Designed promoters

A dataset of novel promoter sequences. A promoter is the stretch of DNA in front of a gene that decides how strongly it is switched on; each of these is 600 bp of DNA written by a genetic algorithm, not copied from any organism and existing in no genome. Each was scored twice — once with the target cell type described to the model and once with the cell type it should stay silent in — and selected on the difference, the margin. It answers the question the natural database cannot: can we build something better than the genome already offers?

832 candidates recorded, 402 clear the selection rule, 64 selected, across 2 on/off cell pairs. Every candidate from every production run is published, passing or failing, with the record of which rules it clears and by how much. Margins here are measured inside one named genomic window — the scaffold — and mean nothing without it, so every record names its own.

Pair codes. Each pair is one ON cell type and one OFF cell type. A1 ventricular cardiomyocyte ON / hepatocyte OFF · A2 skeletal muscle myofiber ON / hepatocyte OFF · A3 ventricular cardiomyocyte ON / skeletal muscle myofiber OFF · A4 CNS neuron ON / hepatocyte OFF · A5 CNS neuron ON / astrocyte OFF. A code is a shorthand for that pair of cell types and nothing else.

There is a second database, and it is strong exactly where this one is weak. Nothing here can be recalled from the model’s training data, because none of it was ever in any genome to be trained on — and for exactly the same reason, no laboratory has ever measured any of it. The natural promoters have the opposite shape: they already exist, so measured laboratory evidence for them exists too, and a good rank there may be partly recall of the model’s training data.

The two are separate releases with separate file names and separate limitation blocks. Do not merge them into one table or one ranking — they carry different caveats. Exactly one quantity is allowed to cross between them — a design's margin against the best natural promoter in the same window — and it has its own comparison view, labelled as the only one.

Limitations

RESEARCH USE ONLY, NON-COMMERCIAL. Not wet-lab validated. Not clinical. Every number in this release is a prediction of the g0-expression model on checkpoint 20260523. Nothing here has been validated in a wet lab, in any cell type, at any length, in any construct. Nothing here is for clinical, diagnostic or therapeutic use. These are design hypotheses offered for collaborative validation, not products.

L1Not wet-lab validated, and not clinicalNo sequence in this release has been synthesised, transfected, transduced or measured in any cell. Every number is a model prediction. Research use only, non-commercial. Nothing here may inform a clinical, diagnostic or therapeutic decision.

L2Selective on the delivered property — by selection, not by learned repressionSELECTIVE ON THE DELIVERED PROPERTY: strong in the target cell type, and measured silent in the off-target cell type, in a scaffold that is silent in both when empty and whose off-target arm is demonstrably live. MECHANISM FOOTNOTE, required wherever the mechanism is at issue: this selectivity is SELECTION, NOT LEARNED REPRESSION. The natural seeds were already silent in hepatocyte, so there was no off-target gradient to descend; what the optimiser had to do was hold that arm silent while raising the on arm, and it did. Active descent of an off arm that started ACTIVE is demonstrated only in a separate, differently-seeded experiment (hepatic 4.844 -> 0.005), whose own caveats travel with it: culling not conversion, and silence not repression. Do not say the model was taught to silence the off-target arm.

L3The ratio is the quantity, not the raw marginA margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two experiment descriptions. Measured: re-running both pairs in a safe-harbour scaffold raised every raw margin and LOWERED the ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x).

L4No cross-model corroborationNo cross-model corroboration exists for these designs. Cross-model validation was attempted and halted: three control gates failed on cell-type coverage and a fourth on a composition confound, and only 4 of 10 A-tier coverage arms -- the five pairs' ON sides and OFF sides, counted separately -- are reachable by any alternative model that can vote positively at all. Every number here comes from one model on one checkpoint.

L5One checkpoint, with a compressed dynamic rangeCheckpoint 20260523 preserves marker-gene ranking but compresses magnitudes. The reportable margin threshold is 0.80 log units, which is close to the full width of the band most real genes occupy (0.03-0.9). Compare deltas, never absolute magnitudes, and never across scaffolds.

