPECULATIVE] if no structural or experimental support

8.3 Required output format for mutation suggestions

For every suggested mutation, provide:

• (a) Mutation in standard notation (e.g. A149F, Cn PhaC1 numbering)
• (b) Primary evidence basis: Experimental / Structural / Class II analogy / Chemical intuition [SPECULATIVE]
• (c) Mechanistic rationale — specific, not generic
• (d) Consistency with existing data in Section 4 — does it contradict anything?
• (e) Confidence: High (direct experimental support) / Medium (structural + analogy) / Low (chemical intuition only)
• (f) Predicted risk: stability, expression, activity loss

8.4 Reasoning I do NOT want

• Statements like “this position is conserved in mcl enzymes” — you do not have alignment data to support this; use only the pairwise comparison in 3.5
• Quantitative predictions of mol% outcomes
• Suggestions violating hard constraints in Section 6.4
• Suggestions already in Section 6.5 “already tested” table
• Filling data gaps with plausible-sounding inventions — flag uncertainty explicitly

8.5 Especially useful prompts for this dataset type

Given the single-enzyme focus, these prompt types will be most productive:

• Gap analysis: “Which tunnel-lining residues (Section 2.4) have never been mutated in the literature (Section 4.7)? For each, give a structural rationale for whether they are likely to affect specificity.”

• Mechanistic interpretation: “Mutation X gave unexpected result Y. Given the structural context of position X (distance to C319, neighboring residues, tunnel role), propose 2–3 mechanistic explanations.”

• Epistasis prediction: “Given that A149F and S325A are individually beneficial, reason about whether their combination is likely to be additive, synergistic, or antagonistic, based on their structural relationship.”

• Experimental prioritization: “I can test 12 variants. Given the mutation database and structural data, design a 12-variant panel that maximizes information gained about specificity determinants.”

8.6 My background

PhD in bioengineering. Comfortable with molecular biology, enzyme kinetics, and microbial fermentation. Less experienced with structural biology and bioinformatics — please explain structural reasoning clearly but do not oversimplify the biochemistry.

9. Session Log

(Prepend full context document + append this log to every session)

Session 2026.05.12 — Mechanistic hypothesis building + Round 1 experimental panel

Prompts used:

1. Mechanistic hypothesis building for PhaC_Cn scl selectivity
2. Suggest single N-terminal mutations to probe that region
3. Suggest single and combinatorial mutations outside the N-terminus
4. Consolidate all suggestions into a Round 1 experimental panel

Key outputs / hypotheses:

Hypothesis 1 — Two-point selectivity model: scl selectivity in PhaC_Cn is established at two structural points:

• Tunnel constriction (A510): Ala510 acts as a steric gate physically excluding acyl chains longer than ~C5 from reaching C319. This is the primary passive size filter. WT Ala at this position is already near-optimal for scl — bulkier substitutions (F, K, L, N, R) are inactive, and smaller substitutions (G, S, T) widen the tunnel and permit mcl incorporation.
• Active site geometry (F420): Residues near the base of the active site, including F420, optimize catalytic geometry for the C4 acyl-enzyme intermediate. F420S gives a 2.4-fold increase in 3HB-specific activity, consistent with relief of steric strain in the binding pose for 3HB-CoA.

Hypothesis 2 — R398 and H481 are positioning/catalytic residues, not selectivity gates: Both are strictly or highly conserved across Class I enzymes regardless of substrate preference. R398 likely contacts the CoA moiety to position the acyl chain for nucleophilic attack; H481 likely contributes to transition state stabilization. Neither is a primary chain-length selectivity determinant. Both are high-risk mutation targets.

Hypothesis 3 — Residual mcl leakiness has an uncharacterized structural origin: WT PhaC_Cn incorporates very low levels of 3HO and 3HDD. Since A510 appears near-optimal for scl exclusion and F420S already improves 3HB activity, the residual mcl leakiness likely originates from dynamic flexibility elsewhere in the tunnel — possibly in the uncharacterized region between R398/H481 (entrance) and A510 (constriction). This region is not yet described in Section 2.4 and should be a priority for PyMOL analysis.

