Have you heard of CRISPR? It’s most often discussed as a revolutionary genome-engineering tool — and rightly so, since repurposing it has transformed molecular biology, culminating in the ability to edit mutations directly in the human genome. But what does a CRISPR system do in nature?
CRISPR–Cas systems are the adaptive immune system of bacteria: they snatch little snippets of invading viral or plasmid DNA, store them as “spacers” between repeats in the bacterial genome, and then use those spacers as guides to find and destroy the same invader next time it shows up.
The protein components of CRISPR–Cas carry out the snatching, storing, and targeted destruction of foreign DNA. Today I want to highlight a recent update published in Nature Microbiology which reveals just how extraordinarily different these systems look across prokaryotes. I’ll explain (1) why this diversity poses a serious challenge to standard evolutionary narratives from a phylogenetic perspective, and (2) why I’m skeptical about phylogenetics-based classification.
Unique CRISPR Systems
In the paper, the researchers report the number of unique CRISPR systems identified:
- 2 classes
- 7 types
- 46 subtypes
- And many ultra-rare variants that together form a “long tail” of CRISPR diversity.
In summary, CRISPR systems come in all shapes and sizes, exhibiting diversity of both gene sequence and system architecture. As an example of architectural differences, Class 1 systems (Types I, III, IV, VII) use multi-protein complexes (Cascade-like) for targeting while Class 2 (Types II, V, VI) systems do targeting with one big superstar protein (Cas9, Cas12, Cas13).
Now, the authors state that classification of these different systems should be based to the “maximum extent possible, on evolutionary relationships.” But they run into a problem taking that approach! In the very next sentence they say there are “no universal markers suitable for comprehensive phylogenetic analyses.” This means the systems are so different there’s no single way to compare them all. The authors attribute the high degree of variation to these systems “evolving fast”, although there’s no independent evidence provided for this claim.
Because of the lack of universal markers, they had to use something called position-specific scoring matrices (PSSMs) rather than individual protein sequences to do a phylogenetic comparison. Even that proved insufficient; they ultimately needed multiple PSSMs because of extreme sequence divergence.
A Challenge for Unguided Evolutionary
Bottom line: this 2025 classification reveals how staggeringly diverse prokaryotic immunity really is. The more we dig, the more exquisite molecular uniqueness we uncover (a long tail of diversity!), although the most common CRISPR architectures seem to have now been identified and aren’t expected to change. The implication: such a significant amount of diversity raises a challenge for unguided evolutionary mechanisms. This is because studies have indicated there are serious time constraints for even one protein to transform stepwise into a radically different one (Axe 2004; Axe 2010; Dilley et al. 2023).
While these findings clearly raise the hurdle for Darwinian evolution, they also add to the ever-growing pile of evidence for design. When you look at the gene diagrams (Figure 1 and 2) in the paper, consider how designed these systems look. They all share modules which carry out the functions of snatching, storing, and targeted destruction, but they achieve those functions in remarkably different ways. That is exactly what a well-designed approach to antiviral defense would look like, since uniform systems would be far more vulnerable to counter defenses by viruses.
The CRISPR Classification Crisis
This big new review on type III CRISPR systems is impressive: tons of work, beautiful figures, lots of new CRISPR genetics. But one thing kept nagging at me while I was reading it: the authors really, really want these systems to fit into a classic branching phylogenetic tree even though the data is basically screaming “Please don’t!”
The authors say outright: there’s no single universal marker gene. And yet… boom, in Figure 3a, here’s this stunning, colorful phylogenetic tree built entirely on one protein (Cas10) that “evolves rapidly.”
So, I’m staring at this lovely tree (Fig. 3a) and thinking: What does this actually tell me? According to the authors, the sensor and effector proteins are scattered all over the tree like confetti. They acknowledge that: “Both sensors and effectors are scattered across the clades of the Cas10 tree, which is suggestive of extensive module shuffling and horizontal gene transfer shaping of the cOA and SAM–AMP signalling pathways in type III systems.” But they justify retention by stating, “Despite the extensive shuffling of signalling pathway components, some trends are notable.”
“Some trends” are cool! But then maybe… don’t present a tidy-looking family tree?
What actually got me excited in this paper was the shout-out to CasPEDIA — the database that classifies Class 2 CRISPR systems (the ones used in gene editing) not by how related their sequences are, but by what they actually recognize and cut, and what kind of activity they have. In other words, function first. That approach seems like it’s listening to what the biology is trying to say.
More Caution Is Necessary
I’m not fundamentally against phylogenetic trees for classification; I just think more caution is necessary. We use comparisons all the time with human-made objects: you can draw a perfectly useful “evolutionary tree” of t-shirts (sleeveless → short-sleeve → long-sleeve → hoodie) or cars (horse carriage → Model T → Prius → Cybertruck) without pretending it reflects actual shared ancestry, and doing that sometimes does help you classify. However, that is not always the case. For example, if you picked thread, color, or shape to draw a phylogeny of clothing, you’d get very different phylogenetic trees. The thread phylogeny might help you choose which color of thread to buy and use, but the shape phylogeny might not.
Given that the utility of a phylogeny for classification depends on the end goal, I’m concerned about the extent to which phylogenetic classification can produce genome-engineering utility for CRISPR systems. I’m not convinced that the sequence differences reliably track function, especially in a way that might be helpful for genome engineering. For example, if sequence differences in Cas10 put two systems into completely different branches, but they still do the exact same biological job… are we actually helping anyone by putting them far apart on the tree? Or are we inadvertently making it harder to spot the patterns that matter?