Designing Genomes from Text Prompts: A Constraint-First Framework for AI-Driven Sequence Design

A new perspective published in PLOS Computational Biology proposes a structured framework for turning natural-language prompts into biologically feasible genome-scale designs, spanning bacteria, archaea, eukaryotes, and viruses. The work, authored by Adhityo Wicaksono and Arli Aditya Parikesit, addresses a growing gap in the field of artificial intelligence-driven biological design: while modern generative models can produce DNA sequences of remarkable length and apparent coherence, the mere act of generation does not guarantee that the resulting sequences can function as living or replicating systems. The authors argue that the field needs an explicit, machine-checkable layer of constraints between what a user asks for and what a model delivers.

Genome-scale sequence generation has advanced rapidly. Models trained on vast repositories of genomic data can now synthesize chromosomes-length DNA strings, and a small number of experimentally viable genome-scale designs have begun to appear in the literature. Yet the authors emphasize that biological feasibility is not a byproduct of statistical plausibility. A generated sequence may resemble natural DNA in composition and structure while failing to encode the regulatory logic, metabolic completeness, or lifecycle coherence required for actual function. The paper positions this problem as one of specification rather than generation: the challenge is to define, in formal terms, what makes a designed genome viable, heritable, and safe.

Central to the proposal is a two-part architecture. The first component, termed a Constraint Compiler, translates natural-language prompts into machine-checkable specifications. These specifications capture the scope of the requested design, the biological context in which it is meant to operate, the objectives it must fulfill, the assumptions embedded in the request, and any governance requirements that apply. By converting human intent into formal, verifiable statements, the compiler creates a contract between the user and the design system, one that can be evaluated programmatically rather than left to interpretation.

The second component is a modular Validator Stack organized into six layers, designated C1 through C6. Each layer addresses a distinct dimension of biological plausibility. C1 evaluates sequence integrity, checking for basic structural soundness. C2 assesses gene and regulatory plausibility, asking whether predicted coding regions and control elements are internally consistent. C3 examines genome completeness, ensuring that essential functions are represented. C4 performs comparative coherence, measuring how the design relates to known biological sequences and systems. C5 evaluates systems-level feasibility, integrating the preceding checks into a holistic assessment of whether the design could function as a coordinated whole. C6 addresses safety and governance, screening for risks and ensuring compliance with applicable norms.

A distinctive feature of the framework is its stance on the relationship between constraints and natural sequence space. The authors argue that constraints should not require designs to remain within the boundaries of sequences found in nature. Instead, constraints should define testable boundaries for function, inheritance, lifecycle coherence, and safe use, including for designs that are new to nature. This position opens the door to genuinely novel biological systems while maintaining a rigorous standard of evaluation. The emphasis is on demonstrable properties rather than similarity to existing organisms.

To probe the robustness of prompt-to-genome systems, the authors propose the use of extreme prompts as stress tests. These deliberately challenging inputs are designed to expose hidden assumptions, conflicting objectives, and refusal points within a design pipeline. Extreme prompts are complemented by adversarial testing, in which the system is probed with inputs crafted to reveal weaknesses, and by formal risk assessment procedures. Together, these practices aim to surface failure modes before they manifest in real designs, rather than after.

The paper also outlines a phased roadmap for implementing the framework. Rather than proposing an all-at-once deployment, the authors describe a progression in which increasingly ambitious design tasks are attempted as validation capabilities mature. Community benchmarks are proposed to measure specification accuracy, biological feasibility, robustness, and governance performance. These benchmarks would allow different research groups to evaluate their systems against shared standards, fostering comparability and cumulative progress across the field.

The scope of the framework is notably broad, encompassing bacteria, archaea, eukaryotes, and viruses. This breadth reflects the authors’ view that the principles of constraint-first design apply across the spectrum of biological entities, from minimal viral genomes to complex eukaryotic chromosomes. It also acknowledges that different domains of life present different validation challenges: viral genomes, for instance, must satisfy constraints related to host interaction and packaging, while eukaryotic genomes involve chromatin organization and developmental regulation.

The ultimate goal articulated in the paper is to move prompt-to-genome systems from impressive generation toward reliable, auditable, and accountable biological design. The authors contend that without a formal constraint layer, the field risks producing sequences that are statistically compelling but biologically hollow, and that without governance integration, it risks advancing faster than its safety infrastructure. By embedding specifications, validation, stress testing, and benchmarks into a single coherent framework, the proposal offers a template for how AI-driven genome design might mature into a discipline characterized by rigor and responsibility.

As genome-scale AI design capabilities continue to expand, the constraint-first perspective presented here provides a conceptual foundation for ensuring that ambition is matched by verification. The framework does not prescribe specific model architectures or biological endpoints; rather, it defines the evaluative scaffolding within which such systems can be developed, tested, and governed. In doing so, it contributes to an emerging conversation about how computational biology can harness the power of generative models while maintaining the standards of evidence and accountability that the life sciences demand.

Subject of Research: A constraint-first framework for AI-driven genome-scale sequence design across bacteria, archaea, eukaryotes, and viruses

Article Title: From prompts to genomes: A constraint-first framework for AI-driven sequence design across life and viruses

Article References: Wicaksono, A., & Parikesit, A. A. (2026). From prompts to genomes: A constraint-first framework for AI-driven sequence design across life and viruses. PLOS Computational Biology, 22(10), e1014865. https://doi.org/10.1371/journal.pcbi.1014865

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014865

Keywords: genome design, artificial intelligence, constraint compiler, validator stack, synthetic biology, sequence generation, biological feasibility, governance, adversarial testing, viruses, prompt-to-genome, computational biology