The review places this shift in historical context. Traditional hybridization, the labor-intensive trial and error of saving seeds from superior plants, constitutes Breeding 1.0. Over the past century, the introduction of molecular markers and genomic selection produced Breeding 2.0 and 3.0, dramatically improving the precision with which useful genetic variants could be tracked and accumulated.

More recently, big-data omics technologies defined Breeding 4.0, expanding the catalog of known crop genes through genome-scale measurement. Yet even with these advances, conventional breeding cycles still span a grueling ten to fifteen years from initial cross to commercial variety. Breeders also continue to struggle with a core difficulty: many agronomically decisive traits are governed by vast networks of genes whose effects intertwine with unpredictable climate interactions, so improving yield, stress tolerance, and quality simultaneously remains an unsolved multi-objective problem.

The authors draw a sharp distinction between the artificial intelligence that has so far dominated agricultural applications and the models now driving the new paradigm. Earlier tools were overwhelmingly discriminative, meaning they evaluated, ranked, or classified genetic variants that already existed in germplasm collections and breeding populations. Such systems resemble a breeder walking a field and selecting the most promising plants from what is already growing there; they can identify the best available candidates but cannot propose anything beyond the existing pool.

Generative artificial intelligence operates differently. By learning the statistical patterns embedded in vast biological datasets, these models can propose novel DNA sequences, proteins, and regulatory elements tailored to specified biological functions, effectively expanding the search space beyond the boundaries of natural variation. Four foundational generative architectures anchor the technical survey.

Variational autoencoders compress complex biological sequences into continuous latent spaces, an encoding that allows researchers to fine-tune gene expression levels in a graded manner, which the authors liken to adjusting a smooth volume knob rather than flipping a binary switch. Generative adversarial networks deploy competing neural networks, in which a generator produces candidate sequences and a discriminator judges their authenticity, and this adversarial training yields sharp, diverse DNA sequences with realistic biological properties. Both approaches give breeders continuous, navigable representations of sequence space where natural germplasm offers only discrete, sampled points.

Diffusion models constitute a third architecture, reconstructing target molecules from random noise through a step-by-step denoising process. The review highlights their unprecedented precision in crafting three-dimensional protein structures and in fine-tuning gene promoters, tasks where the gradual refinement from noise to structure mirrors how a sculptor works a form out of raw material. Large language models form the fourth pillar, and biological variants of them, including genome language models, protein language models, and scientific text models, treat nucleotides as letters, codons as words, and genes as sentences.

Trained on evolutionary data at scale, these models capture long-range evolutionary grammar, the deep constraints that determine which sequence changes a genome can tolerate, and can translate broad agronomic goals into concrete molecular designs. Having established the model families, the review surveys 67 representative generative tools and identifies six critical application areas for crop molecular design. The first is the intelligent design of gene regulatory elements: inventing synthetic promoters and enhancers tailored to activate only under specific environmental triggers, so that a stress-response program switches on when drought or heat arrives without imposing unwanted yield penalties during normal growth.

The second is functional protein innovation, where generative models design customized enzymes intended to enhance photosynthesis and synthetic binders engineered to disarm pathogen effectors, the virulence molecules that pathogens deploy against plant defenses. The remaining frontiers extend generation beyond sequence alone. Small-molecule generation aims to discover eco-friendly plant growth regulators and targeted crop protection chemicals, replacing broad-spectrum agrochemicals with rationally designed compounds.

Gene-editing tool optimization applies protein generation to engineer novel CRISPR-Cas nucleases with ultra-high fidelity and to design multiplex guide RNAs adapted to the large, complex genomes of major crops, where off-target effects and polyploidy complicate editing. Single-cell multi-omics simulation points toward AI virtual cells, abbreviated AIVCs, that simulate cellular responses in silico, such as modeling how root cells behave under salt stress before any plant is grown. Finally, cross-domain knowledge discovery leverages scientific AI to mine decades of agronomic literature and build interconnected gene-trait knowledge graphs, converting scattered publications into a queryable map of relationships.

To organize these rapid developments into actionable breeding practice, the authors propose the DESIGN framework, which decomposes generative breeding into six synergistic capabilities. Discover mines functional patterns from deep genomic and structural spaces. Engineer rationally refines molecular parts under biophysical and cellular constraints.

Simulate evaluates candidate performance across virtual environments before candidates head to the field. Integrate fuses multi-omics, environmental, and literature data into a unified biological semantic map. Generate produces functional genetic elements beyond the boundaries of natural variation.

Navigate guides multi-objective optimization through high-dimensional search spaces to find optimal trade-offs, for example balancing drought resistance against grain yield, a compromise that conventional breeding pursues only through costly, slow iteration. Crucially, the six capabilities do not operate as a linear pipeline. Instead, they are collectively integrated into a closed loop the authors abbreviate LDBT, standing for Learn, Design, Build, and Test.

Data generated at each experimental stage are fed back into the model to iteratively optimize design schemes, fostering continuous synergy between computational modeling and experimental validation. The review illustrates the workflow with a schematic drought-responsive promoter engineering project that integrates multi-omics data, conditional promoter generation, experimental validation, and active learning feedback to support breeding applications. In this example, each round of promoter designs is built and tested in the laboratory, the measured outcomes retrain the generative model, and the refined model proposes improved candidates, demonstrating concretely how the components of the DESIGN framework work together in practice.

The authors are candid about the gap between code and crop field. Translating generative AI into real-world agronomic gains faces substantial hurdles: plant data remain fragmented across institutions and formats, experimental throughput in greenhouse trials limits how many designs can be validated per cycle, and model interpretability remains a concern when breeders must commit resources to AI-proposed sequences. The review also stresses that AI will not replace plant breeders.

Future success, the authors argue, hinges on close human-AI collaboration, with breeders guiding strategic goals and exercising agronomic judgment while AI navigates molecular complexity at scales no human can search. Whether Breeding 5.0 delivers on its promise will depend on closing that loop: richer shared datasets, higher-throughput validation, and models whose reasoning breeders can inspect and trust. If the LDBT cycle matures, the review suggests, the decade-long variety development timeline that has defined modern agriculture could compress into a design process in which the next variety is computed, tested, and refined long before it is ever sown.

Subject of Research: Generative AI applications in crop molecular design and breeding

Article Title: From natural selection to de novo design: Generative AI ushers in "Breeding 5.0" for crops

Article References: From natural selection to de novo design: Generative AI ushers in "Breeding 5.0" for crops. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: generative AI, crop breeding, Breeding 5.0, DESIGN framework, diffusion models, large language models, CRISPR-Cas, gene regulatory elements, AI virtual cells, protein engineering, Science China Life Sciences, natural