Week 14 Review: Bio-design and Bio-fabrication - live from SynBioBeta

Week 14 — Bio Design & Bio Fabrication

The dream of “real engineering” is what’s holding biology back. Bio-fabrication platforms are how we earn it.

About this lecture. Week 14 of HTGAA Spring 2026 was delivered live from SynBioBeta 2026 in San Jose and simulcast back to the MIT classroom and to the global HTGAA cohort. David Kong called it “our first time ever doing this kind of coast-to-coast interaction”. George Church watched from the chat; Joe Jacobson — who co-founded the company whose displays became the bottom layer of the platform Michael Chen would demo twenty minutes later — stood up during Q&A. The week ran with two co-speakers in dialogue rather than two consecutive lectures: Christina Agapakis on bio-design as philosophy and practice, and Michael Chen on bio-fabrication as an actual platform.

Why are we not aleady growing almost anything we can dream of?

Synthetic biology has been promising to “finally engineer life” for at least five hundred years, and a real one for fifty. The promise keeps moving — biofuels, then molecular biology, then proteins, now AI — but the lab-to-market gap looks similar each time. Week 14 reframes that gap. The conceptual answer (Christina Agapakis) is that we have been reaching for the wrong metaphor: there is no such thing as “real engineering,” and treating biology’s translation pipeline as a linear pipe is the failure mode. The practical answer (Michael Chen) is that wherever a parallel-screening primitive does exist — digital microfluidics, cell-free protein synthesis, split-GFP detection — the cycle time of biology compresses to weeks, and that’s what actually moves discoveries forward. Both are necessary; neither is sufficient alone.

For a course called “How to Grow Almost Anything,” this is the closing argument. Almost is the keyword; it is the thesis. You grow what people and cells will let you grow, and you wander to find the niche that actually wants what you made.

Core concepts

Bio-design vs. bio-engineering. Engineering implies linear translation: design → build → ship, with the desire on the market side held constant and treated as a problem to be solved by marketing. Design — Agapakis’s preferred frame — admits that desire on both sides (customers and cells) shifts under you while you build, that the technology and the market co-evolve, and that wandering the valley of death is the work, not a failure mode.

Bio-fabrication. A class of platforms that compress the build-and-test loop of biology by removing some traditional bottleneck — living cells, manual liquid handling, slow assays. Cell-free protein synthesis (CFPS) removes the cell. Digital microfluidics (DMF) removes the pipette. Together with a parallel optical readout, they remove the throughput bottleneck. The Nuclera eProtein Discovery platform shown this week is one concrete implementation.

Cell-free protein synthesis (CFPS). The transcription–translation machinery extracted into an open reactor (typically E. coli S30 or similar). You add DNA, NTPs, amino acids, energy regeneration, and your protein appears within hours — but more importantly, you can now tune the folding environment (chaperones, disulfide-bond-formation enzymes, cofactors, metal ions) at will, because there is no membrane to defend.

Electrowetting on dielectric (EWOD). A digital-microfluidic technique where individual ~nL droplets are addressed and moved across a planar electrode array by switching local voltages — analogous to how an e-reader switches pixels. Nuclera’s cartridges sit on a thin-film-transistor backplane originally engineered for E Ink displays; the company acquired E Ink’s digital microfluidics unit in 2021. The TFT pixel layer was always good at addressing many small things in parallel; what changed is that those small things are now nanoliters of biology.

Split-GFP detection. A 17-aa fragment of green fluorescent protein is genetically fused to your construct; the complementary inactive half is provided in solution; fluorescence appears only when the tag is exposed and the protein is soluble. This is a near-quantitative reporter for full-length, well-folded protein — and the spatial distribution of fluorescence within a droplet reveals aggregation. The same principle underpins the read-out across all eProtein Discovery workflows.

Cycle time. Joe Jacobson’s Q&A intervention: in VLSI semiconductor design you don’t tape out a working chip the first time; you measure progress by spin number, the number of fabrication iterations needed. State-of-the-art VLSI converges in ~3 spins. Several of Michael’s case studies hit that threshold for protein discovery — the question for the field is which classes of bio-deliverable can be brought down to comparable counts, and which inherently can’t.

