Sequence plausibility
Candidates remain compatible with global and local statistical structure learned from the regulatory sequence corpus.
Target-conditioned sequence generation
Scientific Design Engine · Regulatory deoxyribonucleic acid (DNA) design
TrioDNADesign regulatory DNA with explicit cell-type intent
TrioDNA connects a learned DNA sequence prior, cell-aware activity evaluation, model-guided local editing, and population selection. The output is a ranked archive that keeps target activity, off-target context, regulatory motif evidence, and sequence diversity visible.
More search raises score while population diversity narrows
The project curve keeps complex and defined promoter targets visible across six generation budgets and shows the activity–diversity trade-off directly.
Evidence boundary. Per-budget dispersion is unavailable; the figure supports within-method sensitivity only.- Design object
- Regulatory DNA sequence
- Condition
- Requested cell type
- Search
- Generation plus local infill
- Output
- Evidence-bearing archive
01 / Abstract and project definition
Cell-Type-Aware Regulatory DNA Design
TrioDNA treats regulatory DNA design as a constrained population-search problem. Candidate sequences should remain plausible under a learned sequence prior, direct predicted activity toward a requested cell type, control off-target responses, preserve transcription-factor binding site evidence, and retain diversity across the final archive.
The method supports full-sequence generation and bounded local infill. A causal Transformer learns global and local sequence structure; cell activity measurements create directional supervision; compatible predictors return target and off-target score maps; and population search balances reward, regulatory alignment, and inter-sequence diversity.
02 / Scientific problem
A sequence can receive a high score and still be the wrong regulatory design.
One scalar objective cannot fully describe sequence plausibility, cell-type specificity, off-target behavior, transcription-factor binding site patterns, and population diversity. TrioDNA keeps these dimensions visible during generation, evaluation, and archive selection.
Cell-type specificity
Target activity is interpreted beside maximum and aggregate off-target activity and the resulting specificity margin.
Motif fidelity
Transcription-factor binding site frequency patterns provide a reference-facing constraint beyond activity alone.
Archive diversity
Selection returns multiple distinct high-scoring candidates rather than a single isolated optimum.
03 / Method overview
A generative prior proposes sequence variation; cell-aware and regulatory evidence determine what survives.
The tokenizer represents standard DNA bases and control tokens for autoregressive generation and fill-in-the-middle rewriting. Preference alignment biases the prior toward target-favoring sequences. Candidate batches receive cell score maps and sequence-quality audits, while population search creates local edits and selects survivors across a controlled generation budget.
Sequence prior
Autoregressive and fill-in-the-middle training teach full-sequence continuation and bounded local rewriting.
Output · Pretrained DNA modelCell-type alignment
Activity measurements create within-split labels and preferred/rejected sequence pairs that encode directional cell intent.
Output · Aligned sequence policyGeneration and editing
The system samples full sequences or rewrites bounded spans while preserving sequence length and untouched context.
Output · Candidate populationEvaluation and selection
Activity maps, off-target summaries, sequence audits, motif correlation, reward, and diversity drive archive selection.
Output · Ranked sequence archive04 / Architecture
Every handoff has an explicit scientific contract
Inputs, operations, evidence, and outputs stay named so the source of each decision can be reviewed.
- Input A
Regulatory sequence corpus
Normalized sequences establish the global and local sequence prior.
→ - Input B
Cell activity measurements
Compatible target and off-target values define directional supervision and evaluation context.
→ - Model
DNA sequence prior
A causal Transformer learns autoregressive continuation and fill-in-the-middle rewriting.
→ - Condition
Cell-type intent
Preference comparisons bias sequence likelihood toward the requested cellular context.
→ - Search
Candidate population
Full generation and local edits feed a population-level optimization loop.
→ - Select
Evidence-bearing archive
Reward, motif alignment, and sequence diversity determine the retained set.
05 / End-to-end workflow
From scientific input to an evidence-bearing decision
Each stage consumes a bounded object and leaves a reviewable artifact for the next stage.
- 01Prior corpus
Prepare
Normalize regulatory DNA, establish training and validation partitions, and define the sequence context.
- 02Sequence prior
Train
Learn full-sequence and local-span structure through autoregressive and fill-in-the-middle objectives.
