Programmable Biology

1. The tension: program it vs. hold it

Reliability splits two ways. Robustness — the system holds a target despite disturbance. Composability — you stack modules and the result behaves as designed. Most synthetic circuits die on the second: a part that works alone leaks, loads, or retroacts when wired into a cascade. The approaches below come at the problem from opposite ends.

2. Control theory in the cell — Khammash (ETH Zurich)

Import control theory into the cell. Two flavors: external/closed-loop, where a computer watches single cells under a microscope and steers gene expression in real time via light (optogenetics); and embedded, where you build the controller out of molecules so the cell regulates itself.

Antithetic integral feedback (AIF)

Two molecular species — e.g. a sigma factor and an anti-sigma factor — are produced, one tracking the output and one the setpoint, and they annihilate each other (sequestration). That annihilation mathematically performs integration of the error, the textbook mechanism for robust perfect adaptation (RPA): the output returns exactly to setpoint after any disturbance, independent of parameter values.

RPA is a structural property — it does not depend on tuning rate constants, which is exactly what kills most synthetic circuits.

Briat, Gupta & Khammash proved (Cell Systems, 2016) that antithetic feedback guarantees RPA even in noisy networks. Aoki et al. proved it is essentially the unique fundamental topology and built it in E. coli, showing adaptation holds for both population-average and single-cell time-average (Aoki et al., Nature 570:533, 2019). Extensions add proportional and derivative terms — biomolecular PID — to cut variance and improve transient response.

Shipped wins. Optogenetic closed-loop control of the unfolded protein response lifted α-amylase titers ~60% (Nat. Commun., 2025; see also PMID 36914087). "Cyberloop"/"Cybergenetics" platforms now do rapid prototyping of single-cell controllers (ETH cybergenetics). The most theoretically grounded answer to "reliable" in the field, demonstrated in E. coli, yeast, and mammalian cells.

3. Clean abstraction — the substrate is the program

a. dCas9 / CRISPRi — the guide RNA is the software (Qi, Stanford)

Catalytically dead Cas9 (dCas9) is an RNA-guided DNA-binding protein with no enzymatic effect of its own. Point it at a promoter with a guide RNA and it blocks transcription (CRISPRi); fuse it to an activator and it turns genes on (CRISPRa). Targeting is set entirely by the 20-nt guide, so the guide RNA is the program. One universal actuator, swap the gRNA, retarget anywhere. The cleanest software abstraction in synthetic biology.

Orthogonality is the reliability mechanism: many guides act independently, so wide circuits don't each need a bespoke protein. dCas9 NOT-gates are low-burden vs protein-cascade gates; Cas13d/dCas12a multiplex dozens of perturbations at once. A CRISPR/dCas9 "central processing unit" runs logic in human cells (PNAS, 2019); CRISPRi circuits now run in plants (Nat. Biotech, 2024).

b. synNotch — swappable cell I/O (Lim, UCSF)

Gut the natural Notch receptor: swap the extracellular sensing domain for a custom antibody (choose your input — any surface antigen) and the intracellular domain for a synthetic transcription factor (choose your output — any gene program). Antigen binding triggers cleavage that releases the TF, which drives the payload. Input and output are independently swappable modules: a clean sense→compute→actuate abstraction at the cell level.

The signaling is mechanical (force-induced proteolysis), not a leaky enzymatic cascade, so basal noise is low and ON/OFF is sharp. Combine synNotch with a CAR for an AND-gate tumor recognition, the basis for next-gen cell therapies (Roybal et al., Cell 167, 2016). Wire two cells' outputs to each other's inputs and you get programmable multicellular patterning (Toda et al., 2019). It is a cellular device, not a circuit: slow (transcriptional timescales), and per-application genome architecture is engineering, not plug-and-play. Conceptually mature, moving into clinical CAR-T gating.

c. DNA strand displacement — sequence as program (Winfree & Qian, Caltech)

Compute with the physics of base-pairing alone: no enzymes, no cells. A single-stranded input displaces a strand from a partly-double-stranded gate via a toehold, releasing an output that feeds the next gate. Qian & Winfree's seesaw gate is a standardized, reversible motif; networks realize AND/OR/NOT and analog weighted-sum + threshold — neural-network-style computation in a test tube. They built a 4-neuron Hopfield network and 100s-of-gate digital circuits (Qian, Winfree, Bruck, Nature 475:368, 2011).

