TBD

Philip Ball on Google DeepMind's Alpha Genome

Alexander Calder, Spiral with black Stripes (detail)

When the researchers at Google DeepMind produced AlphaFold, an AI algorithm that could predict the three-dimensional shape of protein molecules—a question usually resolved using the painstaking and sometimes intractable experimental method of crystallography—it was widely hailed as a revolution in biochemistry. Proteins are the molecules that coordinate most of the chemical reactions that go on inside out cells, and knowing their shape can be crucial for designing drug molecules that can bind to proteins and thus intervene in our own biochemistry. Two of the researchers involved, including DeepMind’s CEO Demis Hassabis, were awarded the 2024 Nobel prize in chemistry for AlphaFold, signalling to many that AI-based science had truly arrived.

Given this history, it’s not surprising that the announcement of DeepMind’s next big step in AI-based biology, called AlphaGenome, which was published in the journal Nature on 29 January, has drawn a lot of excitement in the media. But whereas AlphaFold was a relatively simple thing to explain in broad outline, AlphaGenome seems to have left the press a bit perplexed. Reporters sense that it is a big deal, but haven’t been sure quite how to talk about it. The confusion reveals a problem in how we think and talk about genetics: the public discourse remains far behind the science. In these times when genomics is increasingly impacting our lives, that is concerning.

The arrival of AlphaGenome also highlights pressing questions about the role of AI in grappling with the fearsome complexities of biology. Some researchers are now hinting that the details of molecular biology and genomics are so overwhelmingly complicated that understanding living things from the bottom up – figuring out how life emerges from lifeless molecules – might be beyond human comprehension. In this view, we must resign ourselves to reliance on the black box of AI, which can deliver answers to specific questions (such as “What is the physiological effect of this gene variant?”) without needing to explain its reasoning. It is important that this notion be resisted. For one thing, there may be limits to what AI can deliver in this regard: we should not succumb to the prevailing habit of overestimating the power and reliability of these algorithms. And past experience, for example in drug design, as shown us that interventions tend to be most effective when we have some mechanistic understanding of how they work. But beyond such pragmatic considerations, a decision to make AI a gnomic oracle rather than a useful tool is unnecessarily defeatist. There’s no denying the headspinning complexity of modern molecular biology, but in recent years some of its general principles have started to come into focus. Some of that complexity, at least, might simply be a mirage conjured by looking at the problem from the wrong angle.

AlphaGenome uses the AI method called deep learning (more on that shortly) to predict the biological effects of person-to-person differences in the chemical makeup of our DNA—our genomes. We know that many of the differences between us, such as our physical and psychological traits and our susceptibilities to disease, are related to these genetic differences: slight variations in the sequence of around three billion chemical “letters” that are strung together, like beads on a string, in the DNA molecules carried by each of our cells. But it is tremendously difficult to figure out what the relationship is between these variations in individual genetic sequence (genotypes) and our traits (phenotypes). AlphaGenome is able to predict some of those associations. In  this regard it is not the first system of its kind. Other AI schemes such as Evo 2, developed at Stanford and Berkeley in collaboration with tech giant Nvidia, and Oracle, created at MIT, have previously supplied similar results.

In principle such a capability means that, if a person has their genome sequenced—a procedure now so routine and relatively cheap that some countries are considering offering it for every newborn baby—it would be possible to identify their risk of developing all kinds of diseases, or vulnerabilities to drug side effects, or the drugs that would work most effectively for them, and much more.

Such a prognostic tool could be immensely valuable for health and well-being, but it would come with complex societal and ethical concerns. Already, IVF embryos can be screened to predictions the health and characteristics of the babies that would be born from them. When this is done to avoid (generally rare) genetic diseases, there is a clear argument in favour. If it is permitted for predictions of, say, intelligence or attractiveness, the case is far murkier—which is why such non-disease-related predictions are illegal in some countries.

At this stage, AlphaGenome is a long way from any of that. It is more of a tool for basic science, and in this regard it is both impressive and sure to be useful. Its focus is on predicting how relatively common “genetic variants”—segments of DNA that have been found to differ systematically between different people, such that we can say “Oh, I have that variant”—might affect, say, the chance of developing cancer linked to a particular gene. In the cases tested by the DeepMind team (using both human and mouse genetic data), the algorithm performs pretty well.

Here, though, is the conceptual stumbling block for many media reporters and other non-experts trying to make sense of this stuff. Nearly all of the regions of DNA that are identified (by comparisons of genome sequences for perhaps many thousands of individuals) as being associated with prevalence of disease, and indeed with any heritable trait, are not actually in our genes at all.

