Week 1 HW: Designed Mini-Protein Epigenetic Modifiers

Sep 22, 2026 · @Rishi Shetty

1. What I would like to biomanufacture, and why?

The product. I propose to biomanufacture a library of designed mini-protein binders — engineered proteins of roughly 60–120 amino acids — that disrupt PRC2 by wedging into the interface between its EZH2 and EED subunits. PRC2 is the complex that deposits and propagates H3K27me3: EZH2 is the catalytic writer, and EED is the reader that binds existing H3K27me3 and allosterically stimulates EZH2 to write more, so the mark drives its own spread along chromatin. Breaking the EZH2–EED association collapses that read–write loop without abolishing basal methyltransferase activity.

Why this interface, and not the EED aromatic cage. My first instinct was to target the trimethyllysine-binding pocket on EED directly. That pocket is thoroughly validated — Qi et al. (2017) reported EED226, a selective allosteric PRC2 inhibitor that binds it, blocks H3K27 methylation at PRC2 target genes and induces regression of lymphoma xenografts, and Novartis has since published a series of related compounds with crystal structures showing an induced-fit pocket formed by rearrangement of the aromatic cage residues Phe97, Trp364 and Tyr365. But that is precisely the problem: it is a small, deep, trimethyllysine-sized cavity, and a folded 60–120 residue protein cannot occupy it. That target belongs to small molecules. Designed mini-proteins excel at the opposite geometry — broad, shallow protein–protein interfaces with no druggable pocket — which is how the mini-protein Omomyc reached clinical trials against MYC, long considered undruggable for exactly that reason. The EZH2–EED interface is that kind of surface, and there is peptide precedent for disrupting it: a stabilised alpha-helix mimetic of the EZH2 peptide selectively inhibits H3K27 trimethylation in vitro and suppresses MLL-AF9 leukemia cell growth. Matching molecule class to target geometry is the reason for the switch.

The problem it addresses. Chromatin does not simply “degrade” with age; it redistributes. Constitutive heterochromatin erodes (H3K9me3 loss, LINE-1 derepression), while facultative marks such as H3K27me3 lose their sharp domain boundaries and spread into regions that were previously unmarked. The net result is a loss of the crisp nuclear compartmentalisation that distinguishes a young nucleus from an old one — and that loss is measurable as a change in chromatin texture in immunofluorescence images, which is the basis of the ImAge epigenetic clock (Alvarez-Kuglen et al., 2024).

Most interventions that shift epigenetic age are blunt. Transient OSKM expression resets a broad swathe of the epigenome at once, at the cost of partial loss of cell identity and a real teratoma risk. What is missing is a set of targeted, dose-controllable, non-genetic tools that push one axis of chromatin state at a time.

Why a protein, and not a small molecule or a synthetic peptide. This distinction is the reason biology is required at all:

• Solid-phase peptide synthesis handles linear peptides up to roughly 50 residues cheaply and at large scale. If my product were genuinely a short peptide, chemistry would beat biomanufacturing on every axis and the proposal would collapse.

• A 60–120 residue folded mini-protein sits past comfortable SPPS range. It needs a defined tertiary fold to present a binding surface with shape complementarity that a flat peptide cannot achieve.

• Designed mini-protein binders can reach high affinities against targets small molecules find undruggable, precisely because they bury a large, shaped interface rather than occupying a small pocket.

• Cell-free expression additionally allows incorporation of non-canonical amino acids — for protease resistance, or site-specific conjugation of a nuclear-localisation or cell-penetrating moiety — which is awkward in a living host.

Who it is for. In the near term the users are ageing-biology researchers: this is a reagent set for testing whether a specific chromatin axis is causal for epigenetic ageing phenotypes rather than merely correlated with them. Longer term the same molecules are candidate therapeutics for tissue-specific rejuvenation, but I deliberately cost the research-reagent scale below, because that is the scale at which the manufacturing question stays honest.

Personal motivation. My bench work is in Prof. Alexey Terskikh’s lab at the Harry Perkins Institute, where I run the ImAge/PLATE-MIEL pipeline: nuclear isolation from mouse tissues, H3K27me3 and H3K27ac immunostaining, and high-content imaging on a PerkinElmer Operetta CLS in 384-well format. I spend my time measuring epigenetic age with considerable precision and have almost no ability to perturb it in a controlled way. This proposal is an attempt to build a perturbation arm to match the measurement arm I already have.

2. What living system would I use?

Chassis: an E. coli cell-free transcription–translation (TX-TL) system. Crude S30 lysate supplemented with an energy regeneration system, NTPs, amino acids and T7 RNA polymerase, run in a continuous-exchange format (CECF) for the prototyping phase.

