AI Designs Viral Genomes That Don't Exist in Nature for the First Time: Breakthrough and Risk Coexist

AI Enters the Realm of Life Design: From Text Generation to Genome Creation

The boundaries of artificial intelligence applications are being continuously pushed. Following breakthroughs in protein structure prediction and drug development, AI can now design viral genomes that have never existed in nature. This advancement marks generative AI's formal transition from processing text, images, and code to directly designing life's most fundamental encoding—nucleic acid sequences.

It's worth reviewing the clear evolutionary trajectory of AI breakthroughs in life sciences. In 2020, DeepMind's AlphaFold2 solved the protein folding problem that had puzzled biologists for 50 years, predicting protein 3D structures with atomic-level accuracy. In 2024, AlphaFold3 further expanded to predict complex structures involving proteins with DNA, RNA, small molecules, and more. In drug development, Insilico Medicine used AI to complete the entire process from target discovery to candidate drug identification in 18 months, compared to the 4-5 years typically required by traditional methods. Together, these advances form AI's evolutionary path from "understanding life" to "designing life," and viral genome design is the latest milestone on this path.

According to related research, scientists have used AI models to generate complete viral genome sequences, and some of these designed sequences have been confirmed to be biologically active in laboratory experiments. This means AI is no longer just analyzing or predicting existing biological information—it can create entirely new, functional genetic material.

Technical Principles Behind AI-Designed Viral Genomes

Nucleic Acid Sequences as a "Language"

The core concept behind this research bears striking similarity to how large language models process natural language. DNA and RNA are essentially sequences composed of four bases (A, T, G, C or A, U, G, C), and can be viewed as a language with specific "grammar" and "semantics."

This analogy is not merely a superficial metaphor—it has deep mathematical foundations. Just as natural language consists of vocabulary and grammatical rules, genomes have hierarchical structure: bases form codons (triplets), codons encode amino acids, and amino acids compose proteins. More importantly, genomes contain extensive long-range dependencies—the distant interactions between promoters and coding regions, regulatory elements and target genes—which are highly analogous to contextual dependencies in natural language. The self-attention mechanism in Transformer architecture excels at capturing these long-distance correlations, which explains why large language model architectures can be transferred almost seamlessly to genome sequence modeling tasks.

By training on massive amounts of real genomic data, AI models can learn the intrinsic patterns of viral genomes—which sequence combinations can encode functional proteins, and which structures can support viral replication and assembly. On this basis, models can "generate" entirely new sequences that never appeared in the training set while maintaining biological plausibility.

The Closed Loop from Generation to Experimental Validation

Notably, this research didn't stop at computer simulation. What truly sparked widespread discussion is that some AI-designed genomes demonstrated functionality in actual experiments. This "generate—synthesize—validate" closed loop represents the most disruptive paradigm in current AI-driven synthetic biology: AI proposes hypotheses, experiments rapidly validate them, dramatically compressing the traditional bioengineering R&D cycle.

Traditional synthetic biology follows the "Design-Build-Test-Learn" (DBTL) cycle, but each iteration can take months or even years. AI's introduction fundamentally changes this paradigm's efficiency. In the "Design" phase, AI can generate thousands of candidate sequences in hours, while human experts might need weeks to design just a few schemes. In the "Test" phase, AI's predictive capabilities can significantly reduce the number of candidates requiring actual validation. Meanwhile, DNA synthesis technology costs (such as enzyme-based template-free synthesis) have dropped more than 1,000-fold over the past decade, dramatically improving the economic feasibility of the "synthesize—validate" closed loop. Currently, synthesizing a complete viral genome (thousands to tens of thousands of base pairs) costs on the order of a few thousand dollars, making high-throughput experimental validation possible.

Application Prospects for AI-Designed Viruses

From a positive perspective, this technology holds enormous application potential across multiple medical and life science fields.

Phage Therapy: A New Path to Combat Superbugs

As antibiotic resistance becomes increasingly severe, bacteriophages—which can precisely infect and kill specific pathogenic bacteria—are receiving renewed attention. AI-designed customized phages could become new weapons against superbugs, offering clinicians treatment options beyond traditional antibiotics.

The history of phage therapy dates back to the 1920s, but it was marginalized by Western medicine with the advent of the antibiotic era, continuing only in Eastern European countries like Georgia and Poland. Today, approximately 1.2 million people die annually from antibiotic-resistant bacterial infections worldwide (according to a 2022 Lancet study), and the World Health Organization warns that humanity may be entering a "post-antibiotic era." The biggest bottleneck in traditional phage therapy is that finding phages matching specific pathogens requires considerable time—screening, isolating, and identifying from natural environments typically takes weeks. AI-designed customized phages can bypass this bottleneck by directly generating targeted phage genomes based on the target bacteria's surface receptor characteristics and defense mechanisms, compressing response time from weeks to days.

