Life Finds a Way

by Andreas Wagner

  While the molecular revolution deepened our view of life immensely, it left the principle behind Wright’s fitness landscapes untouched. Biologists still think of organisms as having some adaptive value or fitness. We still think of an organism’s genotype as a location and its fitness as an elevation in a fitness landscape. And when populations of organisms find creative solutions to a problem they face, such as how to swim efficiently or how to escape predators, we still envision that they explore a fitness landscape and climb its peaks, and we still imagine this landscape as a mountain range in three dimensions because we cannot cram the vast hypercubes of high-dimensional genetic spaces into our limited minds.

  Just like relentless competition enables some humans to win a race to the top, natural selection enables populations of organisms to climb any one peak in a fitness landscape. To do so, they need natural selection. But the complexity of the actual landscapes revealed by the molecular revolution lays bare the fact that natural selection is not enough. The landscape of ammonites had two peaks, and that of passion-vine butterflies had dozens of peaks, but truly complex landscapes can harbor many more. Not only can these peaks have different heights, but landscape topography can vary in many other ways as well. Some peaks might be gently sloped, whereas others might be jagged. The peaks may be isolated and scattered throughout the landscape, or they might form a continuous mountain range connected by ridges and saddles.

  Landscapes like these can limit natural selection’s powers because of selection’s proverbial blindness. When selection works on a population scattered along the face of a mountain, it eliminates all downslope mutants and preserves only those upslope, blindly driving the population toward the nearest peak. A population that starts out at the foot of a low hill can get to the nearest peak—a local peak, in scientific jargon—but that’s also where it will get stuck. In its relentless uphill drive, natural selection will not allow a population to cross the valley separating this local peak from the next higher peak. It will ruthlessly prevent inferior variants from surviving, just like those predators that kill butterflies with rare, off-peak warning colors. Even a population laboring to climb the slope of the very highest or global peak might reach some jagged outcrop on its way up. To rise further it would have to backtrack downhill at least a few steps before resuming its climb, but because of natural selection, it cannot do so. In rugged landscapes like that of Figure 1 in the prologue, the roof of the world might be near but forever out of reach because natural selection is like a powerful engine with a crucial flaw—it can only go uphill.

  Wright was already worried that fitness landscapes have many peaks, but it was not until 1987 that it became clear how bad the problem can be.6 That’s when biologists Stuart Kauffman and Simon Levin estimated the number of peaks under the simplest possible theoretical assumption—that the fitness of different genotypes is drawn at random from some possible range of values. This assumption is as good a starting point as any other given how utterly impossible it would be to measure the fitness of all possible genotypes. Just consider the 10120 genotypes of Wright’s original landscape. Even if every one of the seven billion humans alive today dropped everything they are doing and dedicated the next hundred years to the exceedingly important task of measuring the fitness of fruit flies, and did so at a speed of one fly per second, they would be able to process 1020 flies, a formidable number to be sure, but less than one in a 10100th of all the flies in Wright’s fruit fly landscape.7

  The calculations of Kauffman and Levin show that even in their simplified, theoretical landscape, where every gene has only two possible alleles, about one in every fifteen thousand genotypes would be a peak. That number does not sound too bad, until you compute the total number of peaks. It turns out to be a one with more than four thousand trailing zeroes.8 Thus, not only is the size of fitness landscapes beyond imagination, but the number of peaks can be no less mind-boggling. And to make matters worse, this number of peaks increases explosively with the number of possible genotypes.9

  Only one among all these peaks is Mount Everest, the single global peak among myriad lower ones. Natural selection can fulfill its promise of finding the best-adapted organism only if Mount Everest can be reached through a constantly ascending path, one that could take thousands or millions of steps to traverse. To find out if such a path exists, Kauffman and Levin first computed the average number of steps that it would take a population, starting out at some arbitrary place, to ascend the nearest peak, from which natural selection could go no further. They found that the number of steps to the nearest peak was paltry, smaller than fifteen, and not nearly enough to get a population to Mount Everest.10 Most populations would wind up on the nearest molehill.

  Theoretical calculations like this cannot replace experiments that sketch the true contours of a fitness landscape, count its peaks, and trace all access routes to them. Unfortunately, such experiments will never be able to map any one landscape completely, because genomes have so many variants, but they can focus on smaller regions, such as those where only one gene varies. This is useful because any one gene encodes a protein, and proteins are not only the workhorses of our cells, but also crucial links between the genotype and the phenotype. Each cell harbors thousands of different kinds of proteins, each encoded by a gene, each with a different task. Proteins are among the smallest parts of an organism with a phenotype worth studying.

  Each protein is encoded by a string of DNA letters, and the collection of all possible such strings—also called a space of sequences—is a giant realm of possibility. Think of it as a library of texts that encodes not only all of the countless innovative proteins that evolution has discovered in its history, but also all the proteins that it could discover in the future. It is the space where nature goes to find new parts for its biochemical machines.11

  To map the fitness landscape of this space is to measure how well each of its DNA sequences—or, equivalently, the amino acid sequences they encode—is suited for a specific task. It is to measure the speed at which protein enzymes can cleave a sugar molecule, the pull motor proteins can exert in a muscle, or the rate at which transport proteins can ferry nutrients into a cell. And because this landscape’s topography channels the movement of populations—uphill, always uphill—it also imposes limits on nature’s creativity in finding new and better proteins.