Why cell-free rather than a living host.

• No viability constraint. A PRC2 inhibitor is, by design, a molecule that disrupts chromatin regulation. Expressing it inside a living cell risks poisoning the producer. In an open lysate there is no host to poison, no chromatin to disrupt, and no growth phase to optimise around.

• Speed of the design–build–test loop. Linear DNA templates can go straight into a reaction without cloning or transformation, so a design can be tested in hours rather than days. This is the central argument of Karim and Jewett (2021) and it is what makes a combinatorial screen tractable.

• Open reaction. The reaction is physically accessible, so redox conditions, chaperones and disulfide isomerase can be titrated directly, and non-canonical amino acids can be supplied by adding an orthogonal tRNA/synthetase pair to the pot.

Inputs. Linear or plasmid DNA encoding each mini-protein variant under a T7 promoter, with an N-terminal SUMO solubility tag and a C-terminal His6 for purification; E. coli S30 extract; energy mix (creatine phosphate/creatine kinase or a PANOx-SP–style system); NTPs; 20 amino acids; Mg²⁺ and K⁺ optimised per construct.

Process. T7 polymerase transcribes the template; endogenous ribosomes in the lysate translate it co-transcriptionally; the mini-protein folds in the reaction, assisted by supplemented GroEL/ES and an oxidising glutathione buffer if disulfides are required. In CECF format a semipermeable membrane continuously supplies substrates and removes inhibitory by-products (inorganic phosphate, lactate), extending active synthesis from a few hours in batch to roughly a day.

Outputs. Crude reaction containing soluble mini-protein, which is then IMAC-purified, SUMO-protease cleaved to liberate the native N-terminus, and polished by size-exclusion chromatography. For screening, crude reaction can be used directly without purification.

Assumptions I am making, stated explicitly.

• The mini-protein folds without eukaryotic post-translational modification. Reasonable for a designed binder built from a simple fold, but not free — see section 5.

• The target-binding surface can be designed computationally to useful affinity before any wet-lab work, and my job is to screen the resulting shortlist rather than discover binders de novo.

• E. coli lysate has no chromatin of its own, so there is no confounding on-target activity inside the production reaction.

• For the eventual manufacturing phase I assume the molecule can be moved to a conventional host — see section 4(b).

3. How would I use automation?

I would combine two automation technologies, and the point is that they form a closed loop: the measurement instrument is not a quality check bolted on at the end, it is the optimiser that chooses what gets built next.

(a) Build — Opentrons OT-2 liquid handler. The OT-2 assembles cell-free reactions combinatorially in 384-well plates: one axis is the mini-protein variant (DNA template), the other is reaction chemistry (Mg²⁺, K⁺, chaperone supplementation, redox condition, template concentration). At ~10 µL per reaction, one plate is 384 conditions from under 4 mL of lysate. Manual assembly of that matrix is roughly a day of pipetting with a pipetting-error variance that would swamp the effect I am trying to measure; the OT-2 does it in about an hour with consistent dispense volumes.

(b) Test — automated high-content imaging on isolated nuclei. Nuclei isolated from mouse liver or brain are arrayed in 384-well imaging plates, treated with crude or purified reaction product, then fixed, immunostained for H3K27me3 and H3K27ac, and imaged on a PerkinElmer Operetta CLS. Images go through the PLATE-MIEL segmentation pipeline (GPU-accelerated, run on the Pawsey Setonix cluster), which extracts chromatin texture features and returns an ImAge-style score per nucleus.

A note on assay design: I originally considered DAPI/PI staining as the readout, but propidium iodide is a membrane-integrity dye and is uninformative on isolated nuclei, whose membranes are already permeabilised. DNA-dye texture alone is also a weak proxy for epigenetic state. Histone-mark immunofluorescence is both the correct readout and the one already validated in our pipeline.

(c) The loop. Texture scores from plate n feed a simple Bayesian optimiser that proposes the variant/chemistry combinations for plate n+1, which the OT-2 then builds. Human involvement is limited to plate transfer and periodic review. Each cycle is roughly 48 hours: reaction assembly and expression (~24 h in CECF), nuclei treatment and staining (~8 h), imaging (~4 h overnight on the Operetta), segmentation and analysis (~2 h on Setonix).

Why this particular pairing. Screening designed binders is a high-dimensional search over sequence and reaction conditions, and the informative readout — chromatin texture — is an image, not a number from a plate reader. Automating only liquid handling would still leave imaging and analysis as the bottleneck; automating only imaging would leave me hand-building the matrix. The value is in closing the circuit between them.