Hypothesis 4 — N-terminal domain role is genuinely unclear: Deletion of residues 2–65 (Ye et al. 2008) shows no specificity change, suggesting the extreme N-terminus is dispensable. D153 and G175 are proposed as candidate positions based on Class II pairwise comparison only. G175 is conserved as Gly in both Class I and II, suggesting structural rather than selectivity role. D153 differs conservatively between classes (Asp vs. Glu) and is the stronger candidate. All N-terminal reasoning is [SPECULATIVE].

Round 1 experimental panel:

12 total variants including WT. Fits standard 12-well format with no spares — consider dropping Y440L (priority 10) to keep a spare well if expression failures are anticipated.

Flagged uncertainties / SPECULATIVE tags:

• All N-terminal suggestions (D153E, D153A, G175A) are [SPECULATIVE] — grounded only in AlphaFold model (low pLDDT in this region) and Class II pairwise comparison. No structural data from 5T6O available for this region.
• F318Y and Y440L are predictions transferred from PhaC_Ac (37% identity) — transfer reliability is unknown and should be treated as unvalidated until tested directly in PhaC_Cn.
• F420S mechanistic basis (Hypothesis A: steric relief vs. Hypothesis B: electrostatic contribution) is not resolved by available data — both remain plausible.
• The identity of tunnel residues between R398/H481 and A510 is not characterized in Section 2.4 — this gap limits mechanistic reasoning about the constriction region.
• Epistasis between all combinations is unknown; F420S + A510G is the combination with the most interpretable expected outcome.

Action items:

• Complete PyMOL tunnel analysis (Section 2.5) — residues within 10 Å of C319, focusing on the R398-to-A510 region
• Verify structural position of F420 relative to C319 and catalytic triad in 5T6O
• Order / construct Round 1 variant panel (12 variants listed above)
• After results: add all new data to Section 4 and Section 10 before next session

Session [DATE]

(repeat block)

10. Experimental Results Log

Experiment [DATE / ID]

Variants tested:

Interpretation: [What do these results mean for your mechanistic model?]

Surprises / inconsistencies with predictions: [Critical to record — unexpected results are often the most informative]

Updated hypotheses: [How do results revise your model of specificity determinants?]

Add to mutation database: [Y/N — copy rows to Section 4 as appropriate]

End of context document — v2.2 (single-enzyme / mutagenesis dataset scope) Keep this file updated and prepend it in full to every new LLM session

Notebook

This page logs the notes, thoughts, brainstorming, and planning associated with my individual final project.

Links to section

• Idea brainstorming
• Quorum-sensing based killswitch
• PHA synthase enzyme engineering

Feb 24, 2026

Brainstorming:

• Identification of PhaC analog in Cyanobacterium aponium UTEX 3222 and overproducing or engineering for increased efficiency
  • BLAST/align with known PHA-synthases
  • Compare efficiency / mutations that improved turnover in other PhaC - test analogous mutations (aligned location, similar or different AAs). improved substrate specificity?
  • Site-specific saturation mutagenesis? Would be good use for automation
• Quorum sensing based killswitch (i.e. cell dies if it escapes bioreactor)
  • Has to have some kind of inducible element or won’t grow after initial transformation
  • What’s good at quorum sensing already?
• Something else??? Something in E coli that can be done on Opentron
  • Because it’s more convenient for a final project to be executed in Victoria remotely
• Cyanobacterial expression plasmid across multiple cyano species
  • needs to include E coli machinery for manipulation and production (and conjugation, for relevant species)

Ideas:

