On: Artificial Intelligence used to design brand new viruses
The news today is of a new kind of alchemy, one where the crucible is a silicon lattice and the philosopher’s stone is an algorithm. They have used what they call “artificial intelligence” to design novel viruses, and succeeded sixteen times. The report is breathless, as if the mere act of creation were the marvel. But the act is not new; we have been designing compounds, alloys, and even biological agents through computation for decades. The novelty they trumpet is in the tool, not the principle.
What they have not done, and what they must do, is apply the consilience test. A hypothesis - in this case, the AI’s model of viral structure and function - was trained on a corpus of known viral genomes. It then produced new sequences that, when synthesized, yielded functional viruses. This is a curve fit of the highest order. It explains, or rather replicates, the data on which it was trained. The true test is this: did the model predict any property of these new entities that was not implicit in its training? Did it foresee, for instance, a novel mode of cellular entry, or a stability under conditions never before recorded for its parent strains? If the success is merely that the viruses infect - a property defined by the training set - then they have demonstrated a powerful pattern-matching engine, not a discovery. They have automated induction of the lowest rung.
They speak of “design,” but I question the term. Design implies an understanding of first principles, a causal map from sequence to function. If their machine is a black box that correlates patterns, it is not designing; it is guessing with a high rate of success. This is not a semantic quibble. To name it correctly is to see the peril and the promise clearly. The peril is that we will mistake correlation for mechanism and unleash entities whose deeper behaviors are opaque. The promise, which they have missed entirely, lies in the next step: to take these sixteen successes and ask what unanticipated, cross-domain properties they share. Do they, perhaps, share an unexpected affinity for a particular class of receptor, one common to tissues the original viruses never touched? That would be a thread of consilience - a prediction reaching into unmapped territory.
Until such a test is passed, this is not a revolution in biology. It is a very expensive, and potentially dangerous, proof that machines are excellent students. But a student who can only recite the textbook has not yet learned to think.