What does that mean? Well, here’s the story that often gets told, and which has informed much of the reporting in this case. Genes are stretches of DNA that hold the “code” for making proteins. Molecular machinery in our cells “read” the sequence of a gene and use that information as an instruction for putting together a specific protein, which then does some important job in the cell. This is where AlphaFold comes in: the gene sequence specifies the sequence of building blocks in the protein chain, which in turn specify what shape the chain folds up into. AlphaFold can now predict shape from sequence alone. If there’s a mutation in a gene – a “letter” of the sequence is changed, say – this can change the protein sequence and prevent it from folding quite as it should, so that it doesn’t work so well. The result can be a disease. Both cystic fibrosis and sickle-cell disease are caused by gene variants like this that are relatively widespread in the population.

But only around 2% of all our DNA actually corresponds to protein-coding genes. The Human Genome Project revealed that there are just 19,000-20,000 such genes: considerably fewer than most biologists expected. So what is the other 98% of DNA for?

According to the old story, we once thought that all this other DNA was just “junk”, a term coined in this context in the 1970s. It was accumulated over the course of evolution, for example by viruses inserting their genetic material into ours, and it did nothing useful but just cluttered up our genome the way all our old detritus of a lifetime clutters up our attic. But now (we’re still with the old story here) we have come to realise that this “noncoding” DNA is not “junk” at all but performs many vital tasks. So genetic variants amongst this “not-such-junk” matter too. In place of the dismissive “junk” label, some like to characterize this “non-gene” DNA as the “dark genome”: it is up to something important, but we don’t know what.

The problem is that this story simply isn’t true, for several reasons. For one thing, we have known since the early days of modern molecular biology in the 1960s that there are important bits of DNA (in all organisms) that don’t encode proteins. These have long been known to be involved in “regulating” genes: determining whether or not the proteins they encode are produced. For one thing, different types of cell in our different tissues – skin, muscle, neurons and so on – tend to require different suites of proteins, and so their genes are regulated in tissue-specific ways.

What has changed over the past several decades is that we  have started to appreciate just how incredible complicated gene regulation is in complex creatures like us. There are many distinct DNA regions (“genetic loci”) involved in regulation just for a single gene, some of them a long way away from the gene itself on a particular chromosome. (Our genome is divided into 23 chromosomes, of which we inherit a set from each biological parent.) These regulatory sequences interact with gene-regulating proteins (called transcription factors) and other molecules. Regulation is also affected by so-called epigenetic modifications: chemical groups that get attached to DNA and to the proteins that help to package it in chromosomes. And it depends on the three-dimensional shape of the DNA – for example, which other regions are nearby, and whether it is clumped up compactly or is more open and loose so that the gene-reading molecules can get to it.

And one more thing that has become clear over the past 30 years or so: not all genes encode proteins. Some are used to make RNA molecules that also carry out important jobs in the cell. RNA used to be regarded just as an intermediary in protein synthesis: the DNA sequence was copied into RNA, which was then read to make the protein. But some “noncoding” RNAs are an end in themselves. No one knowns how many there are: there may be between 5,000 and many tens of thousands of them with genuine biological functions, and most of those functions seem to be associated with gene regulation.

Understanding this complexity of gene regulation has become one of the main goals of molecular biology in recent decades. That’s why it is misleading to speak of a “dark genome”. There’s lots we still don’t understand about the roles of noncoding DNA, but we do know an awful lot.

Incidentally, arguments still rage about just now much of our noncoding DNA is truly “functional” and how much is genuinely “junk”. It’s a slightly sterile debate about numbers that misses the more interesting points: not only is there considerably more functional noncoding DNA than we thought at the start of the Human Genome Project, but the distinction between “functional” and “non-functional” DNA has become blurred in any case. That’s not least because DNA that is not being “used” by cells at one stage in evolution might find uses later, while functional DNA at one time can decay into “dead” DNA at a later time.

AlphaGenome, then, represents an attempt to sidestep the difficult job of understanding what all the noncoding DNA is doing to regulate genes, and simple uses deep learning techniques to identify statistical correlations between genetic variants and biological effects. As with all deep-learning AI, the algorithm is “trained” on cases where these links are known, until it can infer correlations that enable it to make predictions about sequences and variants it hasn’t encountered before. In just the same way, AlphaFold avoided the difficult question of howa protein sequence determines its folding process, and simply sought correlations between that sequence and a protein’s shape, using data from proteins whose shapes had already been determined experimentally. So all the complexities of how transcription factors bind to regulatory DNA, how epigenetic marks alter gene activity, how the DNA is itself folded up, and so on, are subsumed into the algorithm’s number-crunching.