1. PhaC protein engineering
   1. Short term aim: Design small library of PhaC variants with expected improvement
   2. Medium term aim: Generate library and test in chassis strain
   3. Long term aim: Develop PHB bio-manufacturing cyanobacterial strain for carbon-neutral/carbon-negative plastic (depending on biodegradation).
2. Quorum sensing based circuit for biocontainment
   1. Short term aim: Design killswitch with genetic circuit to trigger based on quorum sensing.
   2. Medium term aim: Build genetic circuit with expression based on quorum sensing with a measureable output; test circuit in E. coli.
   3. Long term aim: Optimize circuit sensitivity and test with killswitch expression; integrate into bio-manufacturing chassis strains for population-linked biocontainment.
3. Broad cyanobacterial expression plasmid
   1. Short term aim: Design plasmid backbone based off native cyanobacterial plasmids and established E. coli machinery.
   2. Medium term aim: Test expression in multiple cyanobacterial strains (including some previously considered genetically intractable with classic broad-host-range vectors).
   3. Long term aim: Establish protocol for domestication of newly prospected, wild-type cyanobacterial strains using the cyanobacterial plasmid.

Mar 31, 2026

Leaning towards quorum sensing killswitch because it’s more aligned with my prior experience and knowledge, so i think it will take less research on my behalf. since i’m already falling behind on homeworks, i’m worried about how much time it would take to optimize a protein since i have no prior machine learning experience.

Quorum sensing notes:

• auto-inducer: triggers expression
• keep in mind phenolic compounds and other naturally occurring quorum quenching
• also keep in mind the potential for auto-inducer production from other related bacteria; like if biomanufacturing strain escapes bioreactor but lands in soil with existing microbiome - we still want the escaped cells to die
  • maybe we could make a synthetic quorum sensing system for orthogonality: would require biosynthetic pathway for auto-inducer, auto-inducer recognition (inducible promoter, transcription factor, riboswitch, etc.), auto-inducer export pathway to preferentially or rapidly diffuse out of the cell
• for killswitch activation at low population (cells escaped from bioreactor), maybe consider secondary/back-up activation from an environmental signal
• potentially test circuit with fluorescence or colorimetric output first before killswitch/toxin-antitoxin genes

References

1. Miguel, CMTS; Santos, CA; Lima, EMF; et al. Quorum Sensing in Bacteria: From Mechanisms to Applications in Foods. 2026. Current Opinion in Food Science: 101394. DOI: 10.1016/j.cofs.2026.101394.
2. Ream, MJ; Prather, KLJ. Orthogonal quorum sensing circuits enable dynamic regulation in Escherichia coli. 2026. Metabolic Engineering, 96: 104-112. DOI: 10.1016/j.ymben.2026.03.009.

Circuits:

• killswitch trigger for population high:
  • low constitutive expression of toxin
  • autoinducer-triggered expression of antitoxin
  • maybe autoinducer can also trigger expression of toxin repressor
• killswitch trigger for population low (like Paula’s idea for targeted drug delivery):
  • low constitutive expression of antitoxin
  • autoinducer-triggered expression of toxin
  • maybe autoinducer can also trigger expression of antitoxin repressor

Apr 2, 2026

In Victoria node recitation last night, Derek suggested a possibility for cell-free testing of the system to be able to use the Gingko cloud lab instead. Originally i had figured that because i want it to be a killswitch system, that it needs to be in a living cell. Also because traditional quorum sensing systems are dependent on autoinducer concentration within vs outside the cell membrane. But Derek suggested considering instead how to design a quorum sensing system that would be cell-free like on a paper biosensor: that triggers when at a minimum concentration, rather than triggering at the first cell it sees.

Phrasing it that way made me think of the analog computing of the neuromorphic circuits: where inputs are additive (positive or negative). So to reach a minimum concentration rather than the very first thing present, there has to be a counter-actor to the sensed thing present; something that degrades or binds to the autoinducer or signalling metabolite that is expressed constitutively at a low level. So when the signalling molecule concentration is low (low population), it will be all used up by the counter-actor before it can trigger expression of the QS-controlled genes. When the signalling molecule concentration is high (high population), it will outnumber the counter-actor, so it can still trigger the QS-controlled genes.

