Science1 publisher3 min readPublished
Aphid genomes let AlphaFold2 trace 2,400 plant-hijacking proteins to one reworked fold
Stowers Institute scientists fed AlphaFold2 new aphid genomes and got about 2,400 structures of plant-hijacking proteins built on one reworked fold. Two X-ray structures anchor those predictions. What the proteins do inside a plant is still unknown.
The Scientist · Science desk
What happened
- Aphid BICYCLE proteins, a family Stern's Stowers lab had named earlier, matched nothing in sequence databases because their sequences change so fast.
- With Angela Gronenborn's lab at Pittsburgh, the team spent years solving two BICYCLE proteins by X-ray and found a version of the saposin-like fold.
- AlphaFold2 got the structure wrong at first because its search for related proteins came back empty.
- The team then generated about 2,400 high-confidence predictions across seven aphid species, with the saposin-like fold reappearing throughout in repeated, rearranged forms.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- cost For proteins with no cataloged relatives, getting a usable AlphaFold2 model can mean field collection and genome sequencing first, and that time and travel is spent before any computing starts.
- decision Anyone building on the roughly 2,400 models has to decide how far to trust predictions that outnumber experimentally solved structures about 1,200 to one.
- capability Sequence searches could not place this family, but researchers can now compare its members by shape, protein against protein, across seven species.
AlphaFold2's early miss came down to how the tool works. The program's accuracy does not come from machine learning alone. It leans heavily on the evolutionary data it is given [8]. BICYCLE sequences have been rewritten so quickly that database comparisons no longer detect any family resemblance [5], so there was no evolutionary data to give it. "When it searched the database for similar proteins, it couldn't find any," said David Stern of the Stowers Institute for Medical Research, whose lab did the work with Angela Gronenborn's group at the University of Pittsburgh [9][1]. "That part of the AlphaFold2 program was empty. It was actually an empty box." [9]
The design had a control before the AI was involved. The two crystal structures came first, and AlphaFold2 was released the same week, according to Phys.org's report [6][19]. The team then filled the empty box. It collected aphids across Virginia and West Virginia, traveled to Japan for one species and sequenced their genomes [10]. Now the output could be scored against a measured answer. "Lo and behold, AlphaFold2 gave us back the crystal structure that we had solved," Stern said [11].
The scale-up rests on that check. The paper, published Sept. 2 in the Proceedings of the National Academy of Sciences [2], reports roughly 2,400 high-confidence predictions across seven species [12]. Two structures were solved experimentally, which works out to about 1,200 predictions for each one measured directly [18]. The report does not say how many of the rest were tested against experiment.
What the predictions show is one fold used many ways. The saposin-like unit is duplicated, reoriented and decorated differently from one protein to the next [13]. That shared blueprint became visible only after the new genomes filled in the sequence data [10][11].
The speed of change is also the biological puzzle. Proteins at the contact point between a parasite and its host tend to evolve fastest, because each side is under pressure to counter the other's latest move [17]. Aphids inject hundreds of these proteins into plants and use them to build galls, structures of plant tissue shaped to house and feed their young [3]. The researchers suggest the extensive remodeling may give aphids a molecular arsenal for manipulating plants and slipping past their defenses [14].
It is still unknown what any one BICYCLE protein does inside a plant. The structures were meant as a way toward that answer. "One way to try to gain insight into that problem is to solve the 3D structure of these proteins, and that was the motivation for this project," Stern said [16].
The authors say the approach could help with other rapidly evolving proteins, including those involved in immunity, host-parasite interactions and agriculture [15]. In this study the slow step was biological sampling: insects gathered in two US states, plus a trip to Japan for one species [10].
What to watch
- Further experimental structures of BICYCLE proteins beyond the first two, to test whether the predictions hold up outside the crystallized pair.
- Functional tests linking particular versions of the saposin-like fold to gall building or to evading plant defenses.
- Attempts to apply the same genome-first route to fast-evolving immune or host-parasite proteins that also lack database relatives.