A recent study in PNAS by Sahakyan et al.1 investigated the evolution of protein folds in early evolution, prior to the Last Universal Common Ancestor (LUCA). The team developed the Protein Fold Evolution Simulator (PFES), a computational method for simulating the evolution of globular protein folds from initially random genetic sequences. Key steps in PFES are the (i) introduction of random mutations, (ii) evaluation of their fitness effect using AI-based protein structure prediction, and (iii) subsequent selection. It is the second step — fitness evaluation — that presents a significant challenge.
There is a robust literature surrounding the problem of understanding the fitness landscape in protein sequence space. This problem has many significant challenges. For example, the sequence space is astronomically large, experimentally measured fitness is context-dependent and measurements are mostly clustered in the proximity of wild-type sequences, and the protein sequence fitness landscape may be rugged and influenced by epistatic effects.
No Less Challenging
These problems are no less challenging for Sahakyan et al. How is fitness evaluated in an early evolution, pre-LUCA, context with rudimentary organisms and unknown environments? How is fitness evaluated for random or near-random sequences, or disordered protein structures? Nonetheless, Sahakyan et al. use an incredibly simple fitness score for their simulated evolution of single, globular protein-coding genes.
Their fitness score includes metrics of the quality of the AI structure prediction, and a contact number metric which is an indicator of the degree to which the protein structure is folded up. This contact number metric is an aggregate approximation of the structure’s similarity to functional, globular proteins. In general, as the number of contacts increases, the structure approaches a realistic, native-like protein. It is, effectively, a proxy for the progress in evolving a protein-coding gene.
A Key Component
Hence, it is a key component in the fitness score, as it drives the PFES toward native protein structures. In addition to the contact number metric, the fitness score also ensures realism by constraining the lengths of the overall protein structure and any alpha helix and beta strand secondary structures.
Therefore, it is of little surprise that PFES drives random sequences to sequences that form native-like, globular tertiary structures — that is precisely what the fitness score points toward.
Note that the fitness score is based exclusively on structural metrics. This is a significant shortfall of the study — the fitness score does not actually model fitness. There is no scientific evidence, for example, that a mutation occurring in a random sequence causing slightly more contacts would result in improved organism fitness. Yet such a sequence would be assigned a higher fitness score in the PFES. Also, the structural metrics do not map to any particular structure, but rather represent a large number of possible structures, with varying function and fitness implications.
Another problem with this study is its use of AI structure prediction tools. These tools are based on a training set of known protein structures and their amino acid sequences. These structure/sequence combinations are, for the most part, proximate to native, wild-type, proteins. This is an enormous sampling bias and it is likely that the AI predictions are biased toward native-like structures. This would make the job of simulating the evolution of a protein, even from an initially random genetic sequence, much easier.
Simulating an Evolutionary Process?
This combination of (i) using a fitness score that is incorrectly based on the protein structure, and (ii) using protein structure predictors that are subject to enormous sampling bias, means that the PFES is not simulating an evolutionary process. Rather, it is driving initial structures toward folded, globular structures, in the AI structure prediction space.
This is in stark contrast to the study’s claims. Sahakyan et al. claim they have developed a computational tool to “recapitulate protein fold evolution in detail.” Furthermore, they claim that using PFES, they have (i) “shed light on the enigma of the rapid evolution of diverse protein folds at the earliest stages of life evolution,” and (ii) found that in the early, pre-LUCA, world, the “emergence of simple, stable, globular protein folds from random amino acid sequences is relatively easy and could occur quickly.” These claims are unfounded. While a charitable reading is always desirable, it would be misleading to characterize this study as anything less than false science. The researchers designed a simulation tool, and then falsely claimed that it represents the evolutionary process.
Notes
- H. Sahakyan, S. G. Babajanyan, Y. I. Wolf, E. V. Koonin, In silico evolution of globular protein folds from random sequences. Proc. Natl. Acad. Sci. U.S.A. 122 e2509015122 (2025).