For example, a riboswitch that recognizes a small molecule metabolite. The metabolite is produced and exported by cells, and when present, activates the gene of interest (in my switch a killswitch or fluorescent protein). We’d also want to express the riboswitch as an aptamer that is unconnected to the gene of interest to bind the metabolite at low concentrations, until a high concentration of the metabolite is reached and the metabolite outnumbers the loose aptamer and can trigger the riboswitch to activate gene expression.

Apr 3, 2026

Brainstorming and design

Drafts

Title: Population-dependent killswitch to prevent bioreactor escape

Short description: Biocontainment is essential for safe biomanufacturing, but most strains with biocontainment have bespoke systems designed for that particular strain. A population-dependent killswitch would kill any cells that escape the bioreactor where they are being cultivated or harvested. My initial idea is a toxin-antitoxin system expressed under control of a quorum-sensing circuit. Future considerations: safeguards against biocontainment escape through mutation, multiple levels of regulation.

Aims:

1. Design a genetic circuit that controls expression dependent upon cell population density in E. coli. The circuit will be designed with the intent for a final use with a killswitch, but fluorescent or colorimetric outputs might be used for initial design and validation. Validate the circuit with a simulation in Asimov Kernel.
2. Test circuit in E. coli with a measurable output (such as fluorescence).
3. Test circuit with killswitch; integrate into a biomanufacturing chassis strain for population-linked biocontainment.

Companies: Asimov - I plan on using Kernel to design and simulate my genetic circuit.
Basecamp Research - Maybe their AI can help me design overlapping genes to prevent killswitch escape via toxin gene mutation.
Cultivarium - If successful, quorum-based biocontainment could be a useful genetic tool to port to new potential chassis microbes.

Project idea slide

Project description: Biocontainment is essential for safe biomanufacturing, but most strains with biocontainment have bespoke systems designed for that particular strain. A population-dependent killswitch would kill any cells that escape the bioreactor where they are being cultivated/harvested. Initial idea is a toxin-antitoxin system expressed under control of a quorum-sensing circuit. Need to consider safeguards against biocontainment escape through mutation, multiple levels of regulation.

Asimov: I plan on using Kernel to design and simulate my genetic circuit. Basecamp Research: Maybe their AI can help me design overlapping genes to prevent killswitch escape via toxin gene mutation. Cultivarium: if successful, quorum-based biocontainment could be a useful genetic tool to port to new potential chassis microorganisms.

References:

1. Leonard SP; Halvorsen TM; Lim B; et al. Synthetic overlapping genes stabilize genetic systems. 2026. mBio, 17(3):e0272525. DOI: 10.1128/mbio.02725-25.
2. Blazejewski, T; Ho, H-I; Wang, HH. Synthetic sequence entanglement augments stability and containment of genetic information in cells. 2019. Science, 365(6453): 595-598. DOI: 10.1126/science.aav5477
3. Ream, MJ; Prather, KLJ. Orthogonal quorum sensing circuits enable dynamic regulation in Escherichia coli. 2026. Metabolic Engineering, 96: 104-112. DOI: 10.1016/j.ymben.2026.03.009

Apr 7, 2026

Last night in the Victoria node recitation, Derek was really talking about how cool the Kernel-Twist-Nebula-Waters pipeline is, and he mentioned that he was a little disappointed that not many projects seemed like they were fully utilizing it. Especially since he still isn’t back in Victoria, it seems like the only way i’ll be able to get any actual lab data unfortunately, so that’s different from my original plan with the quorum sensing. i ended up messaging Derek on Discourse to ask if it was too late to change my mind, and he said that while technically yes, since no one had signed up on my slide as a mentor yet, i could change it. So i went ahead and came up with a new slide and replaced my original response.

