There are no life lessons to be learned in AI’s Chinese Room

‘AI learns the language of life.’ That’s how an algorithm called Evo, developed at Stanford University in the US, was introduced recently on the cover of Science.¹ A perspective article on the work was a little more specific, being titled ‘Learning the language of DNA’.² DNA being, after all, the secret of life, in the apocryphal words of Francis Crick (which were in fact invented by James Watson to spice up his 1968 book The Double Helix).

It sounds, then, as though the work of biology has been completed (as we’re told everything else is going to be) by artificial intelligence. There’s the small hitch that Evo can’t actually explain the ‘language of life’ to us, because it remains the archetypal black box, able only to accept inputs and churn out answers. But maybe we shouldn’t worry, as long as the machine doesn’t stop?

Linguistic metaphors for how both life and genomes (which, it should not be necessary to point out, are not synonymous) work are common, and in many ways rather apt. But like all metaphors, they have pitfalls – not least, that we might forget they are metaphors at all. And even if there were a language of genomic sequences, today’s AI cannot learn it. This is a perfect instantiation of the Chinese Room thought experiment of philosopher John Searle, proposed to challenge computational models of consciousness. Here a person sits in a room and is fed questions written in Chinese, which they do not understand – but to which they can look up and return perfect answers thanks to a handbook that describes which outputs are appropriate for given sets of input characters. The person is not learning Chinese, nor understanding anything.

The framing aside, Evo a deeply impressive tool. Trained on data from 2.7 million genomes of single-celled prokaryotic organisms and their viruses (phage), it can accurately predict the roles of different DNA sequences for such organisms, including the effects of even single-nucleotide mutations on organismal fitness. And it can work in generative mode to design de novo sequences that could be used to genetically engineer altered and new functions. Achieving something comparable for eukaryotes, and especially for complex metazoans like us, will surely be harder, not least because the very ‘logic of life’ (especially gene regulation) is different for us in some respects. All the same, that doesn’t seem too much to hope for.

This facility could be very useful, but by itself it is all but silent about what is actually going on at the molecular level. Like most AI – and like Searle’s Chinese Room – it’s a tool for prediction, not understanding. That’s why it seems worth pushing back a little on the rhetoric surrounding developments like this – because not only is understanding surely what science is all about but because often only true understanding will lead to effective outcomes. In times when some researchers are suggesting that biology is just too complex for human minds and should be left to AI, it’s a point worth making.

That defeatist capitulation to the machine is challenged anyway by research that unpicks the detailed and often remarkable mechanics of what biomolecules are up to. Take two recent papers on mechanisms of gene regulation. One reason it’s so complicated in humans is that it tends to involve hosts of molecular actors, including teams of transcription-factor proteins along with segments of DNA that help trigger transcription (promoters) and can activate or suppress it (enhancers and silencers). Weirdly, the latter are often ‘distal regulatory elements’ (DREs), far from the site of the gene(s) they regulate, and they seem to be brought into proximity on loops of chromatin.

Angelika Feldmann of the German Cancer Research Center in Heidelberg, Germany, and colleagues have shown that at least some DREs don’t simply zoom in like an on-switch for a promoter, but work through a dynamic dance in which changes in the topology of the chromatin allow the DRE to make repeated, transient interactions, as if giving the promoter a series of little nudges to wake it up.³

Meanwhile, Alexander Stark of the Vienna BioCenter and colleagues have found hundreds of new silencers in fruit flies that seem important for governing cell fate – what type a cell will become – and which have some rather surprising properties.⁴ First, they work alone, unlike the enhancers, which do stuff in combination with a host of other molecules. Second, these silencers repress expression locally as well as distally. And they haven’t previously been recognised because they seem to be in bits of chromatin that are simply inaccessible and which were therefore assumed to be irrelevant.

All this shows that while AI is useful, it’s no substitute for painstaking benchtop experiments to figure out what is really going on at the molecular scale. The two studies also reveal how much we still don’t know about the ‘language of life’ – and which AI won’t tell us. And more: it reminds us to be amazed, thrilled and sometimes baffled by life’s mechanisms, rather than leaving them inside a black box.