DNA/RNA sequence design: mRNA, regulatory elements & strains

Expression of a gene depends on far more than its protein sequence: codon choice, mRNA structure, UTRs, promoters and the host's metabolism all matter. Nucleotide language models and task-specific ML now propose sequences that are scored for expression and stability before they are synthesised, so design-build-test-learn cycles need fewer wet-lab rounds. Typical users are mRNA-therapeutic developers, enzyme and ingredient producers and synthetic-biology teams.

Problem & value

• Today: a protein is back-translated with a codon-usage table (or a vendor tool), a standard UTR and promoter are chosen, and dozens to thousands of variants are screened in the lab. Rules of thumb (codon adaptation index, GC content) correlate only moderately with real expression, and results depend on host, context and delivery.
• AI value: learn the sequence-to-function map from the organisation's own screens plus pre-trained nucleotide models, then pick the variants most likely to work. Gains are measured in fewer screening rounds and higher titre or protein yield; published gains are assay-specific, so benchmark against your own baseline rather than quoting vendor numbers.
• EU context: the EU Biotech and Biomanufacturing initiative and the EU Biotech Act proposal aim to strengthen European biomanufacturing; mRNA capacity (BioNTech, CureVac, Moderna sites in Europe) and industrial biotech (enzymes, ingredients, biochemicals) are the main demand.

Variants

• mRNA therapeutics and vaccines: codon optimisation, 5'/3' UTR design, poly-A and modified-nucleoside choices, secondary-structure and stability prediction, reduced innate-immune side effects.
• Regulatory elements: promoters, enhancers, ribosome binding sites, terminators, riboswitches; cell-type-specific elements for gene therapy (AAV payload size limits).
• Microbial strain engineering: pathway genes, expression balancing, genome edits (E. coli, Bacillus, yeast, Corynebacterium, CHO) for enzymes, proteins and small molecules.
• Gene circuits and cell-free systems: logic, sensors and biosensors, cell-free expression.
• Plants and animals: regulatory edits to support breeding programmes, see SCI-020.
• Private use: none sensible; this is a laboratory-bound, regulated activity.

How it works

1. target + host + constraints
2. candidate generation (rules, generative nucleotide LM, evolutionary / Bayesian search)
3. in-silico scoring (expression, stability, structure, immunogenicity, motif and IP filters)
4. biosecurity and sequence screening
5. synthesis order (DNA vendor)
6. build (cloning, IVT, transfection, strain construction)
7. test (plate assays, MPRA, RNA-seq, titre)
8. data back into the model (active learning)

• Nucleotide language models (Evo 2, Nucleotide Transformer, Caduceus, DNABERT-family) are pre-trained on genomes and give zero-shot likelihood scores and embeddings; Evo 2 can also generate sequence.
• Supervised predictors (CNNs, gradient boosting, small transformers) trained on MPRA or screening data usually remain the most reliable scorers for one defined task and host.
• mRNA-specific tools combine codon choice with minimum-free-energy and structure objectives (for example LinearDesign for joint codon and structure optimisation).
• Closed loop: Bayesian optimisation or active learning picks the next batch; connect to lab automation (SCI-007). Human-in-the-loop: scientists set constraints and approve the order; a biosafety officer signs off on biosecurity screening before anything is synthesised.

Data preparation

Needed data are the host's codon-usage tables, your own assay results (plate-reader, MPRA, RNA-seq, titres) and the sequence designs behind them. In practice these sit in ELN/LIMS entries, instrument exports on lab file shares, Excel sheets of individual scientists and sometimes paper lab notebooks. Typical gaps are measurements without the exact sequence or construct ID, inconsistent units and plate layouts, undocumented assay conditions, failed runs mixed with valid ones, and duplicate or renamed constructs.

Preparation steps: 1. Define one construct ID scheme and link every measurement to its full sequence (see PREP-002 for gap analysis and PREP-004 for data ownership). 2. Export instrument and ELN/LIMS data through APIs or scripted exports (PREP-005), and move Excel results into tidy tables (PREP-011). 3. Record assay metadata (host, medium, time point, replicate) and flag failed runs. 4. Hold out whole sequence families as a test set (PREP-018). 5. Treat unpublished sequences as trade secrets and classify them before any cloud use (PREP-024).