A 500-year history of “synthetic biology”

The phrase keeps getting reinvented. Christina’s slides walked the audience through its predecessors:

1620s — Francis Bacon, New Atlantis. A speculative dispensatory: “we have not only all manner of exquisite distillations and separations… but also exact forms of composition whereby they incorporate almost as if they were natural simples.” Bio-design as thought experiment, three centuries before recombinant DNA.
1865 / 1898 — Claude Bernard, Introduction à l’étude de la médecine expérimentale. Experimental physiology, held to the same rigor as chemistry. Pasteur takes the same impulse industrial.
1912 — Stéphane Leduc, La Biologie Synthétique. A French physician shows osmotic chemical-garden experiments — ink in salt solutions producing plant-like and cell-like structures — and argues that life-like behavior is engineerable from chemistry alone. The book is the first published use of the term on a cover.
1978 — Szybalski & Skalka, Gene 4(3):181–182. A two-page editorial celebrating the Nobel for restriction enzymes: “the new era of ‘synthetic biology’ where not only existing genes are described and analyzed but also new gene arrangements can be constructed and evaluated.”
1979 — Science 206, 9 Nov 1979. A News & Comment piece a year before Genentech’s IPO, quoting E. F. Hutton analyst Nelson Schneider: “At present, the commercial applications of recombinant DNA remain as much shouting as substance, but the field has progressed with great rapidity and is clearly headed for interesting places.” Substitute “AI-driven protein design” and the sentence still works.

Agapakis’s conclusion: the complaint that biology isn’t real engineering yet has been the field’s anthem since well before “biology” had its modern meaning. The complaint is older than the discipline it indicts. If five hundred years of waiting hasn’t produced “real engineering,” perhaps the wait, not biology, is the problem.

Christina’s argument — innovation is downstream of desire

The field tends to externalize its frustrations onto four scapegoats: VCs are too dumb, biologists are too dumb, biology is too complex for human minds, the public hates GMOs. Each has a kernel of truth, but treating them as the diagnosis produces interventions that fail the same way the technology does — more money in the lab end of the pipe, more robots replacing scientists, more public-education campaigns about the difference between GMO and selective breeding. Each move assumes the linear lab → market model is the right model. Christina’s claim is that it isn’t, and never was. Innovation, in any field, is downstream of desire — customers want things, products that thread that desire survive, and the technology evolves to fit the niche rather than the other way around. You don’t bridge the valley of death; you wander it.

Three case studies make the point concrete:

Algae oil. Solazyme (founded 2003) raised on biofuels — algae-derived oil for transportation. Petrol prices ate the unit economics. The company rebranded as TerraVia in March 2016, pivoted to high-value food ingredients, filed Chapter 11 in August 2017, and was acquired by Corbion in September 2017 for ~$20 M + assumed debt. The technology survived, and by 2023 it was Corbion’s fastest-growing line. The consumer face is Algae Cooking Club, a culinary brand whose marketing pitches health (high omega-9, 535 °F smoke point, “no seed oils”) — not sustainability. The TikTok influencer Nara Smith cooks with it on camera. The pivot from biofuels to cooking oil was the win, not the failure.

Spider silk. Bolt Threads pitched it as a Kevlar-grade jacket material. There was no market for $10,000 jackets stronger than Kevlar. B-silk Protein — a yeast-fermented, spider-silk-derived polypeptide — found its niche in cosmetics, where small amounts of a well-defined biopolymer can credibly differentiate a hair-care or skin-care product. Bolt Threads paused Mylo (mycelium leather) production in 2023, but biotech-derived peptides are now a recurring class of premium beauty ingredient. (Christina specifically linked B-silk-style peptides to Lady Gaga’s Haus Labs cosmetics; the broader category claim is well-documented in trade press, but a direct B-silk ingredient-listing in Haus Labs products is [UNVERIFIED] in primary sources. Treat as illustrative of the category rather than a literal ingredient claim.)

Pharma — the cyclodextrin / Niemann-Pick C story. Chris Hempel, the mother of twin daughters Addi and Cassi (both NPC1), read a 2009 mouse paper showing that 2-hydroxypropyl-β-cyclodextrin (HPβCD) clears cholesterol storage and prolongs life in NPC1-deficient mice. She procured the compound — it had decades of safety data as a pharmaceutical excipient — and dose-titrated it into her own children. That single act of off-protocol parental science is what seeded the clinical program. Intrathecal HPβCD was published as a phase 1–2 trial in The Lancet in 2017; the drug is now under FDA regulatory review. Amy Dockser Marcus’s book We the Scientists (2023) is the long-form account. None of this was produced by the canonical “find the target, engineer a drug” pipeline. A mother and a science journalist were. Christina’s gloss: most drugs aren’t engineered into existence; they emerge from clinical understanding and a lot of wandering.

Key takeaway — wandering is the work. The companies that survive are not the ones that bridge the valley of death in one bound; they are the ones that have built the muscle to wander it. The wandering produces the pivot, and the pivot produces the niche that actually wanted the thing you made.