An Accelerator for Gene Therapy and Vaccine Development

Many gene therapy approaches rely on engineered viral vectors to deliver therapeutic genes, and AI can design more efficient, safer vectors with lower immunogenicity. In vaccine design, AI can also help optimize antigen sequences, dramatically accelerating response speed to emerging infectious diseases.

One of the core challenges in gene therapy is how to safely and efficiently deliver therapeutic genes to target cells. Adeno-associated virus (AAV) is currently the most commonly used gene therapy vector, with several AAV-based gene therapy drugs already approved, such as Zolgensma for spinal muscular atrophy (priced at over $2 million for a single treatment). However, existing AAV vectors have limitations including high immunogenicity, limited tissue tropism, and small packaging capacity (approximately 4.7kb). AI can break through these limitations by designing entirely new capsid protein sequences, creating vector variants that natural evolution never produced, achieving more precise tissue targeting and lower immune rejection—thus driving gene therapy's expansion from rare diseases to common diseases.

An Exploration Tool for Fundamental Life Sciences

Additionally, this technology provides powerful tools for basic research, helping scientists understand the boundaries of viral evolution, explore "all possible forms of life," and expand our understanding of biological principles. By generating viral sequences that don't exist in nature and observing their behavior, researchers can systematically explore which regions of sequence space correspond to functional life forms, thereby revealing fundamental principles of how life operates.

Biosecurity Risks and Ethical Challenges That Cannot Be Ignored

However, this breakthrough also raises serious safety and ethical questions.

The Dual-Use Dilemma

Biosecurity is the primary concern. If AI can design functional viral genomes, then theoretically it could also be misused to design pathogens with enhanced transmissibility or virulence. This "dual-use" dilemma is an unavoidable challenge for all powerful biotechnologies. Lowering the technical threshold for pathogen design could mean risk spreading to a broader population.

Discussions on governing "Dual-Use Research of Concern" (DURC) can be traced back to the 2011-2012 controversy over H5N1 avian influenza "gain-of-function" experiments. At that time, research teams in the Netherlands and the United States independently achieved airborne transmission of H5N1 virus between mammals, triggering a year-long research moratorium and intense global debate. Subsequently, the U.S. established multi-layered review mechanisms including Institutional Review Boards (IRBs) and the National Science Advisory Board for Biosecurity (NSABB). However, AI-generated viral genomes present entirely new governance challenges: traditional regulatory frameworks target "research activities in laboratories," but once an AI model is released, its "research" can be conducted on any internet-connected computer, making territorial jurisdiction and institutional review extremely difficult.

AI Model Openness and Safety Guardrails

Currently, many powerful biological sequence generation models are open-source or partially open. How to establish effective "guardrail" mechanisms that prevent models from being used to generate dangerous sequences while promoting open science is a problem the industry must address urgently. Some organizations have begun exploring screening tools for DNA synthesis companies to identify and intercept potentially dangerous orders.

Global DNA synthesis industry security screening currently relies primarily on voluntary protocols established by the International Gene Synthesis Consortium (IGSC). The protocol requires synthesis companies to compare customer orders against controlled pathogen sequence databases and verify customer identity and institutional legitimacy. However, this system has obvious vulnerabilities: it is voluntary rather than mandatory, covering only about 80% of global synthesis capacity; comparison databases may not recognize AI-designed novel sequences (because these sequences have no direct homology with known dangerous sequences); furthermore, the emergence of benchtop DNA synthesizers could completely bypass centralized screening systems. In 2023, the U.S. government issued an executive order requiring federally funded research to use screened synthesis providers—an important step toward mandatory screening—but global-level coordination still has a long way to go.

Finding Balance Between Innovation and Regulation

AI designing viruses that don't exist in nature both demonstrates generative AI's remarkable potential in life sciences and once again pushes biosecurity governance into the spotlight. This is not merely a technical achievement—it's a governance issue that the research community, regulatory agencies, and society as a whole must face together.

Attitudes toward this technology are complex and cautious: on one hand, there's anticipation for revolutionary breakthroughs in medical fields like phage therapy and gene therapy; on the other, there's deep concern about the possibility of misuse. At a time when AI capabilities continue to explode, establishing matching safety frameworks and ethical norms may be more urgent than the technical progress itself. This requires building closer dialogue mechanisms between technology developers, biosecurity experts, policymakers, and the public to find a dynamic equilibrium between open innovation and risk management.