L6Scaffold-boundEvery score here was measured with the 600 bp design substituted for the proximal promoter of one fixed real gene TSS window (FAM47C). That is a real gene TSS window, the same ~9.2 kb shape the model was trained on, and it is not the locus a design would be used at. A margin from this scaffold is not comparable in magnitude with a margin from another. The identical 600 bp ALB promoter scores -0.012 in one admitted scaffold and +4.844 in another.

L7Not measured in an AAV cassetteA transgene cassette in a plasmid is out of distribution for this model. These designs were evolved and scored in a genomic window, not in a cassette. A cassette-optimised ranking did not transfer in this project: the best design from a cassette-environment run ranked 63rd of 64 on the genomic axis.

L8The on-target reporting floor excludes nothingBoth pairs report an on-target floor derived from a small panel of human genomic control promoters scored in this same scaffold before any design was scored. In both pairs `floor / negative_on_max` is below 1, which means the floor is cleared by some random DNA in this scaffold and carries no weight. The banded margin criterion is the one doing the work. Both counts are reported so 'both criteria met' cannot be read as two independent checks.

L9Novelty is bounded against the run's own seedsEdit distance is measured over the 600 bp region against the 25 natural seeds of that run. It bounds recall of those seeds, not of every promoter in the genome.

L10A third pair was attempted and no pair selected from itA third pair (CNS neuron ON vs hepatocyte OFF, A4) was generated, and no selected pair came out of it: nothing among its candidates is selected, so it has no pair record beside A1 and A2. Its 128 candidates are nevertheless published, in the harvested tables, each marked with the criterion it fails — see L11 for why a record is not a design that passed, and L13 for the measurements that stopped this one. A failure that is countable is worth more than a failure that was deleted.

L11The six criteria are recorded, not enforced — a record in this release is not a design that passedThis release publishes EVERY candidate from every production generation run, each carrying its own record of the six selection criteria: which of them it meets, and by how much. The criteria are recorded rather than enforced. A candidate that fails one is published marked as failing it instead of being deleted, so the criteria themselves are unchanged and what changed is what happens to a candidate that misses one. Three different counts follow, and they are not interchangeable: the number RECORDED is how many candidates were generated and says nothing about any criterion; the number CLEARING THE SELECTION RULE is how many meet all six; the number SELECTED is the curated per-pair selection. Check whether a record was selected, and read its `gate_record`, before quoting it.

L12Six design environments, and only the ratio crosses between themThe harvested records span FAM47C (the scaffold of record, 384 records), FOXB2 (superseded, 192), CCR5site_plus and AAVS1_plus (safe-harbour integration coordinates, both rejected as design environments, 128 and 64), Keppel19_minus (the A5 integration site, 64: its designs reach a median margin of 3.961 there and 0.090 re-scored in FAM47C, with a rank correlation of 0.0724 between the two, which is why A5 is not selected) and a plasmid cassette (out of distribution, segregated and never tabulated). A raw margin is meaningless without its scaffold named. Only the design-over-nature ratio against the best natural promoter in the SAME window and the SAME run crosses a scaffold boundary, and every run carries that baseline.

L13The A4 records are published on a hollow admissionThe 128 A4 (CNS neuron ON / hepatocyte OFF) records are published with their criteria recorded rather than enforced, and nothing among them is selected. FAM47C's admission for A4 is carried by exactly one neuronal control, SYT1, by 0.0014 log units — and SYT1 is also A4 seed 7, so the environment was admitted by the optimiser's own starting material rather than by an independent control. All 64 FAM47C candidates are more active in hepatocyte than that run's own random-DNA negatives. The 64 AAVS1_plus candidates were evolved in an environment whose EMPTY scaffold is already active on the neuronal arm, and they collapse to a median 0.802 when re-scored in a real gene TSS window. All three measurements travel with every A4 record: the admission carried by a single control, SYT1, by 0.0014 log units and that control is also A4 seed 7; all 64 FAM47C candidates more active in hepatocyte than their own run's random-DNA negatives; and the 64 AAVS1_plus candidates collapsing to a median 0.802 in a real gene TSS window.

