If U.S. engineers built a spaceship with hyperdrive, and a foreign country managed to reverse-engineer it and figure out how it works, who should get the credit? What is the bigger accomplishment: reverse-engineering a futuristic craft, or designing one from scratch?
DeepMind is a leader in artificial intelligence (AI). Its geniuses managed to beat humans at the popular name Go using its AlphaGo algorithm. Its AI systems have now reached 90 percent success at predicting how a protein will fold. A blog post from DeepMind explains why this is a big deal:
In his acceptance speech for the 1972 Nobel Prize in Chemistry, Christian Anfinsen famously postulated that, in theory, a protein’s amino acid sequence should fully determine its structure. This hypothesis sparked a five decade quest to be able to computationally predict a protein’s 3D structure based solely on its 1D amino acid sequence as a complementary alternative to these expensive and time consuming experimental methods. A major challenge, however, is that the number of ways a protein could theoretically fold before settling into its final 3D structure is astronomical. In 1969 Cyrus Levinthal noted that it would take longer than the age of the known universe to enumerate all possible configurations of a typical protein by brute force calculation — Levinthal estimated 10^300 possible conformations for a typical protein. Yet in nature, proteins fold spontaneously, some within milliseconds — a dichotomy sometimes referred to as Levinthal’s paradox. [Emphasis added.]
Simple by Comparison
Reverse-engineering a hyperdrive looks simple by comparison. In 1994, a professor started a contest for AI specialists named CASP: Critical Assessment of [protein] Structure Prediction. Every two years, contestants try to predict a protein’s fold from its amino acid sequence alone, without knowing the fold in advance. Before now, scores achieved 20 to 40 on the Global Distance Test (GDT), a measure of the distance between predicted amino acid positions versus actual biological positions. DeepMind achieved an average score of 60 with AlphaFold in 2018. They increased it enormously this year to 92.4. The blog entry pictures how closely the predicted fold matches the actual fold for two cases. They appear to overlap very closely.
This computational work represents a stunning advance on the protein-folding problem, a 50-year-old grand challenge in biology. It has occurred decades before many people in the field would have predicted. It will be exciting to see the many ways in which it will fundamentally change biological research.
Achieving this success drew inspiration from biology, physics, and machine learning, along with leading experts in protein structure. The team constructed a neural network to approach the challenge, solving small clusters of amino acids then using deep learning methods to explore how they might join up. Even so, the CASP contest uses relatively simple proteins called domains. AlphaFold has more trouble figuring out proteins that interact. Nature News says,
The network also struggles to model individual structures in protein complexes, or groups, whereby interactions with other proteins distort their shapes.
Nevertheless, the success represents “a gigantic leap” that “will change everything,” Ewen Callaway writes. In what ways? John Moult, a professor at the University of Maryland who co-founded CASP, explains in The Scientist,
“This will change medicine. It will change research. It will change bioengineering. It will change everything,” Andrei Lupas, an evolutionary biologist at the Max Planck Institute for Developmental Biology in Germany who helped judge the contest, tells Nature, adding that AlphaFold took only 30 minutes to produce the structure of a protein his lab had been trying to figure out for 10 years.
Writing in Science Magazine, Robert F. Service adds,
Knowing those shapes helps researchers devise drugs that can lodge in proteins’ crevices. And being able to synthesize proteins with a desired structure could speed development of enzymes to make biofuels and degrade waste plastic.
Lest We Forget
This is great news, and rightly applauded. But we need to remember that this folding problem that has baffled humans for 50 years is solved rapidly in living cells at every moment. Levinthal noted that proteins routinely “fold spontaneously, some within milliseconds” inside the cell. A few need help from chaperones to find their native fold, but many go directly from 1D amino acid sequence to 3D functional protein.
That’s not all. The cell has repair enzymes, too, that can dismantle improperly folded proteins and fix them or replace them if they are irreparable. In our hyperdrive spacecraft analogy, it is like having robots able to detect failed components, pull them out, fix them, and re-insert them. How does a cell without eyes and brains do this? Think of the sophistication of algorithms able to perform such operations!
The concept of Search bears heavily in intelligent design theory. If the search problem is complicated enough, succeeding requires additional information beyond what blind search can achieve in the time available. Finding a marked atom in a galaxy, for instance, would take far longer than the age of the universe to succeed. William Dembski proved in his book No Free Lunch that no evolutionary algorithm is superior to blind search, unless auxiliary information is provided. The catch-22, though, is that the searcher needs to search through all possible sources of the auxiliary information to know which one is correct. Dembski likened it to finding a buried treasure by digging at random on an island. That kind of blind search is highly unlikely to succeed if the island is big enough. If the searcher is handed a map, he could go directly to the spot with that auxiliary information. All well and good, like in the movie It’s a Mad Mad Mad Mad World, where the information provided was accurate from somebody who knew.
The Absence of Foresight
Evolutionary algorithms, though, having no foresight, could generate a billion treasure maps, only one of which might be correct. The search problem then switches to finding the correct map out of billions of treasure maps. If a book on the shelf tells you which map is the correct one (more auxiliary information), how does the searcher know? The searcher would have to check a billion books offering random answers with that information, only one of which might be correct. Each added piece of auxiliary information must be checked out by another search. That’s why no evolutionary algorithm is superior to blind search. The only way to speed to the buried treasure is to trust a source that knows and test the information by digging there. That information must come ultimately from a mind — someone who knows the right answer.
Returning to the protein folding problem, we have seen that the search space for protein folds is vast beyond comprehension, like an island as big as the universe. Observing cells routinely folding proteins quickly and accurately, one can conclude therefore that a mind was behind the information. That conclusion is certified by watching AI experts using their minds to reverse engineer protein folds. AI is not inventing sequences that will fold; it is trying to figure out how a given sequence will produce an observed functional fold. Inventing a fold de novo is the harder problem (Stephen Meyer discusses functional folds in Signature in the Cell, pp. 99ff.)
What About Evolution?
Does evolution come into the story at all? Some of the workers in the CASP contest are evolutionary biologists. Thinking that some proteins evolved into other proteins through mutation and natural selection, they believe that similar proteins are connected by common ancestry. This belief, they feel, can allow them to “evolve” new proteins with similar folds. Callaway says, “Some applications, such as the evolutionary analysis of proteins, are set to flourish because the tsunami of available genomic data might now be reliably translated into structures.” That is intelligent design, however, not Darwinian evolution; take the word “evolutionary” out of “evolutionary analysis,” otherwise it is an oxymoron. There are limits to the amount of variation a fold can endure and preserve function. (Douglas Axe discusses these limits in Undeniable, p. 80ff and 180-182. See coverage by Evolution News here and here about simplistic evolutionary approaches to solving the folding problem, and why they miss the mark.)
If DeepMind’s AlphaFold algorithm succeeds in designing new enzymes, it will be through intelligent design, not blind search. It will build on the information in working proteins, extending that knowledge by design. Materialistic evolutionary processes have no such foresight.
In short, DeepMind’s achievement is laudable, but the real prize goes to the designer of protein systems: their encoding in DNA, their translation in the ribosome, their spontaneous (sometimes chaperone-assisted) folding, their functions, their interactions, and their repair mechanisms. All those get perfect scores when not harmed by random mutations that degrade information. AI has not even begun to imitate those capabilities. Any higher scores through AI in the future will be attained by intelligent design, not evolution. The news only underscores the superior knowledge built into the molecular basis of life.