  Unfortunately, even this library of protein texts is too large to explore completely—there are more than 10130 proteins with one hundred amino acids each, and many proteins are much longer. Therefore, experimenters must focus either on shorter strings, of which there are fewer, or on a smattering of paths through the landscape. Even that requires technologies to manufacture numerous DNA and protein strings, and it had to wait until the first decade of the twenty-first century, when such technologies became efficient enough, eighty years after Wright birthed the landscape idea.

  Some such paths through the vast library of protein texts lead to innovations that save lives—not those of humans, but of our lethal enemies, disease-causing bacteria. These bacteria have discovered proteins called beta-lactamases that disarm the offensive weapons—antibiotics—that doctors use to kill bacteria. Named after beta-lactam, a ring of atoms that occurs in antibiotics like penicillin, beta-lactamases can destroy this ring and defuse these antibiotics. Because beta-lactamases save bacteria from death, they spread like wildfire through bacterial populations—courtesy of natural selection—while endangering the lives of patients, who are left helpless when overrun by a bacterial infection. Innovations like beta-lactamases are nature’s defenses in an endless arms race between medical researchers, who constantly develop new offensive weapons, and bacteria, whose vast populations scour nature’s DNA libraries for new ways to neutralize these weapons.

  An especially important offensive weapon is cefotaxime, a broad-spectrum antibiotic capable of destroying many different kinds of bacteria. It is on the World Health Organization’s List of Essential Medicines. Alas, it may not remain
there much longer, because of a simple disturbing fact: disarming cefotaxime requires nothing more than a few tweaks in today’s beta-lactamase proteins.

  A conventional beta-lactamase disarms cefotaxime slowly, too slowly to help bacteria survive the hefty doses prescribed by doctors. But it turns out that with only five letter changes in such a beta-lactamase the protein is rendered one hundred thousand times more efficient at destroying cefotaxime.12 The new protein variant is a high peak—although perhaps not the highest—in the fitness landscape of proteins that can destroy cefotaxime. How hard is it to reach this peak? Is it the top of a craggy mountain or a smooth sugar cone? The ideal experiment to answer these questions would manufacture all protein variants around this peak, measure their ability to destroy cefotaxime, and find out which of them are lower outcrops that could stop selection’s march. Alas, their number is too large. There are more than a trillion proteins that differ from conventional beta-lactamase by five or fewer amino acids, more than can be easily manufactured with current technologies. But even though we cannot map every local peak and valley, we can still get a glimpse of the whole by following individual paths.13 Here is how.

  Imagine you are blind, standing at the foot of a mountain and wanting to climb it. You cannot see the best path to the peak, but you can distinguish an uphill from a downhill step, so you feel your way as you take step after step. If the mountain is perfectly smooth, then every path of uphill steps will eventually lead to the peak. Some of these paths will meander uphill in serpentines, others might spiral slowly toward the peak, while yet others will lead straight up, but every single one will eventually get you there. Not so if that mountain is craggy. In that case, most paths will get you stuck on some outcrop below the peak, and only a few—not necessarily the straight uphill ones—may lead all the way up. To find out how craggy the mountain is, you can do this: try the same climb multiple times, with sequences of steps in different directions, but all uphill, and count how often you get stuck. If all your attempts succeed in reaching the peak, the mountain is perfectly smooth. If you get stuck every time, it is maximally rugged.

  In 2006, Daniel Weinreich, then a postdoctoral researcher at Harvard University, translated this very idea into an experiment on beta-lactamases, tracing paths from the original beta-lactamase protein to its cefotaxime-destroying variant in which five amino acid letters are changed. Each letter change is one step toward the peak. Because five letters can change in different orders, different paths lead up the peak, just like you can transform the word BOLT into GOLD by editing two letters in different ways, either from BOLT to MOLD to GOLD, or from BOLT to (the meaningless) GOLT to GOLD. There are 120 different orders in which five different amino acid changes can occur, each of them a different path toward the cefotaxime peak. Weinreich and his collaborators synthesized all proteins along each path and measured their ability to destroy cefotaxime in order to identify which paths are dead ends.

  Most of them are, it turns out. More than 90 percent of the paths lead some way up the peak but then encounter a protein that cannot be improved by a single further step. And because natural selection prohibits backtracking downhill, evolution’s uphill climb would end right there.14

  A dozen further experiments run by other researchers like this have climbed peaks elsewhere in the vast landscapes where molecules and organisms evolve. They created bacteria that can grow and divide faster on the same diet, HIV viruses that can infect human cells more efficiently, and enzymes that manufacture new self-defense chemicals for plants. And they reveal a similar topography, one not as hopelessly rugged as the landscapes studied by theorists like Kauffman and Levin, but still much more rugged than a sugar cone.15 Among the many paths that lead toward a peak, only a few reach it. On the remaining paths, natural selection dead-ends below the highest peak—sometimes far below. In evolution’s landscapes, a climber’s risk of getting stuck near base camp is very real.