L14Both selected scaffold admissions rest on a single carrier — disclosed, not re-gatedThe advisory hollow-admission check (reported under `scaffold.admission.hollow_admission_check` on every record) fires on both SELECTED pairs, not only on the pair that was dropped. **A2 / FAM47C** is admitted on ONE carrier of five, `mck__published_cassette`, clearing its own banded floor by 0.0302 — a mouse synthetic fusion with no genomic TSS of its own, while all three HUMAN genomic members of the same panel failed. **A1F / FAM47C** has a single-member panel by construction: `ctnt__region_600bp`, i.e. TNNT2, which is also A1 seed 11 — structurally the same defect that disqualified A4/FAM47C (SYT1 alone, by 0.0014, and SYT1 is A4 seed 7), one pair earlier and in a weaker form. **No verdict moves and nothing is re-gated:** the four admission criteria were fixed before any scaffold was scored, and adjusting them after seeing where they landed is exactly the failure they exist to prevent. Read an admission here as passed, not as broad, and read A4's exclusion as the strong form of a weakness the selected pairs share. **A2's weakness is narrower than it first reads.** A 110-pair triage table re-derives admission under a mechanical ON panel — the ON context's own committed marker genes as 600 bp genomic windows, a rule whose inputs all predate that work — and under it A2 / FAM47C is admitted by THREE human genomic carriers, NEB, MYF6 and TNNI1. So the single-mouse-fusion carrier is a property of the curated panel we happened to use, not evidence that no human genomic promoter can carry this environment: a second, independently-derived and stronger basis for the same admission. It re-gates nothing, and it does not transfer to A1F, whose one-member panel is a genuine single-carrier case under both rules. Triage is not admission — that table's panel differs from the curated procedure, so a green cell in it may not be quoted as a curated admission.

Designed promoters

Every record is subject to L1–L14 above.

832 candidates recorded. 402 clear the selection rule. 64 selected. Every candidate from every production generation run is here, passing or failing, each carrying its record of the six selection criteria: which it meets, and by how much. A record here is not a design that passed. Being recorded here does not mean a candidate cleared the selection rule. Each record's glyph strip and margin axis state which rules it clears and by how much.

Why 402 clear the rule and 64 are selected

402 is 12 separate shortlists, not one. Each production run produced its own, and two of them are the runs this project publishes. Selected is a comparability grade, not a quality grade: the other 338 clear the same rule, several of them at larger raw margins, and they are not in the selection because they were measured in a different design environment or belong to a pair we do not publish — a margin is a property of the window it was measured in, so those numbers cannot be read down one table beside these.

run	clears the rule	why it is not in the selection
`A1FAM_floorrule`	17	selected — the run of record for A1, in FAM47C
`A2_prod_genomic_module`	47	selected — the run of record for A2, in FAM47C
`CCR5_A1_margin`	50	not in the selection — in CCR5site_plus (REJECTED as a design environment; adopted for A2's control panel); recorded role superseded scaffold
`A1FAM_margin`	48	not in the selection — in FAM47C (scaffold of record); recorded role sibling of the selected A1 run, same scaffold, margin fitness
`A1_prod_genomic_module`	45	not in the selection — in FOXB2 (superseded design environment); recorded role superseded scaffold
`A5_Keppel19_margin`	45	not in the selection — in Keppel19_minus (rejected for A5); recorded role attempted, selected, NOT in the release
`A1_prod_genomic_module_floorfix`	34	not in the selection — in FOXB2 (superseded design environment); recorded role superseded scaffold
`A4_AAVS1_margin`	34	not in the selection — in AAVS1_plus (rejected for A4); recorded role attempted, not selected
`CCR5_A2_margin`	25	not in the selection — in CCR5site_plus (REJECTED as a design environment; adopted for A2's control panel); recorded role superseded scaffold
`A1_prod_genomic_long`	23	not in the selection — in FOXB2 (superseded design environment); recorded role superseded scaffold, long run
`OFFARM_FAM47C_margin`	20	not in the selection — in FAM47C (scaffold of record); recorded role off-arm descent experiment
`OFFARM_FAM47C_floorrule`	14	not in the selection — in FAM47C (scaffold of record); recorded role off-arm descent experiment

338 not selected, plus 64 selected, is 402. Every candidate from every one of these runs is in the harvest below with its full record of the six criteria, passing or failing; none of them is deleted.