This is the substrate where reliability is most rigorously characterizable: the only "machine" is Watson-Crick hybridization, whose thermodynamics and kinetics are predictable from sequence. You can design, simulate, and trust large circuits in a way you cannot in a living cell.

The cost: leak (spurious displacement without the correct input) dominates error and limits depth. Reactions are slow (hours), one-shot (consumed fuel, no clean reset), and confined to the tube: not autonomous, not in vivo. Scaling to thousands of gates is a synthesis-purity and leak-suppression grind.

4. Predictable physics as a prototyping layer

Cell-free TX-TL — the breadboard (Noireaux, Murray)

Run transcription + translation in a cell extract, without live cells: a biomolecular breadboard. Dial DNA template, polymerase, and resource concentrations directly; no membrane, no growth, no evolution. Prototype a circuit in hours instead of the days-to-weeks of cloning into cells. Linear DNA works, so no cloning step: a four-component genetic switch was prototyped in under 8 hours (Siegal-Gaskins et al., ACS Syn Bio, 2014; protocols, 2013).

5. Different computational paradigms inside the cell

Analog / cytomorphic computation (Sarpeshkar, Dartmouth)

Stop forcing biology into digital logic. The Boltzmann exponentials of subthreshold transistors have the same mathematical form as mass-action chemical kinetics, so map analog electronic circuit motifs directly onto molecular ones (cytomorphic mapping), and conversely simulate cell networks on analog electronic supercomputers. Computation runs on continuous concentrations, not 0/1 states. The argument: analog is more resource-efficient than digital at moderate precision — a few molecules compute a logarithm or multiplication that would need many digital gates (Sarpeshkar, Phil. Trans. R. Soc. A, 2014).

The catch: analog lacks the noise-restoration that makes digital composable, so errors accumulate and deep cascades are fragile without feedback. The most concrete deliverable is hardware: digitally programmable cytomorphic chips that simulate arbitrary biochemical reaction networks (Woo et al., 2018). More a unifying design framework than a turnkey reprogrammable platform.

Recombinase logic & memory (Endy, Stanford; Lu, MIT)

Use site-specific recombinases (serine integrases) to flip or excise DNA segments at defined sites. Each flip is a permanent, digital edit written into the genome itself, so the circuit computes and remembers in one stroke. State survives cell death and is readable by PCR/sequencing. Bonnet's amplifying gates couple the DNA flip to transcriptional gain for clean ON/OFF (Bonnet et al., Science 340:599, 2013); Siuti/Lu integrate logic and memory in living cells (Nat. Biotech 31:448, 2013); hierarchical design tools compose reliable multi-input logic (Nat. Commun., 2019; Sci. Rep., 2017).

6. Synthesis — who solves "reliable AND programmable"

7. Orthogonal bets — computing in voltage and form

The approaches above all live in the genetic-circuit frame: engineer DNA → transcription-factor logic → cell behaves like a programmed FPGA. The bets below don't fit that frame, and don't fit the soup/LLM frame (a vat that statistically maps inputs to outputs) either. They compute in membrane voltage, in the physical dynamics of tissue, in chemical kinetics, in the substrate's own physics. Computation here is a property of matter and electricity, not of edited code or learned weights. Maturity is uniformly lower; read the TRL column honestly.

Süel's bacterial K⁺ waves universalize Levin's bet below the animal kingdom: bioelectric signaling is not exclusive to brains. If voltage is a computational medium in microbes with no neurons and no genome editing, the electrical layer of all living matter is in play; brains are just the famous case.