Description: Current bioplastic production is too expensive to compete commercially with petroleum plastics. Bioplastics, such as PHA, produced by cyanobacteria are one possible solution because they require minimal feedstocks (atmospheric carbon dioxide, sunlight, non-potable water), but processing can be expensive (harvesting, dewatering, purification). Cyanobacterium UTEX 3222 is a new potential chassis for PHB bioplastic with lower processing intensity due to its phenotypic properties (planktonic grown and settling, potential native PHB production). This strain can be further improved by engineering PHA synthase for increased substrate specificity, to preferentially produce PHB (to decrease purification steps). Existing data on PHA synthase substrate preference and specificity can be used to train an ML tool, which could recommend mutations that could be tested in a cell-free expression system.

Companies: Basecamp Research: Machine learning expertise and datasets Biofabricate: experience with biomaterials marketing and maybe scale-up Boltz.bio: ML for protein design (but it’s for drug discovery) Cultivarium: specifically for aim3, UTEX 3222 is currently considered genetically intractable

Ginkgo: Cell-free protein expression system Twist: DNA synthesis for my mutant library Waters: MS to identify PHA (general) and PHB (specific) production

References

These are for using mass spectrometry for PHA analysis, for the Waters step of the pipeline.

1. Khang, TU; Kim, M-J; Yoo, JI; et al. Rapid analysis of polyhydroxyalkanoate contents and its monomer compositions by pyrolysis-gas chromatography combined with mass spectrometry (Py-GC/MS). 2021. International Journal of Biological Macromolecules, 174: 449-456. DOI: 10.1016/j.ijbiomac.2021.01.108
2. Johnston, B; Radecka, I; Chiellini, E; et al. Mass spectrometry reveals molecular structure of polyhydroxyalkanoates attained by bioconversion of oxidized polypropylene waste fragments. 2019. Polymers, 11(10):1580. DOI: 10.3390/polym11101580
3. Conners, EM; Bose, A. State-of-the-art methods for quantifying microbial polyhydroxyalkanoates. 2025. ASM Applied and Environmental Microbiology, 91(9):e00274-25. DOI: 10.1128/aem.00274-25 These are for machine learning for enzyme engineering.
4. Landwehr, GM; Bogart, JW; Magalhaes, C; et al. Accelerated enzyme engineering by machine-learning guided cell-free expression. 2025. Nature Communications, 16: 865(2025). DOI: 10.1038/s41467-024-55399-0
5. Qiu, S; Saeed, H; Leonard, W; et al. Machine learning for enzyme catalytic activity: current progress and future horizons. 2026. Briefings in Bioinformatics, 27(1): bbag002. DOI: 10.1093/bib/bbag002
6. Cui, H; Su, Y; Dean, TJ; et al. Enzyme specificity prediction using cross-attention graph neural networks. 2025. Nature, 647: 639-647. DOI: 10.1038/s41586-025-09697-2
7. Link to sci comm article about Paper6: https://chbe.illinois.edu/news/stories/new-AI-tool-helps-match-enzymes-substrates This last one is about cell-free PHB synthesis.
8. Satoh, Y; Tajima, K; Tannai, H; et al. Enzyme-catalyzed poly(3-hydroxybutyrate) synthesis from acetate with CoA recycling and NADPH regeneration in Vitro. 2003. Journal of Bioscience and Bioengineering, 95(4): 335-341. DOI: 10.1016/S1389-1723(03)80064-6

Apr 9, 2026

Talking to Derek about the timeline in our Victoria node recitation last night, he suggested that everything for lab work will probably need to be ordered in the next two weeks to get data in time for final project presentations. He also said that if we want to run anything on the Ginkgo Nebula cloud lab, we need to talk with Ronan and see if he has the capacity for it. Given this timeline, i am almost definitely not going to be able to figure out any ML-guided protein engineering before the final ordering. What i’m thinking instead is to design initial constructs of PhaC from C. necator, PhaC from UTEX 3222, and a rational design for a UTEX 3222 PhaC mutant, all designed for cell-free expression. The reaction will probably include the monomer, since it’s simpler than using the full 5-enzyme cell free system from Satoh et al (reference 8 above) that used acetate as the feedstock, but I will need to double check the energy and CoA regeneration.