Downloads: designs.tsv — the selected records, flat; harvest.tsv — every recorded candidate with its six criteria, passing or failing; designs.fasta — the sequences themselves.

Taking these to a vendor: the synthesis handoff carries the same 64 designs as a submittable FASTA and a per-construct order sheet, with the synthesizability read beside every one. Order-ready means a vendor's intake would accept the file. It does not mean validated.

How to read a record

Margin — scaffold-bound. The RATIO is the quantity, not the raw margin. A margin is scaffold-bound: a scaffold with more dynamic range lifts the designs and the natural promoters together. The comparable number is a design's margin divided by the best natural promoter's margin measured in the SAME scaffold, in the SAME run, under the SAME two experiment descriptions. The re-run that demonstrates it is limitation L3 above.
The four marks on the axis. Left to right by value, not by importance: this run's composition-matched random-DNA negatives at their most active; the record's own noise floor, the margin it has to beat to be reportable at all; the best natural promoter measured in the same window, the same run and under the same two verbatim experiment descriptions; and the record itself. Only the ratio between the last two crosses a scaffold boundary.
Both experiment descriptions, unedited. Every record carries the ON and the OFF description exactly as it appears in the model’s training data. A few contain small oddities — a typo, or a detail that disagrees with another line of the same record. Those are left alone on purpose: the description is how the model identifies which experiment is being asked about, and correcting one can change the answer. Why, on the overview.
The glyph strip is in a fixed order and means the same thing on the natural page. A grey square is not a failure — it means no measurement exists.
Controls are never averaged. They are reported individually, per run, because the panel members are different species, different lengths and different construct routes.

the check ran and this record cleared itthe check ran and this record is marginal, or clears it with a caveatthe check ran and this record did not clear itno measurement exists — not a failure, an absence

The criteria are not equally informative. The one that removes the most records is also the least resolvable: 228 of the 334 off-arm failures in this corpus fail by less than 0.0181, the width by which one score moves between the two implementations of this same checkpoint, so a narrow failure is an undetermined sign and not a near miss. And the ON-target reporting floor excludes nothing in this scaffold — L8 above.

The two selected pairs

A1 — cardiomyocyte vs hepatocyte

Ventricular cardiomyocyte ON / Hepatocyte (liver parenchymal cell) OFF. An intravenous AAV dose loads the liver first, so a cardiac gene therapy that uses a ubiquitous promoter expresses its transgene in hepatocytes at high level. The promoter is where that de-targeting has to happen. The field standard for the cardiac arm is cTnT (the TNNT2 promoter).

17 designs selected from this pair. In the FAM47C scaffold their margins run 6.3228 down to 5.9169, which is 1.351× down to 1.2643× the margin of the best natural promoter measured in that same window and the same run (TNNT2, 4.6801 in FAM47C) — a ratio of two margins on the log scale, not a fold change in expression. The empty FAM47C window scores 0.0325 ON / -0.0339 OFF, so the window is not carrying the signal; this run's 12 composition-matched random-DNA negatives reach a margin of 0.1416 in the same window.

The same result off the log scale. Exponentiating the two margins puts them back on the normalised TPM scale, where the best design separates its two arms 5.21× as far as the best natural promoter in the same FAM47C window does — a fold-change in normalised TPM, never raw TPM, since exponentiating inverts the logarithm and the +1 and nothing inverts the quantile normalisation, and read against the one defect of this checkpoint that this conversion magnifies: Checkpoint 20260523 preserves marker-gene ranking but compresses magnitudes. Only the ratio of the two converted numbers carries any meaning; neither of them alone is a TPM anybody measured, which is why neither is printed. The log-scale figure — 1.351× on ln(quantile-normalised TPM + 1) — is the one the release files and every other page quote, and it stays the headline.

Scaffold admission: pre-named ON-standard panel of 1 member(s), 1 clearing C1 and 1 carrying C2. SINGLE-CONTROL ADMISSION: one member is the whole panel and it carries both criteria. So the admission says this scaffold expresses that one known on-target element far above its negatives, with a reportable margin. It does NOT say the scaffold expresses on-target elements in general -- one control cannot show that. Read it as the weakest form of a passed admission. Note also that this element is normally in the pair's natural_baseline as well, so the admission control and the baseline are not independent of each other.