Michael’s argument — bio-fabrication as a parallel-screening primitive

Nuclera’s eProtein Discovery platform is built by stacking three primitives:

Digital microfluidics (DMF) via electrowetting-on-dielectric (EWOD). The substrate is the same TFT array that drives e-paper displays — Nuclera and E Ink partnered from 2018, and Nuclera acquired E Ink’s DMF unit in May 2021. E Ink remains the exclusive supplier of TFT backplanes. The pixel grid that addresses ink dots in your e-reader is now addressing nanoliter droplets of cell-free reactions.
Cell-free protein synthesis (CFPS). Eight pre-formulated reaction blends per workflow give eight different folding environments — varying chaperones, disulfide-bond-formation reagents (PDI + GSSG), cofactors, metal ions, and proteases. The reactions are linear over many orders of magnitude (nL on cartridge → mL or L scale-up) because they don’t consume oxygen.
Split-GFP detection. A 17-aa tag is fused to every construct; the soluble GFP complement is provided in solution; fluorescence reports full-length, soluble, well-folded protein. The spatial pattern within a droplet distinguishes three phenotypes — homogeneous bright (good expressor), homogeneous dim (low but well-behaved), and heterogeneous bright (aggregating). The software filters out heterogeneous hits so that “well-behaved” actually means well-behaved.

The product is a 24-hour parallel screen:

Workflow	Constructs × cell-free blends	Total expression conditions	Purification follow-up
Soluble	24 × 8	192	Top 30 → StrepTag MagBeads
Membrane (nanodisc)	11 × 8	88	All 88

User input: pipette DNA into a cartridge. User output: a heat-map of the best expression + purification conditions for each construct, ready to scale up either in CFPS at liter scale or in BL21(DE3) + T7 E. coli — the platform claims good correlation with that scale-up route, especially when disulfide-bond folding agents are part of the recipe.

What this enables, illustrated by Michael’s case studies:

Targeted protein degradation at AbbVie. A CRISPR-knockout screen identified one E3 ligase responsible for ubiquitinating a specific target. Mapping which of the target’s 35 lysines were modified required 48 mutants across three rounds (15 mutants × 3–4 AA changes, 20 × 2 AA, 12 × single AA). Conventional pipelines: 6–12 months. With eProtein Discovery: 2–3 months. A 3–4× compression of a real drug-discovery campaign.
An ABC transporter at Imperial College. DNA arrival to low-resolution cryo-EM grid in approximately a week — the platform produced the membrane protein in nanodiscs ready for QC, structure determination followed.
FFAR1 (a GPCR) at Diamond Light Source / Andrew Quigley’s group. The first active GPCR demonstrated on the platform; ligand-induced stabilization verified by nano-DSF. (High-resolution structure is reportedly being pursued — [UNVERIFIED] as a published result at the time of this writing.)
Bayer Crop Science membrane proteins. Aiming for pest-specific herbicide/pesticide targets: 8 of 9 membrane proteins recovered on the Nuclera platform versus a much lower hit-rate on E. coli, insect, yeast, and a competing eukaryotic cell-free system.
Ribbon Bio × Scala Biodesign — a restriction enzyme. Scala’s computational stability-design algorithms (Scala Biodesign was founded by alumni of the Sarel Fleishman lab at the Weizmann Institute) generate variant designs; Ribbon synthesizes them; Nuclera screens them. The case study highlighted in the slide: 9× yield improvement and +14 °C thermal stability in ~1–2 weeks on a specific restriction enzyme.

Forward-looking: a $100-per-antibody CFPS-based screening service is being introduced for AI-ML antibody-discovery pipelines that need to validate thousands of zero-shot designs cheaply — full-fat IgGs via heavy + light chain co-expression, aglycosylated but otherwise comparable to CHO-produced material.

David’s bridge — the planetary-scale classroom

Between Christina and Michael, David Kong reframes HTGAA itself as a distributed, planetary-scale synthetic-biology experiment. ~1,500 learners globally; physical nodes from Hartnell College in Salinas to the Biopunk Community Lab in San Francisco; published in Nature Biotechnology with student authors who had no prior wet-lab experience.

Three signature projects from the year:

The phage-therapy lysis-protein design challenge. Students worldwide designed lysis proteins (DNA synthesized by Twist), screened on Nuclera systems at MIT. A previewed cell-free assay where each pixel of a video is the fluorescence read-out of one student’s design.
Neuromorphic genetic circuits with Ron Weiss. Analog-behavior circuits assembled, transfected, and tested by robots — innovation distributed globally, execution centralized to lab automation.
The biopixel art experiment. An r/place-inspired live biopixel canvas — 1,536-well plate, edit one pixel, cool-down, edit again — open across HTGAA and SynBioBeta. The winning image gets printed in live E. coli via Ginkgo’s cloud lab and shown at the SynBioBeta closing ceremony. The pedagogical anchor is Papert & Solomon (1971), “Twenty Things to Do with a Computer” (MIT AI Memo 248 / Logo Memo 3) — Papert’s first thing was a Logo turtle drawing shapes; HTGAA’s first thing for the cloud lab is “grow an artwork.”

