When a Genome Model Produces a Working Virus

For much of the recent history of artificial intelligence in biology, the dominant question has been one of prediction. Given a protein sequence, can a model infer its structure? Given a genome, can it identify functional regions, predict a phenotype or classify an organism? Given sufficiently large collections of biological observations, can computational methods recover patterns that would otherwise remain difficult to see?

An emerging class of experiments asks a subtly different question. Rather than asking a model to interpret an existing biological object, researchers are asking whether it can propose a new one.

Recent work on the generative design of bacteriophage genomes provides an unusually clear example of this transition. Researchers led by Samuel H. King and colleagues used genome-scale generative models to produce candidate genomes related to the small bacteriophage ΦX174. Candidate sequences were synthesised and experimentally tested, resulting in 16 viable bacteriophages capable of infecting Escherichia coli. The work was initially reported as a preprint and has attracted renewed attention as an indication of how rapidly generative approaches are moving from biological sequence generation towards experimentally testable biological design. [1]

The result is interesting not because artificial intelligence has somehow learned to manufacture life independently, as some of the more excitable descriptions of the work might imply, but because it exposes a more consequential scientific question: what does a computational model need to represent about a biological system in order to produce something that works?

From sequence to system

DNA can readily be represented as a sequence of characters. A functioning bacteriophage cannot.

Even a comparatively small viral genome encodes an interconnected biological system whose components must collectively permit genome replication, expression, assembly, host interaction and propagation. A sequence can therefore be syntactically plausible while being biologically useless. The difference between the two is precisely where this experiment becomes interesting.

Genome language models extend an approach already familiar from protein language modelling. Rather than treating nucleotides merely as independent observations, they learn statistical relationships across very large collections of biological sequences. Earlier work has demonstrated that long-context models trained on bacteriophage genomes can capture information useful for tasks including genome generation and functional prediction. [2]

The more recent phage-design work takes the proposition further. A generated sequence is no longer evaluated only according to a computational metric. It is synthesised, introduced into a biological context and subjected to a rather unforgiving test: does the proposed genome give rise to a viable phage?

Most candidates did not.

That fact is at least as informative as the successful examples. Generating a sequence that resembles biology is not equivalent to generating a biological system capable of functioning. Of the hundreds of candidate genomes taken forward for experimental evaluation, only a minority produced viable phages. [1]

The distinction matters because biological design ultimately cannot be reduced to producing sequences that are statistically convincing. A useful design system must increasingly contend with function, interaction and constraint.

Prediction and design are different problems

The vocabulary surrounding generative biology can obscure this distinction.

A predictive model is generally evaluated by asking whether it correctly infers some property of an existing system. A design model faces an additional burden. Its output must enter the physical world and behave sufficiently like the intended biological object.

This introduces layers of constraint that may be absent from the optimisation objective itself. Genes must function in relation to other genes. Regulatory regions must operate in an appropriate context. Protein products must interact. Structures must assemble. The resulting system must encounter the host environment and remain biologically viable.

In that sense, successful generation provides indirect evidence that the model has captured at least some of the constraints embedded within naturally occurring genomes.

It does not necessarily follow that the model understands those constraints in anything resembling a mechanistic scientific sense.

That distinction is important.

A model can learn that certain sequence relationships are associated with viable biological systems without possessing an explicit representation of why those relationships matter. Indeed, subsequent analysis of generative genome design has suggested that much of the apparent efficiency of current approaches may arise from effective exploration relatively close to known evolutionary sequence space rather than unconstrained invention of radically novel biology. [3]

This is not a criticism of the achievement. It is a useful description of where the technology presently appears to sit.

Generative models may already be becoming useful biological optimisers before they become comprehensive biological reasoners.

Why bacteriophages matter

Bacteriophages are particularly interesting objects on which to explore these questions.

They are biologically sophisticated enough to require coordinated genome-scale function, yet sufficiently compact to make whole-genome experimentation tractable. Their dependence upon bacterial hosts also creates an immediate connection to one of the most persistent problems in infectious disease: antimicrobial resistance.

Interest in phage therapy substantially predates modern artificial intelligence. Bacteriophages can kill bacteria with considerable specificity, and engineered phages have long been investigated as possible therapeutic agents, delivery systems and tools for manipulating bacterial populations. Advances in genome engineering, including CRISPR-based approaches, have progressively increased researchers’ ability to alter host range, introduce functional payloads and investigate phage biology. [4,5]

Artificial intelligence potentially changes the search problem.