The control promoters from this same run, individually

Never average these. cTnT is chicken; MHCK7 and CK8e are mouse synthetic fusions with no genomic TSS; TBG-as-shipped is a 669 bp composite; hSYN1 is the only control whose human gene is in the model's test split; and the lengths differ, so rows must not be compared with each other or with a 600 bp design.

control	species	bp	ON	OFF	marginin FAM47C	verdict
`ctnt__published_cassette`	Gallus gallus (chicken)	429	4.2812	-0.0172	4.2984	reportable
`ctnt__published_cassette_genomic_core`	Gallus gallus (chicken)	414	4.2812	-0.0245	4.3057	reportable
`ctnt__region_600bp`	Homo sapiens	600	4.6562	-0.0239	4.6801	reportable
`mck__published_cassette`	synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)	792	1.4609	-0.0294	1.4903	reportable
`mck__published_cassette_ck8`	synthetic construct (mouse Ckm-derived)	450	1.3906	-0.027	1.4176	reportable
`mck__published_600bp`	synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)	600	0.2061	-0.0289	0.235	not_a_finding
`mck__region_600bp`	Homo sapiens	600	0.0615	-0.0334	0.0949	not_a_finding
`hsyn1__published_cassette`	Homo sapiens	470	0.1079	-0.0168	0.1247	not_a_finding
`hsyn1__published_cassette_genomic_core`	Homo sapiens	456	0.1187	-0.0105	0.1292	not_a_finding
`hsyn1__region_600bp`	Homo sapiens	600	0.1113	-0.0081	0.1194	not_a_finding
`tbg__published_cassette`	Homo sapiens (synthetic composite)	669	0.064	1.3828	-1.3188	not_a_finding
`tbg__published_cassette_tbg_bare`	Homo sapiens	410	0.0376	1.2812	-1.2436	not_a_finding
`tbg__published_600bp`	Homo sapiens (synthetic composite)	600	0.0737	2.0781	-2.0044	not_a_finding
`tbg__region_600bp`	Homo sapiens	600	0.0576	1.3906	-1.333	not_a_finding
`nppa__region_600bp`	Homo sapiens	600	0.4473	-0.0344	0.4817	not_a_finding
`nppb__region_600bp`	Homo sapiens	600	0.1235	-0.0165	0.14	not_a_finding
`myh7__region_600bp`	Homo sapiens	600	0.3984	-0.0444	0.4428	not_a_finding
`ryr2__region_600bp`	Homo sapiens	600	0.2051	-0.0161	0.2212	not_a_finding
`des__region_600bp`	Homo sapiens	600	0.0554	-0.0267	0.0821	not_a_finding

A2 — skeletal muscle myofiber vs hepatocyte

Skeletal muscle myofiber ON / Hepatocyte (liver parenchymal cell) OFF. The same delivery problem as A1 for systemically dosed muscle programmes: the vector reaches the liver whatever the target is, so the liver arm is the one that has to be silent. The field standards for the muscle arm are MCK-derived elements (MHCK7, CK8e), which are striated rather than skeletal-restricted.

47 designs selected from this pair. In the FAM47C scaffold their margins run 4.5822 down to 4.1451, which is 2.3083× down to 2.0881× the margin of the best natural promoter measured in that same window and the same run (TRIM54, 1.9851 in FAM47C) — a ratio of two margins on the log scale, not a fold change in expression. The empty FAM47C window scores 0.0095 ON / -0.0339 OFF, so the window is not carrying the signal; this run's 12 composition-matched random-DNA negatives reach a margin of 0.063 in the same window.

The same result off the log scale. Exponentiating the two margins puts them back on the normalised TPM scale, where the best design separates its two arms 15.40× as far as the best natural promoter in the same FAM47C window does — a fold-change in normalised TPM, never raw TPM, since exponentiating inverts the logarithm and the +1 and nothing inverts the quantile normalisation, and read against the one defect of this checkpoint that this conversion magnifies: Checkpoint 20260523 preserves marker-gene ranking but compresses magnitudes. Only the ratio of the two converted numbers carries any meaning; neither of them alone is a TPM anybody measured, which is why neither is printed. The log-scale figure — 2.3083× on ln(quantile-normalised TPM + 1) — is the one the release files and every other page quote, and it stays the headline.