4. How much would I need to make, and how long would it take?

(a) How much?

Scale definition. A preclinical efficacy study in mice: 100 animals dosed over a 10-week cyclic regimen, mirroring the dosing logic used for partial reprogramming studies.

So the pilot scale is sub-gram, not 100 g. This is worth stating plainly, because it inverts the intuition: the hard problem here is not making enough material, it is making the right material. If the product ever reached human trials the number changes character — at 5 mg/kg in a 70 kg adult that is 350 mg per dose, so 100 patients × 10 doses ≈ 350 g, a thousandfold jump that would not be run on cell-free at all (see below).

Screening demand, which is actually larger than dosing demand. The discovery campaign needs 5,000 variants × 10 µL reactions ≈ 50 mL of total reaction volume, plus repeat rounds — trivial in volume, but it is where nearly all the time goes.

(b) How long?

Discovery phase (cell-free + Opentrons + Operetta).

• 5,000 variants ÷ 384 wells per plate ≈ 13 plates per round

• 1 round ≈ 48 h (assembly → expression → staining → imaging → analysis), with 2 plates running in parallel

• 13 plates ≈ 7 working days per round; 5 optimisation rounds ≈ 5–7 weeks

Production of 375 mg by cell-free.

• Assumed CECF yield: 0.5 g/L for a novel designed protein. Published cell-free yields span roughly 0.5–2 g/L for well-behaved model proteins; a first-pass designed binder belongs at the low end. This is an assumption, not a measurement.

• Purification recovery: 40% (IMAC → SUMO cleavage → SEC, three steps at ~75% each)

• Reaction volume required: 0.375 g ÷ (0.5 g/L × 0.40) = 1.9 L

• At 10 mL per CECF device, that is ~190 reactions, or ~19 reactions per run across 10 parallel units

• Each CECF run ≈ 24 h synthesis + 12 h turnaround → ~2 weeks including downstream processing

Total: roughly 2 months from design shortlist to dosed animals.

Cost is the real constraint, and I want to name it rather than hide it. Cell-free reagent cost runs on the order of hundreds to low thousands of dollars per litre once NTPs, energy system, amino acids and template are counted. At 1.9 L that is a four-to-five-figure reagent bill for a mouse study — acceptable. At the 350 g human scale it would require ~1,750 L of reaction and the reagent cost alone becomes prohibitive.

Prototype vs manufacture — these are different platforms, and I think that is the honest conclusion.

Once a single winning sequence is fixed, the properties that made cell-free attractive — no cloning, no host toxicity, open reaction — stop mattering, and its cost per gram becomes the dominant term. A toxic product would be the one case for staying cell-free at scale; I would test host toxicity early precisely to settle that question.

5. What don’t I know yet?

Unknown 1 — Does the screening assay measure anything real?

This is the deepest problem in the proposal. An isolated nucleus in a tube has no acetyl-CoA, no S-adenosylmethionine, no ATP regeneration and no chromatin turnover. Whatever my imaging screen reports, it is not “rejuvenation” — at best it reports binding, or direct enzymatic inhibition of residual PRC2 activity, over a few hours. The gap between that proxy and an actual shift in epigenetic age in a living tissue is the single largest assumption in the whole design.

What I would do: run the screen in parallel on isolated nuclei and on live primary cells from the same animal, and ask whether the two rank the same variants in the same order. If the rank correlation is poor, the cheap high-throughput assay is worthless and the whole automation loop has to be rebuilt around live cells, which slows every cycle by an order of magnitude. I would answer this with ~20 variants before committing to 5,000.

Unknown 2 — Delivery, which is a harder problem than folding.

My earlier draft listed folding stability as the main risk. On reflection, delivery is worse. The molecule must survive serum proteases, escape renal clearance (proteins below the ~60 kDa filtration cutoff clear in minutes), cross the plasma membrane, then cross the nuclear envelope, in a specific tissue. Each of those is an unsolved problem for designed mini-proteins, and a molecule that works beautifully on naked nuclei may never reach chromatin in vivo.

What I would do: measure plasma half-life and nuclear delivery early, with a fluorescently labelled inert variant, before optimising potency at all. If nuclear delivery is below a few percent, the dose calculation in section 4 is wrong by orders of magnitude and the project pivots to a conjugate or nanoparticle formulation.

Unknown 3 — Whether the biology is even the right target.