To my mind, the slightly bewildered tropes about junk DNA and dark genomes are a reflection of how all this complexity regarding our genome has been largely ignored in popular discourse on genomics. DNA is still typically regarded as code for making proteins—or worse (and this outdated cliché was much in evidence in the reporting), as a “blueprint” for making us. Marginalia’s New Biology project is aiming to bring together experts from diverse fields in biology to seek better narratives that capture and convey our new understanding.

One reason why the ongoing reliance on old and outmoded narratives is concerning us that it can lead to unrealistic expectations about what AlphaGenome will do: in particular, that it will eventually predict the organism—all we can expect of the individual person—from its genomic sequence. A similarly obsolete view of proteins—that the “protein-folding problem was the key roadblock to drug discovery—prompted some naïve reporting about how AlphaFold would revolutionize drug discovery. (It hasn’t and won’t, though it will be a useful tool among others.)

The truth is that it is precisely the understanding that has emerged in the past two or three decades about genomics and gene regulation that has undermined the simplistic “blueprint” idea of the genome. All the complexity of noncoding DNA and RNA, of genome architecture, of epigenetics, of how (as the AlphaGenome team explore) genetic variants can influence exactly which protein sequence actually gets made from a given coding gene, and of how all of this depends on context—on tissue type, on environmental signals, on past history, and so on—means that the genome might be best seen not as an “instruction manual” but in the way Nobel laureate geneticist Barbara McClintock described it in 1983: “ as a highly sensitive organ of the cell that monitors genomic activities and corrects common errors, senses unusual and unexpected events, and responds to them.”

All of this is reflected in the admirably clear discussion by the DeepMind team of the limitations of AlphaGenome. In short, they have so far, and understandably, gone for the low-hanging fruit: the predictions that are relatively straightforward. For one thing, they have considered only how genetic variants affect protein-coding genes. They say “We have not benchmarked the model on personal genome prediction, which is a known weakness of models in this space”—in other words, it’s one thing to predict what the average effects of a variant might be, quite another to say what it will be in a given individual. (Often the effect of a genetic mutation will depend on the “genetic context”: what all the other variants are. This is in part what makes it hard to predict whether a person with a known disease-linked variant will develop the disease, or how badly.) Most importantly, it is far from clear how AlphaGenome will fare for predicting complex traits that typically involved many (perhaps hundreds of) genes, not least because, as the researchers say, “these phenotypes involve broader biological processes (including gene function, development and environmental factors).” Let me put that more simply: such predictions will inevitably be of limited value, because the relevant information is not all in the genome sequence.

AlphaGenome is sure be a valuable tool for research, as is AlphaFold. Whether it will lead to clinical advances remains to be seen, and its applicability will be limited by its very nature. But to fully assess developments like this, we need to have a better public narrative about what genomes are and what they do. What’s more, it would be foolish to suppose that a “black box” AI like this can save us from the daunting task of understanding how the genome works at a fundamental level. But who would want to forego that anyway, given the intellectual richness and even the beauty that, for all its messiness, modern molecular biology is revealing?

There’s a broader issue at stake here too that pertains to the place of AI in scientific research. Too often now, we hear claims of how AI will replace human scientists and, in doing so, will solve our most pressing problems: curing all disease, solving climate change, figuring out how to conduct nuclear fusion. Such claims are never supported by experts in those fields—and that’s not because they want to convince us that they are indispensable after all. Rather, it is generally because they know that the bottlenecks are not ones AI can address. Knowing protein structures is not the fundamental obstacle in drug discovery. Knowing how to make better materials for green technologies is not the key stumbling block for tackling climate change. Combatting disease is not merely or even primarily about pharmaceutical interventions. (As oncologist Siddhartha Mukherjee remarked recently, “Cancer’s end is not conceivable.”) And science itself is not simply about producing theories or models that churn out predictive numbers. Machines can be immensely useful tools, but there are many things that we can only do for ourselves, and we cannot afford to lose the motivation to do so. 

---

Philip Ball is a scientist, writer, and a former editor at the journal Nature. He has won numerous awards and has published more than twenty-five books, most recently How Life Works: A User’s Guide to the New Biology; The Book of Minds: How to Understand Ourselves and Other Beings, From Animals to Aliens; and The Modern Myths: Adventures in the Machinery of the Popular Imagination. He writes on science for many magazines and journals internationally and is the Marginalia Review of Books' Editor for Science. Follow @philipcball.bsky.social