Scaffold admission: pre-named ON-standard panel of 5 member(s), 2 clearing C1 and 1 carrying C2. The admission is real but it rests on a single member's margin: the other panel members' margins are in c2.member_margins below and most of them are not findings. Read the admission as passed, not as broad.

The control promoters from this same run, individually

control	species	bp	ON	OFF	marginin FAM47C	verdict
`ctnt__published_cassette`	Gallus gallus (chicken)	429	1.7031	-0.0172	1.7203	reportable
`ctnt__published_cassette_genomic_core`	Gallus gallus (chicken)	414	1.7109	-0.0245	1.7354	reportable
`ctnt__region_600bp`	Homo sapiens	600	0.3164	-0.0239	0.3403	not_a_finding
`mck__published_cassette`	synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)	792	0.8008	-0.0294	0.8302	reportable
`mck__published_cassette_ck8`	synthetic construct (mouse Ckm-derived)	450	1.2109	-0.027	1.2379	provisional_within_noise
`mck__published_600bp`	synthetic construct (mouse Ckm-derived enhancer/promoter + mouse alpha-MHC enhancer)	600	0.3027	-0.0289	0.3316	not_a_finding
`mck__region_600bp`	Homo sapiens	600	0.032	-0.0334	0.0654	not_a_finding
`hsyn1__published_cassette`	Homo sapiens	470	0.042	-0.0168	0.0588	not_a_finding
`hsyn1__published_cassette_genomic_core`	Homo sapiens	456	0.0342	-0.0105	0.0447	not_a_finding
`hsyn1__region_600bp`	Homo sapiens	600	0.0398	-0.0081	0.0479	not_a_finding
`tbg__published_cassette`	Homo sapiens (synthetic composite)	669	0.0291	1.3828	-1.3537	not_a_finding
`tbg__published_cassette_tbg_bare`	Homo sapiens	410	0.0287	1.2812	-1.2525	not_a_finding
`tbg__published_600bp`	Homo sapiens (synthetic composite)	600	0.0361	2.0781	-2.042	not_a_finding
`tbg__region_600bp`	Homo sapiens	600	0.028	1.3906	-1.3626	not_a_finding
`des__region_600bp`	Homo sapiens	600	0.0344	-0.0267	0.0611	not_a_finding
`myl1__region_600bp`	Homo sapiens	600	0.0206	-0.0454	0.066	not_a_finding

What was attempted and not selected

A4 — not in the release (recorded status attempted, not selected). Both candidate design environments fail, and they fail differently. FAM47C's admission for this pair is carried by one control, SYT1, by 0.0014 log units; the AAVS1_plus candidates collapse to a median 0.802 when re-scored in a real gene TSS window. 0 selected.
A5 — not in the release (recorded status attempted, not selected). The only environment of twelve that admitted this pair is a safe-harbour integration site, and the margin the run produced there does not exist anywhere else. Same shape as A4/AAVS1_plus, worse.
CCR5-A1 / CCR5-A2 — not in the release (recorded status superseded). Higher raw margins, WORSE design-over-nature ratio in both pairs (A1 1.351x -> 1.168x, A2 2.308x -> 1.256x): the scaffold lifted the natural baseline more than it lifted the designs, so the ratio and not the raw margin is the quantity.

Their candidates are in the harvest above rather than deleted. The per-candidate measurements that stopped each pair are published as a machine-readable file: download not_selected.json.

What this release was built from

Every record carries both of its experiment descriptions verbatim, and the full set — with complete SHA-256 digests and measured paraphrase floors — is in contexts.json (JSON download). They are reproduced verbatim and must not be corrected; why is on the overview.

Every record in designs.jsonl carries a reproduce block: the scaffold window and its checksum, the coordinates the 600 bp design is substituted into, the request body and the expected score. The window discipline those steps assert — and what silently goes wrong when it is not asserted — is on the overview, under exactly how a window is built.

The release as a whole — the model id, revision, checkpoint, unit and endpoint, the file inventory and the limitations block — is in MANIFEST.json (JSON download).