Yes, DeepMind crunches the numbers – but is it really a magic bullet? | John Naughton

The most important development of the week was not Facebook losing its appeal against the EUR2.4bn European Commission fine for abusing its search monopoly to the disadvantage of its competitors to its shopping site. DeepMind, a London-based subsidiary of Google (or its holding company, Alphabet), was moving into the pharmaceutical industry via a new company called Isomorphic Labs. Its goal is to "reimagine the entire drug discovery process starting from the first principles using an AI-first approach".
Because they are interested in the first principles of things, let's first clarify what AI means. It does not refer to anything artificially intelligent. Instead, it refers to machine learning, of which DeepMind is an acknowledged expert. AI is a classic example Orwellian newspeak that has been adopted by tech companies to clean up a data-gobbling and energy-intensive technology. It has, like most digital things, both socially beneficial and dystopian uses.

DeepMind's new venture seems to be more socially beneficial. It is because DeepMind's researchers discovered that their technology could play an important part in solving a fundamental problem in biology: protein folding.

Proteins are complex, large molecules that do the bulk of the work in living organisms. Proteins are essential for almost every bodily function, including the ability to contract muscles, sense light, absorb food, and run the immune system. The three-dimensional structure of a protein determines what it can do.

The problem is that, although we now know the genetic sequence of a protein's DNA, it doesn't tell us how the protein will be folded into the 3D form that determines its function. It is very difficult to predict structure from the sequence. In 1969, American molecular biologist Cyrus Levinthal calculated that there were 10 possible foldings for a typical protein. This would be more than the time it takes to unravel the universe.

Every two years, researchers working on the folding problem hold a competition to predict the structure of proteins. DeepMind's Alphafold machine learning software, entered the competition in 2018 and surprised many other teams. Nature, who is normally quite serious, said that it was "head and shoulders above" the 100 participating research teams in the 2020 competition last November.

It is easy to see why DeepMind believed it was onto something. One, it could be a huge benefit for medical diagnosis and treatment to quickly determine the structure of protein. For example, protein misfolding may play a significant role in Parkinson's Disease. However, the tech industry is vast, but healthcare is much larger. If Alphabet is looking for a way to make search more efficient, Isomorphic Labs could be a smart investment. It might help to ease some DeepMind geeks' concerns about working for a surveillance capitalism monopoly.

AI is a classic example Orwellian newspeak that has been adopted by the tech sector

DeepMind's recent foray into molecular Biology will be celebrated by machine-learning enthusiasts as proof that traditional methods of science, such as formulating hypotheses and testing them, gathering data, or performing peer review, are obsolete. If this is what enthusiasts believe, they don't understand science or machine learning. DeepMind's brilliant demonstration is that smart algorithms combined with brute-force computing can be used to accomplish specialized tasks such as gaming, which has clear rules and criteria for success. Predicting protein structures is another example. It remains to be determined if the combination works in less structured domains.

This doesn't mean the technology can't be used in many scientific areas. It excels at extracting patterns from huge data sets. This is why it is loved by both particle physicists as well as astrophysicists. Many observers consider it a valuable tool for discovery. Machine learning allows you to find correlations in large amounts of data - sometimes unexpected and strange ones, like when Walmart discovered that customers who had bought pop tarts from their local grocery stores before hurricane warnings hit.

These correlations might prove useful for a supermarket retailer. However, correlation is not the same as causation, as every GCSE maths student will tell you. The per capita margarine consumption in Maine is closely related to the Maine divorce rate. The number of people who drown in swimming pools closely correlates with the number films in which Nicolas Cage has been seen. These correlations may seem amusing or striking, but they do not contribute to our understanding. To understand how an effect is caused, we must first know what it is. Not a fancy neural network, but a scientific method is necessary.

What I have been reading

Overload

The Singularity is Here is an essay by Ayad Akhtar, Atlantic, about what surveillance capitalism is doing for us.



Streetwise

Memory Lanes: Google's Map of Our Lives by Sirin Kale is a charming feature in the Observer magazine that focuses on memories triggered when one searches Google Streetview and sees people from their past.



Ready to take-off

Mario Vargas Llosa: My Fear of Flying has been written by Mario Vargas Llosa on the FSG Work in Progress site.