What changes when AI moves from reading viral genomes to designing them?

Why in the News

Researchers at Stanford University and the Arc Institute used Artificial Intelligence (AI) to design complete genomes of bacteriophages, viruses that infect bacteria. Of 285 AI-generated designs physically synthesised and tested in the laboratory, 16 produced functioning phages, and some overcame bacterial resistance that had defeated the original virus. Humans have been synthesising viral genomes and deliberately modifying viruses for decades, so what is new is not the physical manufacture of a virus. AI has entered the design stage of biology, deciding what the genome should be rather than executing a design a human specified. The tension is that the same capability that could transform antimicrobial resistance research, vaccines and therapeutics could also accelerate harmful biological engineering.

What is a genome language model?

1. What it is: A genome language model is a machine learning system trained on genetic sequence data rather than on text, and the two used in this experiment were Evo 1 and Evo 2.
2. How it works: The principle resembles a large language model, except that instead of learning patterns in words, it learns patterns in DNA.
3. What it reads: It studies the genetic alphabet of A, C, G and T across vast numbers of genomes, and then generates new genetic sequences from the patterns it has learned.
4. How it was specialised: For this experiment the models were further trained on thousands of bacteriophage genomes related to ΦX174, so the sequences they generated stayed within a known biological family.

What did the Stanford-Arc experiment actually do?

1. The design step was handed over: The scientists already knew the ΦX174 genome and already knew how to synthesise viral DNA and recover functioning phages. What changed was who, or what, proposed the genome.
2. The output was constrained, not open-ended: AI did not invent a completely unrelated virus from nothing. It generated previously unseen ΦX174-like whole genomes within a known biological framework.
3. The build step was conventional: Scientists selected some of these sequences, physically manufactured the DNA and introduced it into E. coli. Where the genetic instructions were biologically coherent, the bacterial machinery produced new phage particles.
4. The yield: Of the 285 designs tested, 16 succeeded in producing functioning phages.
5. Some designs beat the natural virus: Combinations of AI-designed phages overcame resistance in E. coli strains against which the original ΦX174 failed.
6. The most significant result was combinatorial: An AI-designed phage successfully combined a viral protein with other genetic changes in a way conventional engineering had struggled to achieve, which suggests the system can identify multiple genetic changes that work together across an entire genome.

How did biology get from reading genomes to writing them?

1. Phages are old and abundant: Bacteriophages, literally “bacteria eaters”, have been known for more than a century and are among the most abundant biological entities in nature.
2. Reading came first: In 1977, one particularly small phage, ΦX174, became the first complete DNA genome to be sequenced.
3. Writing came next: By the early 2000s, scientists had shown that viral genetic material could be synthesised from known sequence information and used to recover functioning viruses.
4. Deliberate modification followed: The controversial influenza gain-of-function experiments of 2011-12 showed that genetic changes could modify important properties such as transmission in experimental animals.
5. The unresolved dilemma: That research highlighted a dilemma that remains open, since the same science that can improve pandemic preparedness may also create biosafety and biosecurity risks.
6. Design is the fourth step: The progression runs from reading viral genomes, to writing them, to modifying them, and now to AI helping decide what should be written.

What does this open up in medicine?

1. Phage therapy is the nearest application: Antibiotic resistance is steadily eroding conventional treatment options, and bacteriophages offer another way of killing bacteria.
2. Specificity is the limitation: A phage effective against one bacterial strain may fail against another, and bacteria can also develop resistance to phages.
3. The search model has limits: Researchers have traditionally searched nature and phage libraries for suitable candidates, or modified existing viruses, which caps the available options at what already exists.
4. The question changes: Generative biology moves medicine from asking whether the needed phage can be found to asking whether it can be designed.
5. The applications extend well beyond phages: AI can assist the design of vaccine antigens, antibodies, therapeutic proteins and the viral vectors used to deliver genetic treatments, and may eventually help optimise oncolytic viruses that selectively attack cancer cells.
6. The real shift is broader than viruses: The larger revolution is AI becoming capable of designing biological function, rather than AI making viruses.

Is the simplicity of the target a safeguard, or is the risk the acceleration?

1. The reassuring reading: ΦX174 is an exceptionally simple bacteriophage, while dangerous human viruses are vastly more complicated.
2. Human pathogens are harder targets: They must negotiate receptor binding, host range, tissue tropism, replication, immune escape and transmission, each of which is a separate design problem.
3. Complexity is not a defence: Human scientists already understand much about these determinants, and decades of virology, reverse genetics and gain-of-function research have linked many genetic changes to viral behaviour.
4. AI does not need to rediscover virology: Its power lies in integrating what humanity already knows, examining vastly more combinations than humans can explore manually, and accelerating the path from hypothesis to experimental design.
5. The concern is capability amplification: The relevant question is not whether an untrained individual can ask today’s chatbot to generate a pandemic virus. It is whether increasingly capable AI could make a knowledgeable and well-equipped laboratory substantially more effective at designing biological systems.
6. A low success rate is a temporary comfort: The yield reported above is low, but digital systems can generate enormous numbers of candidates, so a low success rate is reassuring only while the number of attempts remains small.

How must biosecurity change?