  For billions of years, proteins have produced a steady stream of innovations, but another kind of innovative molecule—RNA—has been at it even longer. This ugly duckling of molecular biology, long thought to be a mere carbon copy of DNA, helping to make protein, metamorphosed into a swan when biochemists discovered in the 1980s that it can do so much more.

  Like protein, RNA can catalyze chemical reactions, but unlike protein, its letter sequence also stores the same kind of heritable information as DNA. Its special talents have helped cast RNA as the leading actor in some of the invisible dramas that play out inside every living cell. For instance, RNA collaborates with proteins in an enzyme called telomerase by helping to maintain pesky chromosome ends called telomeres, which tend to shorten over time, a bit like fraying shoelaces, except their fraying has more serious consequences. When it goes unchecked—for example, because the telomere maintenance crew is too slow—cells quickly stop to divide, age, and die. That’s bad, but even worse is when that telomerase is hyperactive, because then cells can start to divide uncontrollably and become cancerous.

  Another RNA-equipped biochemical machine is just as remarkable because it opens a window into the very early history of life itself. This is the protein-manufacturing ribosome, a hugely complex apparatus comprised of several RNA strings and more than fifty proteins. Among all these molecules, RNA plays the most important role because it is one of the ribosome’s RNA molecules—transcribed from a special RNA-encoding gene—that performs the crucial task: stringing amino acid parts together, letter by letter by letter, to build a protein string. The ribosome is one of several clues that early life was an RNA world, one where RNA ran the show the way proteins do now.

  Another telltale remnant of this sunken empire is a bizarre process that allows some genes to encode multiple proteins. Once a gene’s DNA has been transcribed into an RNA carbon copy, a cell sometimes deletes short pieces of that copy and splices the remaining pieces together. When the same gene is transcribed twice or more, these deletions can occur in different places, creating different RNA transcripts that are translated into different proteins. Alternative splicing, as biochemists call it, is a nifty mechanism to create proteins with different functions from the same gene. It’s as if you created many shorter variants of a long poem by combining a few lines here and there, using a different combination each time. In human language, most such variants would be garbled and nonsensical, but in the chemical language of proteins, they can encode meaningful, useful proteins. And even though alternative splicing may seem bizarre, it can be quite important. For example, it produces variants of a human protein required to detect sound. These variants help tune cells in our inner ear to perceive sounds of different frequencies.16 No alternative splicing, no Bach, Bartok, or Beethoven.

  For this kind of creative editing, complex organisms like humans need another complex biochemical machine called the spliceosome, but simpler ones like bacteria don’t.17 What’s more, in some bacterial genes the transcribed RNA string itself—without any help from proteins—can do the job. It discards part of its own text and splices whatever is left into a new, shorter string. Such a wonder molecule is not only an RNA enzyme—biochemists call it a ribozyme. It is an RNA enzyme that modifies itself. Think of it as a poem capable of rearranging itself on the written page.

  Just like proteins, RNA molecules are texts written in a molecular alphabet—four nucleotide letters instead of proteins’ twenty amino acid letters—that form a library vast beyond imagination. Some of this library’s texts are capable of self-splicing, and one of them occurs in the genome of an otherwise unremarkable soil bacterium called Azoarcus. Eric Hayden, a young researcher in my laboratory at the University of Zurich, used the Azoarcus ribozyme as a base camp to climb a nearby peak in its fitness landscape.18

  Eric knew that this RNA molecule could use its self-splicing powers to join itself to another string of RNA with a specific letter sequence, while it would fail miserably at joining itself to a third string with a different letter sequence. But in earlier experiments, Eric had found
a more flexible ribozyme that could self-splice with both strings. This ribozyme was at the peak he wanted his molecules to climb, a peak that was only four letter changes away from the Azoarcus ribozyme. Eric synthesized all RNA molecules that lie between the Azoarcus ribozyme and this peak, which allowed him to study every single one of the twenty-four possible pathways that lead up the peak. He discovered that only one pathway actually leads all the way up—the others are impassable for natural selection, because they lead through valleys in the landscape. These experiments taught us that RNA fitness landscapes can be just as rugged as those of proteins.

  Researchers like Eric explore a fitness landscape the hard way, painstakingly synthesizing the molecules on all different paths to the peak. (His molecular journeys took more than a year of dedicated laboratory work.) Other researchers use automated synthesis technology to manufacture huge collections of molecules. This technology can enable them to catalog an entire library of molecules. The drawback is that this works only for small libraries comprised of molecules much shorter than the two-hundred-plus letters of beta-lactamase and the Azoarcus ribozyme.

  Actually, small is relative. In one such study, researchers at Harvard University created all possible RNA molecules that are twenty-four nucleotide letters long—more than 280 trillion of them—and scoured this library for molecules with a skill that is essential to all life: they can attach themselves to other, energy-rich molecules.