Apr 10, 2026

Enzyme sequence choices

PhaC enzyme from Cupriavidus necator was chosen as my wild-type. I used the amino acid sequence from Uniprot and codon-optimized it in Benchling for Escherichia coli. This was from homework 2 i think, and i just used that one.

For the mutant: I found a review paper (Ref1) that identified Ala510 in PhaC_Cnecator as having a role in substrate specificity: with A510M, A510Q, and A510C all increasing promiscuity (M, C both sulfur-containing residues; Q, C both polar residues; M, Q both larger residues); a related PhaC from Chromobacterium sp. USM2 found that changing the analogous A to M/W/V (all non-polar residues, larger than A) increased promiscuity; and the same PhaC from Chromobacterium sp. USM2 found that changing the analogous A to S (similar size, but polar) increased substrate specificity (towards short-chain-length PHAs, like PHB). It’s surprising that A->S had the oppposite effect as A->C, for these two different PhaC variants from different bacteria. But since I didn’t have time to read a lot more, I figured A510S was a good construct to test against the PhaC_Cnecator wild type to start with.

I tried to identify the PhaC sequence from UTEX 3222 to test as well, but I was unable to, as of yet. While the paper in which UTEX 3222 was prospected said the authors identified the genes encoding the PHA biosynthesis enzymes (Ref2), the genes weren’t annotated on the full genome sequence assembly, and I got no results BLASTing either C. necator PhaC or cyanobacterial PhaC from Synechocystis sp. PCC 6803 or the more closely relatated Microcystis aeruginosa sp. PCC 7608SL. I also tried BLASTing PhaE (another PHA biosynthetic enzyme) from a few different cyanobacterial strains as well, with still no results. All BLAST searches were tBLASTn to search for the nucleotide gene sequence within the genome assembly from the amino acid sequences from the various known PhaC protein sequences. The genes were also not listed amongst the biosynthetic genes listed in the Supplementary Information from the UTEX 3222 paper. I was out of ideas at this point, and on a time constraint, so to get constructs added to the order list today, I decided to move forward just with the PhaC_Cnwt and PhaC_CnA510S for now. If/when I get my ML design program working, maybe I can email George Church (or whoever on the author list did the genome annotation) to try to run it through, but I can definitely start with the C. necator one.

Construct design

Derek sent me a message asking me to order constructs today. So I went into Kernel to design PhaC_Cnecator with T7 promoter, RBS and terminator for cell-free expression because the E. coli based cell-free expression kits I found online from both NEB and Thermo Fisher both used T7 polymerase. In Kernel, I used a T7 promoter, T7 RBS, and T7 terminator from the iGEM repository. I chose promoter Bba_Z0251 from the many options because it had a lot of documentation on its iGEM registry page, and matched the full consensus sequence (from the T7 promoters iGEM page). I chose RBS Bba_Z0261 from the many options because it was analyzed by the same iGEM team as the promoter I used. I used T7 terminator Bba_K731721 from the many options because it most closely matched a quick google search for the T7 terminator sequence. While Kernel did have a genetic part for PhaC_Cnecator in the Uniprot repository, it actually didn’t have a nucleotide sequence associated with it. So i copied the promoter, RBS, and terminator sequences from Kernel into Benchling, where I used the previously codon-optimized PhaC_Cnecator sequence from homework 2. I used Benchling’s translation tool to identify the Ala at position 510, and changed a single nucleotide to change A510 to A510S (Ala: GCC; Ser: TCC) for the mutant.