Instead of beginning exclusively with an existing phage and modifying it experimentally, computational systems may increasingly help identify candidate architectures or sequences before laboratory construction. Recent reviews of the field already envisage AI-assisted host prediction, phage selection and engineering becoming part of the development of future synthetic phage therapeutics. [6]

The attraction is obvious. Bacterial diversity is enormous, resistance can emerge rapidly, and a naturally isolated phage will not necessarily possess the combination of host range, stability, safety and therapeutic properties required for a particular application.

A computational system capable of navigating even part of that design space could therefore become useful.

But the experiment also illustrates why biological design cannot end with generation.

The laboratory remains part of the computation

One of the most important aspects of the work is also the least futuristic: the generated genomes had to be built and tested.

Biology supplied the final validation layer.

This relationship between computation and experiment is likely to remain central to generative biology. Computational models can narrow a search space, identify candidate designs or discover relationships that would be difficult to enumerate manually. Experimental biology then exposes those proposals to constraints that the model may incompletely represent.

The unsuccessful candidates are therefore not simply failures. They are information.

In an iterative design system, experimental results can reveal which computationally plausible proposals cease to function when confronted with physical biology. Those observations can, in principle, inform subsequent modelling and design.

The resulting process looks less like an autonomous artificial intelligence designing organisms and more like a progressively tighter design–build–test–learn cycle, in which computation changes the scale and character of what can be proposed.

This is already a familiar principle within synthetic biology. Generative models potentially make the design stage considerably more expansive.

Representation still matters

There is, however, a deeper problem beneath generation.

A genome language model principally encounters biology through sequence. Biological researchers encounter the same system through many additional forms of knowledge: molecular structure, host phenotype, regulatory mechanism, evolutionary history, experimental observation, environmental context and causal explanation.

These representations are not interchangeable.

If the objective is simply to generate viable sequences, statistical representation may sometimes be sufficient. If the objective becomes more demanding — for example, to design a biological intervention with specified behaviour, explain why it should work, compare alternatives, identify failure modes and retain the evidence supporting each design decision — then richer forms of biological representation become increasingly important.

This is one reason Noviota is interested in the progression from representation to modelling, reasoning and design.

Generative models demonstrate that extraordinary amounts of biological structure can be learned implicitly from sequence. The complementary challenge is to determine how biological knowledge can be represented explicitly enough that computational systems can reason over mechanisms, relationships, evidence and constraints rather than merely reproduce their statistical signatures.

These are not competing approaches. The more interesting future may lie in their combination.

From generated genomes to designed biological systems

It is tempting to describe the production of viable AI-generated phages as the arrival of artificial intelligence capable of designing life. That conclusion is premature.

What the experiment demonstrates is narrower and, scientifically, perhaps more interesting.

A computational model trained on biological sequence can generate complete genome-scale proposals, some of which survive synthesis and experimental selection sufficiently well to produce functional biological entities.

That moves generative biology across an important boundary.

The output is no longer merely a prediction about biology. It becomes a proposal for biology.

The next questions are therefore likely to become progressively harder. Can generated systems be designed for specified functions rather than general viability? Can their behaviour be anticipated across different biological contexts? Can computational systems explain why one design succeeds while another fails? Can sequence-based models be combined with explicit mechanistic representations? Can the provenance of design decisions, model outputs and experimental evidence be retained as biological designs evolve?

And, inevitably, how should increasingly capable biological design systems be governed?

The answers will require more than larger models.

They will require better representations of biology, more sophisticated experimental validation, clearer connections between computational inference and biological mechanism, and methods for preserving the evidence and assumptions that sit behind a proposed design.

The recent bacteriophage experiments should therefore be understood neither as evidence that biological design has been solved nor as a curiosity about artificial intelligence generating viruses.

They are an early indication that the boundary between computational description and biological construction is beginning to move.

For computational biology, that is a development worth watching carefully.

A related perspective

The movement from computational proposal to physical biological object also raises questions about how complex biological designs are specified, compared and reproduced. Our colleagues at Molecular Precision examine this problem from the perspective of structured biological design and nanomedicine, including the role of formal representation in making increasingly complex interventions inspectable and reproducible.