8. Memory & modules — programming the layer above the cell

Zoom out from single circuits to the substrate services a program needs. Two of them: MEMORY — durable, writable, readable, inheritable state, so a computation done once stays done; and ORTHOGONAL MODULES — private information channels (replication, transcription, translation, delivery) that run inside the host without touching or being touched by native machinery. Recombinase state machines (§5) are the digital-latch corner of memory; here is the rest of the stack the report surfaces, and the private wiring underneath it.

a. Molecular recorders — SCRIBE, DOMINO, CAMERA

Where recombinases flip whole DNA blocks, base-editing recorders write at single-nucleotide resolution, accumulating an analog "tape" of a signal's magnitude × duration. SCRIBE (Farzadfard & Lu, Science, 2014) expresses ssDNA on input; a coexpressed recombinase drives targeted point mutations that accumulate across a population as an analog function of input strength × time. DOMINO uses base editing (CDA–nCas9–UGI, C→T) as a read/write DNA-state operator with ordered, cascade logic (AND/OR gates that record event sequence). CAMERA is a base-editor route to preprogrammed logic circuits with no double-strand break. The cell logs its own history in its genome, and the log preserves order.

b. Retron-based recorder — Retro-Cascorder

Retrons are bacterial elements whose reverse transcriptase makes a short DNA "receipt" (msDNA) from an RNA template. Retro-Cascorder (Schubert/Shipman et al., Nature, 2022) puts retron tags under different promoters; when a promoter fires, its tag is transcribed, reverse-transcribed to DNA, and integrated in temporal order into a CRISPR array — a time-stamped ledger of transcriptional events, not just exposure. Closer to logging the cell's internal program execution. Demonstrated in bacteria; not yet a routine mammalian tool.

c. Epigenetic writers — dCas9-DNMT3A, and the fastest path to clinic

Fuse dead Cas9 to chromatin effectors to write heritable, mitotically stable state without cutting DNA. The durable recipe pairs a DNA methyltransferase (DNMT3A / DNMT3A–3L) with a repressive histone writer (KRAB for H3K9me3, or Ezh2 for H3K27me3). Co-targeting DNMT3A + Ezh2/KRAB silenced HER2 for >50 days (~57 cell divisions) and flipped the locus to heterochromatin — "hit-and-run" silencing that persists after the editor is gone (Nucleic Acids Res., 2022). The edits are heritable through human hematopoiesis (PNAS, 2023).

d. OrthoRep — a private, hyper-evolvable genome

A cytoplasmic linear plasmid (p1) in yeast, replicated by a dedicated engineered error-prone DNA polymerase (TP-DNAP1) that does not touch the chromosome. Genes on p1 mutate ~100,000× faster than chromosomal genes (up to ~10⁻⁵ subs/base vs genomic ~10⁻¹⁰), enabling hands-off, continuous in-vivo directed evolution of any user gene (Ravikumar, Arrieta & Liu, Cell, 2018). Extended to bacteria as BacORep (phage GIL16 machinery in B. thuringiensis, Nat. Chem. Biol., 2023). An orthogonal mutation/replication channel — a private genome you can hyper-evolve without destabilizing the host.

e. Orthogonal ribosomes & recoded genomes — a private translation channel

Build a translation system that reads codes the natural ribosome ignores (Jason Chin, MRC-LMB). Ribo-Q1 decodes quadruplet codons on a dedicated o-mRNA; Ribo-T covalently tethers the subunits so the orthogonal ribosome never swaps with wild-type; orthogonal aaRS/o-tRNA pairs charge non-canonical amino acids onto the o-code. On the genome side, Syn61 (2019) recoded E. coli to 61 codons; Syn57 pushes to 57 via ~101,000 codon replacements across a 4 Mb synthetic genome, freeing up to 7 codons for non-canonical amino acids (Science 2025; bioRxiv May 2025). Commercial vehicle: Constructive Bio (genome writing → novel encoded polymers). A fully private translation channel — the cell builds molecules natural biology can't read, sequestered from the host proteome. The strongest literal instance of "program a layer the cell can't touch."

9. For a builder — where the systems work is

This page maps the science; it does not pick what to build. The build direction — standalone Rust systems artifacts with intrinsic value, the contribution path, and where to work — lives in What to Build.