- 03Labels and pairs
Align
Convert compatible activity measurements into target-aware labels and preference comparisons.
- 04Candidate batch
Generate
Sample new sequences or create bounded local edits from selected parents.
- 05Score map
Evaluate
Score target and off-target responses and attach sequence-quality and regulatory evidence.
- 06Population history
Search
Select survivors, create offspring, and repeat under a controlled generation budget.
- 07Sequence archive
Archive
Balance reward, motif alignment, and inter-sequence diversity under the experiment contract.
06 / Experimental design
Questions, protocols, and comparison boundaries remain attached
Every study below describes what was varied, what was measured, and where interpretation must stop.
Cell-specific 200 base-pair design
For HepG2, K562, and SK-N-SH, each target and seed scores an archive of 8,192 candidates and retains 256 through motif-aware selection. Five seeds report mean and standard deviation for reward difference and motif correlation.
- Cell types
- 3
- Sequence length
- 200 bp
- Seeds
- 5
- Archive per target and seed
- 8,192 → 256
Generation-budget sensitivity
Complex and defined promoter targets are evaluated across six generation budgets. Raw top and medium activity summaries are read beside sequence diversity; 20 generations is the selected operating balance.
- Targets
- 2
- Budgets
- 5–100 generations
- Confirmation seeds
- 5
- Selected operating point
- 20 generations
07 / Results and evidence
Read the signal with its evidence boundary
Quantitative summaries, structural views, and scientific interpretation remain separate layers.
Enhancer dual-gate passes
TrioDNA exceeds the protocol-matched Ctrl-DNA control baseline on reward difference (ΔR) and motif correlation for every shared cell
Largest motif-correlation gap
SK-N-SH five-seed mean difference versus the Ctrl-DNA control baseline
Complex-target diversity gain
Five-seed mean versus the TACO reinforcement-learning promoter baseline
Defined-target diversity gain
Five-seed mean versus the TACO reinforcement-learning promoter baseline
All three enhancer targets clear both matched gates
For HepG2, K562, and SK-N-SH, TrioDNA exceeds Ctrl-DNA on reward difference and transcription-factor binding site motif correlation under the same cells and evaluation protocol. The largest gaps occur for SK-N-SH: +0.084 reward difference and +0.466 motif correlation.
Boundary. These are predictor and set-level motif results; no position-level motif hits, cellular measurements, or wet-lab enhancer activity are shown.Search depth increases raw score while diversity contracts
Across six budgets, complex-target raw top score rises from 20.55 to 31.70 while diversity falls from 57.0 to 52.8. Defined-target top score rises from 20.05 to 37.17 while diversity falls from 58.0 to 52.0. Twenty generations is retained as the operating balance.
Boundary. Per-budget standard deviations are unavailable in the curated snapshot, so this curve is within-method sensitivity rather than a significance analysis.Protocol-matched diversity remains above TACO
In five-seed confirmation, TrioDNA diversity is 54.80 ± 2.48 for the complex target and 54.40 ± 4.41 for the defined target, compared with 52.80 ± 3.49 and 49.60 ± 1.79 for TACO.
Boundary. Only the shared diversity metric is directly comparable; the reported activity scales differ and remain separate.Reading key
Scientific abbreviations
- TFBS
- Transcription-factor binding site, used here through set-level motif-frequency correlation.
- ΔR
- Reward difference, a target-aware activity summary used by the enhancer evaluation.
- TACO
- A reinforcement-learning promoter-design baseline used for the protocol-matched yeast diversity comparison.
08 / Limitations and provenance
What the current evidence can establish
Only results using compatible target models and evaluation protocols are described as direct comparisons.
Predictor evidence is protocol-specific
Direct comparison requires the same target models, cell definitions, sequence length, selection procedure, and evaluation protocol.
Some dispersion is unavailable
The yeast budget curve lacks per-budget standard deviation in the curated public snapshot.
Activity scales differ
TrioDNA raw oracle summaries and TACO paper-facing activity ceilings use different scales; only protocol-matched diversity is compared directly.
Biological validation is pending
The current evidence does not establish enhancer or promoter activity in cells, tissue specificity, safety, or therapeutic performance.