H3K27me3 redistribution is well documented as a correlate of ageing. Whether reversing it is causally beneficial, neutral, or harmful is not established. PRC2 is essential in many contexts; partially inhibiting it could plausibly accelerate a dysfunctional phenotype rather than rescue it, and there is no dose at which I currently know the sign of the effect.

What I would do: use existing small-molecule EED and EZH2 inhibitors as chemical stand-ins to test the directional hypothesis before manufacturing anything. If a tool compound does not move ImAge score in the desired direction in primary cells, there is no point building a better one.

Smaller unknowns worth flagging: the hit rate of the computational design stage itself — a 2025 assessment of RFdiffusion-designed binders found many candidates failed to express detectably or showed no measurable binding against several protein targets, which is a large part of why the automated screening loop in section 3 is necessary rather than optional; cell-free yield for this specific sequence class (0.5 g/L is an assumption, not a measurement); whether the fold requires disulfides and therefore an oxidising reaction environment; and whether crude lysate components interfere with the immunostaining readout when unpurified reactions are applied to nuclei.

AI use declaration

Client: Claude (Anthropic), Opus 5, accessed via claude.ai. Session dated 22 September 2026.

Prompts used, in sequence:

• Uploaded the assignment brief and asked for proposal ideas grounded in my research background (epigenetic ageing, chromatin imaging, nuclei isolation).

• “Any other ideas with regards to this topic? Scrutinize every aspect of this idea.” — requested a second, broader set of proposals.

• Pasted my own first draft of this concept (recombinant epigenetic modifiers, cell-free chassis, Opentrons with DAPI/PI screening, 100 g at 1 g/L over 14 days) and asked for help brainstorming why it would and would not work.

• Asked for a complete draft I could edit and submit.

** I have already researched extensively on these topics in the past so the prompts might look a little lacking.

How the output was used. The concept, the choice of chassis, the automation pairing and the connection to my own ImAge pipeline were mine from the outset. The AI’s substantive contributions were critical rather than generative: it identified that “chromatin degradation” mischaracterises the biology (ageing chromatin redistributes rather than degrades), that propidium iodide is uninformative on isolated nuclei, that a sub-50-residue peptide would be better made by solid-phase synthesis than by biology, that my original 100 g figure corresponded to a clinical rather than a preclinical scale, and that my yield and timing figures were internally inconsistent. It also drafted the prose of this document, which I have reviewed and edited.

Numbers to verify before presenting: the dose assumption (5 mg/kg) is a conventional starting point rather than a literature value for this molecule class; cell-free yield and purification recovery are order-of-magnitude assumptions, and are flagged as such in section 4.

References

 Alvarez-Kuglen, M., Ninomiya, K., Qin, H., Rodriguez, D., et al. (2024). ImAge quantitates aging and rejuvenation. Nature Aging, 4, 1308–1327. https://doi.org/10.1038/s43587-024-00685-1

 Asin-Garcia, E., Fawcett, J. D., Batianis, C., & Martins dos Santos, V. A. P. (2025). A snapshot of biomanufacturing and the need for enabling research infrastructure. Trends in Biotechnology, 43(5), 1000–1014.

 Farhy, C., et al. (2019). Improving drug discovery using image-based multiparametric analysis of the epigenetic landscape. eLife, 8, e49683.

 Hamilton, N. A., Pantelic, R. S., Hanson, K., & Teasdale, R. D. (2007). Fast automated cell phenotype image classification. BMC Bioinformatics, 8, 110.

 Karim, A. S., & Jewett, M. C. (2021). Toward sustainable, cell-free biomanufacturing. Current Opinion in Biotechnology, 69, 136–144.

 Kim, W., Bird, G. H., Neff, T., et al. (2013). Targeted disruption of the EZH2–EED complex inhibits EZH2-dependent cancer. Nature Chemical Biology, 9. https://doi.org/10.1038/nchembio.1331

 Li, L., et al. (2017). Discovery and molecular basis of a diverse set of Polycomb repressive complex 2 inhibitors recognition by EED. PLOS ONE, 12(1), e0169855.

 Moqri, M., et al. (2023). Biomarkers of aging for the identification and evaluation of longevity interventions. Cell, 186, 3758–3775.

 Petzold, C. J., & Mukhopadhyay, A. (2025). From bench to biofactory: High-throughput technologies and automated workflows to accelerate biomanufacturing. Current Opinion in Biotechnology, 94, 103320.

 Qi, W., Zhao, K., Gu, J., et al. (2017). An allosteric PRC2 inhibitor targeting the H3K27me3 binding pocket of EED. Nature Chemical Biology, 13, 381–388.

 Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620, 1089–1100.