The undercurrent connecting all three is the recent OpenAI × Ginkgo / GPT-5 closed-loop CFPS optimization result. Over six rounds of closed-loop experimentation — 36,000 unique CFPS reaction compositions across 580 automated plates — GPT-5 designed experiments, Ginkgo’s cloud lab executed them, and the system cut cell-free protein synthesis production cost by 40 % with a parallel 27 % increase in titer (sfGFP at $422/g final, ~57 % reagent-cost improvement). The optimized reagent mix is already on Ginkgo’s shelves. HTGAA’s response: an internal “AI cobot,” trained on ten years of HTGAA material, tentatively to be called George (in honor of Church, who was watching from chat). Students will work with the cobot to design their next round of cell-free experiments after their first round comes back.

Cycle time — Joe Jacobson’s bridge

Joe Jacobson stood up in Q&A and dropped the single connecting concept. In VLSI semiconductor design, no one ships a chip on the first tape-out; you measure progress by the spin number — the number of fabrication iterations needed. State-of-the-art VLSI converges in about three. Michael’s protein-expression case studies hit roughly that threshold. Antibody discovery is approaching it — current zero-shot designs report up to ~20 % hit rates for detectable affinity, but typically still need 1–2 follow-up engineering rounds to reach functional sub-nM affinity. The discipline that separates “could engineer biology in principle” from “did engineer it in practice” is converging the spin number.

George Church, jumping in, gave the day’s strongest forward marker for the high end of the difficulty curve: baby KJ Muldoon at CHOP/Penn — severe CPS1 deficiency, a bespoke CRISPR base-editing therapy from the Kiran Musunuru lab, first dose at 6–7 months of age in February 2025, published NEJM 2025. Seven months from diagnosis to a bespoke therapy. Church’s hope: “that’s the rule, not the exception, going forward — or maybe even faster.” That is what a converged spin number looks like at the absolute frontier of bio-fabrication.

Pitfalls, controls, and how to know it worked

For Christina’s frame. The natural pushback is that some technologies do drive their own market — mRNA-LNP, CAR-T, deep learning itself. Push and pull both occur. The honest reading of the talk is don’t mistake supply-side push for the default mode of the field; it is the exception, and most projects assume push when they should be looking for pull.

For Michael’s platform. The fluorescence read-out is a soluble-expression proxy, not an activity assay. A construct that lights up green is full-length and unaggregated; it isn’t necessarily functional. Activity, binding affinity, structure — those are downstream, and the platform shortens the time to get there but does not replace the assay. Michael was explicit that high-resolution cryo-EM of GPCRs in a week, for example, is not what the platform delivers (despite some customer hopes).

For the AI / closed-loop frame. GPT-5 + Ginkgo’s lab cut sfGFP cost by 40 %; that is a quantitative improvement on a quantitative objective. It is not a discovery — sfGFP was already a working protein. The harder open question is whether the same closed-loop architecture handles the messy, unbiased-exploration discovery problems where the objective itself is not pre-defined.

Course resources

HTGAA 2025/2026 course site: How to Grow Almost Anything.
Nuclera eProtein Discovery technology page: nuclera.com/technology — including the May 2021 announcement of Nuclera’s acquisition of E Ink’s digital microfluidics unit.
Ginkgo × OpenAI press release, Feb 2026: Ginkgo Bioworks announcement and OpenAI blog: GPT-5 lowers the cost of cell-free protein synthesis.
Christina Agapakis’s agencies: Oscillator (her personal hub), American Wetware (newer venture, “biology has a design language and we know how to learn it”).
Baby KJ Muldoon / personalized CRISPR therapy: CHOP press release, Penn Medicine press release.

A note on cross-week threads. Two pieces of Week 14 plug directly into the MS2 L-protein engineering project that I am carrying across the group-project arc. Christina’s cells-have-desires frame is a useful counterweight to a strictly LLR-driven design strategy (see Week 5’s r ≈ 0 finding on the L-protein) — the under-represented protein family wants what it wants, and the design language has to listen. Michael’s split-GFP / soluble-fraction read-out is exactly the right quality metric for the MS2-L variant library at the screening stage: full-length, soluble, well-folded, scaled cheaply. Both are pragmatic guidance for the final-project build.

Page created as topic guide. Lecture delivered 2026-05-05; page last updated 2026-05-26. Contact Fiona for further discussions and questions and to hear how bloody incredible she found the whole HTGAA course! Kudos to the whole team and nodes involved!