Expect roughly 30 to 40 percent of the total effort in data preparation. Without clean sequence-to-measurement pairs, model scores are unreliable whatever the model. Public data (for example MPRA datasets) can bootstrap a first model, but your own assays decide the final quality.

Implementation paths

Which path for whom

• Enterprise: Path B, because design data and unpublished sequences are core IP, but hosted nucleotide models (BioNeMo/NIM in an EU region) avoid building GPU operations; move to C if sequences must never leave the site.
• Developer: Path C, because open-weight models such as Evo 2 plus ViennaRNA and BoTorch give full control and reproducibility at HW3 level.

Path A — Buy (off-the-shelf)

• Design platforms and services: Ginkgo Bioworks (Datapoints, model API), Asimov (genetic design for cell and gene therapy and CHO), Twist Bioscience (design and synthesis), Benchling (R&D data and design workflow), Profluent and Cradle (protein-first, sometimes with DNA features).
• EU gene-synthesis options: Eurofins Genomics (DE) and Azenta/GENEWIZ EU sites offer codon-optimised gene synthesis; check where design data and sequences are processed, as pre-publication IP is sensitive.
• Choose A when: you want a result, not an in-house modelling team, and the design problem is standard (secreted protein in CHO or E. coli).

Path B — Build on cloud APIs (recommended default)

• Hosted nucleotide models: Evo 2 through NVIDIA BioNeMo / NIM microservices, and hosted endpoints from model providers; Arc Institute provides model access and weights.
• Orchestration: Python with Biopython, a Bayesian-optimisation library (BoTorch/Ax) and the LIMS or ELN (Benchling API); LLM agents (llm-general) help with literature, constraint specification and experiment records, but not as the sequence scorer.
• Choose B when: you have own assay data and a small ML team, and sequences are not so secret that they must stay on-premises.

Path C — Self-host (open source)

• Models: Evo 2 (open weights, 7B and 40B parameter variants; 1 Mb context), Nucleotide Transformer, Caduceus, DNABERT-2, RNA-FM and similar RNA models; Enformer / Borzoi-class models for expression prediction from sequence; ViennaRNA and LinearDesign for structure.
• Stack: PyTorch, vLLM-style serving or the Evo 2 reference code, BoTorch, Snakemake or Nextflow for pipelines.
• Hardware: the 7B models fit on HW3; the 40B model needs HW3/HW4 with quantisation or multi-GPU. Fine-tuning needs HW4.

Path D — Custom training

Worth it when you own a large proprietary screening dataset (thousands to millions of sequence-measurement pairs in your host and format). Fine-tune a pre-trained nucleotide model or train a task CNN/transformer on MPRA-style data. Quality and reproducibility of the assays matter more than model size; hold out whole sequence families when validating.

State of the art (2026)

• Foundation models: Evo 2 (Arc Institute, NVIDIA and partners) was trained on genomes across all domains of life, is released with open weights and code, and shows zero-shot variant-effect prediction and genome-scale generation. Generation of functional, novel complete genomes has been reported only for small phages, with extensive wet-lab validation.
• Benchmarks: zero-shot likelihoods help for variant effects, but for designing UTRs and promoters in a given host, supervised models trained on matching data still often win. Community benchmarks and competitions (for example the Random Promoter DREAM Challenge) show this.
• mRNA design: structure-aware codon optimisation (LinearDesign) and ML on MPRA data for UTRs are standard in leading mRNA companies; in vivo translation and immunogenicity do not always follow in-vitro predictions.
• Limits and failure modes: hallucinated sequences with plausible statistics but no function; poor transfer across hosts, cell types and delivery formats; training-data bias towards well-studied organisms; long-range effects (chromatin, context) are weakly captured; codon "optimal" can mis-fold proteins (co-translational folding needs slow codons).

Shared resources & combinations

• Same stack (scientific-fm, GPU HW3): SCI-001 protein design (design the protein, then the nucleotide sequence), SCI-009 genomics and variant interpretation (shared DNA models), SCI-017 multi-omics, SCI-020 breeding.
• Upstream/downstream: SCI-002 drug discovery (payloads, targets), SCI-007 self-driving labs (closed-loop build-test), the shared LIMS/ELN data layer.
• Shared infrastructure: active-learning loop, MPRA data pipelines and GPU inference are reusable across these use cases.