Then I exported the FASTA files for both constructs and uploaded them into the Twist portal for clonal genes. I couldn’t remember if it mattered using linear gene fragments or clonal genes within a plasmid, but I went with clonal plasmid because previous experience with linear fragment orders from Twist were pretty low concentration. I remembered from lecture that Ronan preferred us to use chloramphenicol for an antibiotic marker if needed, so I decided to use the pTwist-Chlor-HighCopy cloning vector. Twist’s interface found both genes to be complex, so I used its internal codon optimization to fix this issue: I identified the organism as E. coli, did not omit any restriction enzyme recognition sites, and selected the promoter, RBS, and terminator regions as sequences that should not be changed. To my surprise, these sequences were not identical except the one point mutation; they were optimized differently, but I suppose it doesn’t really matter. Then I exported the full constructs (including plasmid) GenBank files from Twist and re-uploaded into Benchling to generate the link for adding the spreadsheet.

After meeting with Derek to explain, he suggested using linear gene fragments instead of clonal, so I re-did the Twist ordering bit to generate prices and optimized fragment GenBank files. I elected to leave the adaptors on because I assume those will give long enough arms, but I’m not really sure. Derek said he’d check with Ronan.

Experimental design

I checked that Millipore Sigma does in fact carry my substrate, I think. The substrate being the PHB monomer: 3-hydroxy-butyryl-CoA. However, it’s very expensive, so I’ll see about also ordering the DNA and cheaper substrates for the 5-enzyme biosynthetic pathway with CoA recycling that was in one of the papers I found (Ref3). After comparing the even just of all the substrates I’d still need for the full pathway, it’s cheaper just to order the original substrate (since I’d still need at least a little bit of CoA, which is still expensive on its own) then to also get a bunch of additional DNA. Derek mentioned that I need to figure out what kind of purification is needed for Waters to analyze my PHB product at the end of the reaction.

Reference

1. Chek, MF; Hiroe, A; Hakoshima, T; et al. PHA synthase (PhaC): interpreting the functions of bioplastic-producing enzyme from a structural perspective. 2018. Applied Microbiology and Biotechnology, 103: 1131-1141. DOI: 10.1007/s00253-018-9538-8
2. Schubert, MG; Tang, T-C; Goodchild-Michelman, IM; et al. Cyanobacteria newly isolated from marine volcanic seeps display rapid sinking and robust, high-density growth. 2024. ASM Applied and Environmental Microbiology, 90: e00841-24. DOI: 10.1128/aem.00841-24
3. Satoh, Y; Tajima, K; Tannai, H; et al. Enzyme-catalyzed poly(3-hydroxybutyrate) synthesis fro macetate with CoA recycling and NADPH regeneration in Vitro. 2003. Journal of Bioscience and Bioengineering, 95(4): 335-341. DOI: 10.1016/S1389-1723(03)80064-6

Apr 13, 2026

After Derek talked to Ronan, he suggested going back to clonal genes. I used my original PhaC_Cn construct, but then I decided to just have the point mutation for the mutant and have the rest of the sequence be identical. So I copied the PhaC_Cn-pTwist construct into a new DNA sequence in Benchling, and made the point mutation (GCA->TCA), and then verified quickly in Twist that this sequence is still simple and the same price.

Updated idea slide for cell-free synthesis Aim 1 instead of LLM/ML

Apr ?, 2026

Wanted to get started on the AI/ML aspect, so i started looking into that. My original plan had been to use this as an opportunity to expand my Python capabilities and play around with scikit-learn, but unfortunately, i realized i just didn’t have the time for that before the end of the course. After hearing Victoria node TAs Derek and Piyush talk about the AI tutor they developed for the course, I thought about maybe working with an LLM like Claude instead since that would require less initial coding on my end, and could still work because proteins are not dissimilar to language.

Apr 27, 2026

Did some reading, and it turns out that people have used LLMs for protein design before.

INPUT REFERENCES!!

May 7, 2026

Made an account for Claude.ai and