1. Current screening looks for resemblance: Traditional DNA-synthesis screening often asks whether an ordered sequence resembles a known pathogen or toxin.
2. Resemblance fails against generated sequences: A previously unseen sequence generated inside a known family may not resemble anything on a watchlist while still doing the same thing.
3. Screening must move to function: In the age of generative biology, screening must also consider what a sequence might actually do, not simply whether it looks dangerous.
4. Over-restriction has its own cost: Claude Fable 5 was initially deployed with strong safeguards around biology, chemistry and cybersecurity, and legitimate scientific work could sometimes trigger a fallback to a less capable model.
5. The correction points to graduated access: Those safeguards have since been refined to reduce false-positive biology fallbacks while more sensitive capabilities remain restricted, which points toward graduated, auditable access under institutional and security controls.
6. Model refusal is not a strategy: Biosecurity cannot rest entirely on what an AI model agrees or refuses to answer, so safeguards are needed throughout the chain: AI systems, DNA-synthesis providers, laboratories and institutional biosafety oversight.

Why does this matter for India?

1. Frontier AI becomes scientific infrastructure: If frontier AI becomes central to drug discovery, genomics, vaccines, protein engineering and experimental design, access to advanced AI becomes part of national scientific infrastructure.
2. Sufficiency and compulsion are different things: Smaller and specialised models will be sufficient for many tasks, but a country should choose a small model because it is sufficient, not be forced to use one because somebody else owns the frontier.
3. Restricted access compounds over time: If researchers elsewhere receive trusted access to highly capable biomedical models while Indian scientists depend on restricted public versions, the disadvantage accumulates across drug discovery, vaccines and antimicrobial resistance.
4. The investment exists but needs a scientific arm: India is already investing through the IndiaAI Mission and indigenous foundation-model programmes, and that ambition should extend to scientific and biomedical AI, secure compute and high-quality datasets.
5. Trusted access needs a framework: Legitimate researchers need a defined route to stronger capabilities, which requires an institutional trusted-access framework rather than case by case negotiation with model providers.
6. The two goals are not separable: AI sovereignty without biosecurity would be reckless, and biosecurity without AI sovereignty could leave the country scientifically dependent.

Challenges to AI-designed genomes

1. Sequence screening cannot see intent: Order screening matches against known pathogen sequences, so a generated sequence within a benign-looking family passes even where its function is hazardous. Eg. Screening protocols built around named agents on an export control list match those names, so a functionally equivalent sequence outside the list is not flagged. Fix. Require DNA-synthesis providers to run function prediction alongside sequence matching, with a reporting duty on flagged orders.
2. Automated laboratories compress the safety window: Combining generative design with robotic experimentation shortens the interval in which oversight can intervene. Eg. Future systems may compress months or years of literature review, modelling and experimental planning into much shorter cycles. Fix. Mandate institutional biosafety committee sign-off at the design stage rather than only before physical synthesis.
3. Volume defeats low success rates: A weak per-attempt success rate becomes a strong aggregate capability once attempts are cheap and unlimited. Eg. The design pool in this experiment was generated computationally, so the number of candidates was bounded by compute rather than by laboratory effort. Fix. Impose volume-based reporting thresholds on synthesis orders from a single requester within a stated period.
4. Model safeguards obstruct legitimate research: Blunt refusal policies block the research they were meant to protect, which pushes scientists toward unsupervised alternatives. Eg. Legitimate scientific queries triggered fallback to a less capable model under initial biology safeguards. Fix. Operate tiered credentials, where verified institutional researchers receive higher-capability access under audit logging.
5. Governance is nationally fragmented: Biosecurity rules stop at borders while synthesis orders and model access do not. Eg. The 2011-12 gain-of-function controversy produced divergent national moratoria rather than a common standard. Fix. Negotiate a common minimum synthesis-screening standard through the Biological Weapons Convention review process.
6. India lacks a biosecurity institution for generative biology: Existing oversight bodies were designed for genetically modified organisms and field trials, not for computational design of pathogens. Eg. The Genetic Engineering Appraisal Committee and the Review Committee on Genetic Manipulation are structured around organism release rather than sequence design. Fix. Create a statutory biosecurity review function covering generative design, synthesis orders and model access, reporting jointly to the Department of Biotechnology and the Ministry of Electronics and Information Technology.

Conclusion

The experiment does not show that AI can casually manufacture dangerous human viruses. It shows something more precise: computers are beginning to move from analysing biological information towards proposing biological designs that scientists can physically build, a capability that serves therapeutic research and harmful engineering alike. The answer is neither prohibition nor unrestricted access, but controlled acceleration, with safeguards rising as capability and risk rise. The unresolved question is no longer whether AI should be allowed to understand biology, but how to govern it once understanding biology becomes the ability to design it.

Back2Basics: IndiaAI Mission

1. What it is: The IndiaAI Mission is the national artificial intelligence programme approved in 2024 with an outlay of ₹10,371 crore, implemented by IndiaAI under the Ministry of Electronics and Information Technology.
2. Its stated vision: “Making AI in India and Making AI Work for India”, built around seven pillars covering compute, applications, datasets, foundation models, skills, startup financing and safe and trusted AI.
3. Compute pillar: It operates a national AI compute grid with over 38,000 graphics processing units, offering up to 40 per cent lower compute costs to eligible users.
4. Safety arm: The IndiaAI Safety Institute is its national trust framework, covering bias mitigation, privacy, explainability and AI governance.

Matching Previous Year Question

“[2026] Which of the following statements with regard to genetic medicine is/are correct? 1. Genetic medicines correct/compensate for the faulty genes responsible for disease. 2. Engineered viruses and lipid nanoparticles are used as carriers of the genetic medicine. 3. Genetic medicines alter the entire DNA sequence. (a) 1 only (b) 2 and 3 only (c) 1 and 2 only (d) 1, 2 and 3 ANSWER: C”