Effort, cost & KPIs

• Path A: weeks to start; project or service pricing, plus synthesis cost per construct.
• Path B: 3-6 months for a working design loop; costs are dominated by assays and DNA synthesis, not by GPU or API usage.
• Path C/D: 6-12+ months with ML and bioinformatics staff; GPU cost moderate, data generation the main investment.
• KPIs: hit rate of top-N predicted designs versus baseline, rounds to target titre or expression, rank correlation between predicted and measured expression, cost per useful variant, cycle time of the design-build-test loop.

People & change

Start with management and the R&D department lead, who decide on scope, budget and what counts as secret IP. Involve the scientists who run the assays early, as they own the data and must trust the scores; the biosafety officer and legal/compliance sign off on sequence screening and synthesis ordering. IT and IT security join when data leave the site (API access, vendor contracts, see GUIDE-010 and GUIDE-013).

Works council: usually not required, since the system analyses sequences and assay data, not employee behaviour or performance (§87(1) no. 6 BetrVG). Still inform the works council in good time about AI-supported work processes (§90(1) no. 3 BetrVG), especially if usage logs or the ELN are shared. No personal data are processed, so a DPIA is normally not needed; ask the DPO to confirm if user logs are kept.

AI literacy: support scientists with short training on what nucleotide models can and cannot predict (GUIDE-006). Pick a pilot with one construct family and one assay, compare predictions with wet-lab results, and report misses openly. Typical concern: "the model replaces my judgement". Address it by keeping scientists in charge of constraints and the final synthesis order, and by showing uncertainty. See GUIDE-009 and GUIDE-008.

Risks, regulation & compliance

• EU AI Act: minimal risk in typical research and industrial use (no Annex III category); AI literacy (Art. 4) applies. Medical-product context is governed by sector law, not by the model as such.
• GDPR: generally not applicable to microbial and synthetic sequences; becomes relevant if training data include human genomes or patient-derived cell lines (genetic data is special category data, Art. 9).
• Biosecurity (dual use): generative models can in principle propose harmful sequences. Evo 2 developers excluded eukaryotic viral genomes from training data. Screen all orders with sequence-of-concern screening, use synthesis providers that follow IGSC or IBBIS screening standards, keep access logs; do not publish or share design workflows for pathogenic agents.
• Biotech regulation: contained use of GMMs (Directive 2009/41/EC), deliberate release of GMOs (Directive 2001/18/EC), gene-therapy and mRNA products (EMA ATMP / medicinal-product rules, GMP); EU rules on new genomic techniques for plants: see SCI-020.
• IP: nucleotide models may reproduce training sequences; run freedom-to-operate checks on generated promoters and genes; record AI use for inventorship questions.
• Mitigations: validated assays, held-out test sets, version-controlled model and data lineage, biosafety sign-off.

Getting started

1. Pick one well-defined target (for example UTR pair for one mRNA, or a promoter for one host) and define a measurable assay.
2. Establish the baseline: codon-usage-table design plus a standard UTR; collect 200-2000 variants if no data exist.
3. Score the baseline variants with Evo 2 zero-shot and a simple supervised model; check correlation with measured values before trusting either.
4. Run one active-learning round; compare hit rate with random selection.
5. Add biosecurity screening and IP checks to the ordering workflow before scaling.

Sources

1. Arc Institute - Evo 2 announcement - https://arcinstitute.org/news/blog/evo2
2. Evo 2 code and weights - https://github.com/ArcInstitute/evo2
3. Brixi et al. - Genome modelling and design across all domains of life with Evo 2 (bioRxiv) - https://www.biorxiv.org/content/10.1101/2025.02.18.638918
4. Zhang et al. - Algorithm for optimized mRNA design improves stability and immunogenicity (LinearDesign, Nature 2023) - https://www.nature.com/articles/s41586-023-06127-z
5. Dalla-Torre et al. - Nucleotide Transformer (Nature Methods 2024) - https://www.nature.com/articles/s41592-024-02523-z
6. NVIDIA BioNeMo - https://www.nvidia.com/en-us/clara/biopharma/
7. Random Promoter DREAM Challenge (de Boer et al., Nature Biotechnology) - https://www.nature.com/articles/s41587-024-02414-w
8. International Gene Synthesis Consortium - Harmonized Screening Protocol - https://genesynthesisconsortium.org/
9. Directive 2009/41/EC on contained use of genetically modified micro-organisms - https://eur-lex.europa.eu/eli/dir/2009/41/oj
10. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj