The Borderlands of Foldability: Lessons from Simplified Proteins

Abstract

Proteins make complex life possible, yet our understanding of their emergence remains limited. What are the information limits of protein folding, and how did the first proteins emerge? Protein simplification studies – in which contemporary folds are built from limited alphabets, symmetrized, fragmented, or shortened – have provided key insights into these questions. These studies use design constraints to address the discoverability of, and connectedness between, protein folds. By considering various environments, such as high salt concentrations or peptide-nucleic acid coacervates, the role of context in the emergence of folded domains is explored. Taken together, these studies support the early emergence of protein folds and reveal the existence of highly connected and readily traversable regions of sequence-structure space.

From so simple a beginning

Virtually all biochemical processes rely on biopolymers that adopt well-defined structures, even if only transiently. Consequently, it is difficult to envision complex life without molecular folding at its core. For life on Earth, folded molecules are comprised predominantly of nucleotides or amino acids, with the latter forming proteins, the chief workhorse of the cell. The central dogma emerged to harness the special properties of folded molecules, and, by the time of the last universal common ancestor (LUCA), a repertoire of complex protein folds (see Glossary) had been established¹. The profound necessity of proteins for contemporary biochemical functions, paired with the prebiotic accessibility of simple peptides², suggests that proteins and peptides date back to the earliest stages of life on Earth.

The very first proteins and peptides were unlike their contemporary counterparts, with properties that were shaped by the limitations of inchoate translation machinery and prebiotic chemistry. Specifically, the first proteins are thought to have been short in length and comprised of amino acids that were either abundant in the surrounding environment or producible by a primitive metabolism (see Box 1 for a discussion of amino acid alphabets employed in simplification studies). The extent to which the first proteins could have folded and performed biochemical functions, given these constraints, is a core question of the field of protein simplification³ (Figure 1A). By constructing simplified protein variants, often at the edge of foldability, simplification studies have yielded insights into the control and design of protein folds⁴, the mechanisms of protein fold evolution^5–7, the environments that support fold emergence^8–11, and the traversability between folds¹². Taken together, these studies clarify how complex, contemporary folds could have come from “so simple a beginning”.

Box 1. What was the primitive amino acid alphabet?

We describe the history of amino acid incorporation into proteins and peptides as having had an hourglass shape (Figure B1), in which the earliest peptides contained diverse amino acid types that were later winnowed down into an alphabet of ~10 types^38,79 upon the emergence and establishment of the translation system. Over time, the genetic code expanded to include more biosynthetically complex amino acids, ultimately arriving at the canonical set of 20 genetically encoded L-α-amino acid types. Now, in the era of biotechnology, the amino acid alphabet is poised to expand once again, this time to include exotic non-canonical amino acids⁸⁰.

While protein fold simplification studies are rarely contextualized to be at any specific stage in the evolution of the genetic code – reflecting a scientific focus on folding and fold traversability, as well as a desire to sidestep the profound complexities of the translation system and its evolution – a Crick-inspired stage of intermediate genetic code complexity is frequently envisioned. Analyses of metabolic centrality and biosynthetic mechanism³⁸, the structure of the genetic code⁸¹, carbonaceous chondrite composition⁸², and spark disk charge experiments⁸³ are used to infer the likely amino acid composition of proteins at this stage. These studies are broadly consistent with each other, and the amino acids A, C, D, E, G, I, L, P, S, T, and V are generally considered to have been the primary constituents of early coded proteins^8,9,11,54. Several non-canonical amino acids – such as ornithine, 2,4-diaminobutyric acid, and homoserine – may have been part of the genetic code³⁸ at this stage as well, and are periodically included in alphabet simplification studies²¹, particularly those that exclude the canonical basic amino acids.

Prebiotic chemistry experiments indicate that diverse non-canonical amino acids were likely present in the surrounding environment, and thus available for incorporation into the very first peptides and proteins, particularly by a weakly selective emergent translation machinery or prebiotic chemical process. Chief among the non-canonicals are D-α-amino acids. Reports of heterochiral peptides that bind iron-sulfur clusters⁴⁴ or phase separate with RNA^71,84 suggest the potential for broad biochemical roles.

Figure B1. Amino acid diversity in peptides and proteins over time.

The history of amino acid incorporation into peptides and proteins is thought to have an hourglass shape: The earliest peptides likely comprised diverse amino acids, including D-amino acids. Later, with the establishment of the genetic code, a subset of approximately 10 amino acid types was selected for incorporation into proteins. Over time, the genetic code expanded into the 20-amino acid alphabet found in contemporary biology. Now, in the era of biotechnology, the amino acid alphabet is poised to expand once more.

Figure 1. An overview of protein simplification studies.

(a) The primary types of protein simplification studies: symmetrization and fragmentation of rotationally symmetric folds, minimization of asymmetric folds, and amino acid alphabet reduction. Key studies of (b) symmetrization and minimization and (c) alphabet reduction. Fold names are taken from the Evolutionary Classification of Domains database¹³ where possible. In panel b, coloring emphasizes sequence repetitions and homo-oligomerization. The length of the repetitive element and the oligomeric state are indicated. In panel c, the number of amino acids associated with each construct (excluding purification tags and the N-terminal Met) is indicated. The presence of aromatic and basic amino acids are indicated with a benzene ring and a blue plus sign, respectively. If an experimental structure of a simplified variant is not available, a closely related natural protein or a predicted structure is presented instead. Abbreviations for each protein fold are as follows: TIM-barrel (triose-phosphate isomerase barrel); HUP (HIGH-signature proteins, UspA, and PP-ATPases); HhH (helix-hairpin-helix); DZBB (double-zeta β-barrel); PRTase (phosphoribosyltransferase); HTH (helix-turn-helix); R3H (named after the conserved RxxxH motif). PDB accession codes are indicated where available. For proteins without a deposited structure, structure models were generated from the simplified sequences using AlphaFold3¹⁴. All structure figures were prepared using PyMOL (www.pymol.org).

Symmetric folds evolved from the duplication and fusion of oligomerizing peptides

In 1966, Richard Eck and Margaret Dayhoff hypothesized that contemporary ferredoxins were produced by the duplication and fusion of a short peptide composed of just four types of simple amino acids (Ala, Asp or Pro, Ser, and Gly)¹⁵. Their insight – that contemporary folds evolved via a simple and common mechanism of accretion – has become an organizing principle of protein structure evolution, particularly for rotationally symmetric folds¹⁶, and multiple groups have successfully retraced Eck and Dayhoff’s proposed evolutionary trajectory in the laboratory.

Fully symmetrized folds – that is, idealized proteins with sequence repetitions that correspond to the rotational symmetry of the fold – represent the immediate product of a fragment duplication and fusion event. Thus far, fully symmetric variants of the β-trefoil^5,17,18, β-propeller^6,19, TIM barrel²⁰, (HhH)₂-fold²¹, ferredoxin²², and double-ψ β-barrel²³ have been reported. Biophysical characterization of various fully symmetric folds has revealed them to be reasonably well-behaved – soluble, stable proteins with cooperative folding transitions. For example, the designed β-trefoil Symfoil-4P, which has a perfectly repetitive sequence, exhibited a Δ_uG = 55.1 kJ/mol, 34 kJ/mol higher than the wild-type protein from which it was derived⁵. Studies of the β-trefoil also suggest that fully symmetric proteins may be more robust than symmetric proteins with non-repetitive sequences to perturbations such as circular permutation²⁴. Furthermore, it has been proposed that there is an algorithmic bias in evolution, which favors high symmetry even in the absence of selective pressure²⁵. Taken together, these properties suggest that sequence-repetitive rotationally symmetric folds are viable evolutionary intermediates.

Fragmentation studies (Figure 1B), where the foldability of a subdomain from a rotationally symmetric fold is characterized, demonstrate that short peptides can form symmetric folds via homo-oligomerization. Peptides ranging from 28-74 residues in length have been shown to oligomerize into a β-trefoil⁵, a TIM barrel²⁶, a β-propeller^6,19, an (HhH)₂-Fold¹⁰, a ferredoxin⁷, and small β-barrel proteins^12,23. Only in the case of the TIM barrel were disulfide bonds required for stabilization²⁶. The immediate driving force for duplication and fusion of an oligomerizing peptide was likely to alleviate folding constraints⁶, consistent with fully symmetric forms being foldable and capable of a high degree of thermostability. Once fused, subdomains could evolve with greater independence, allowing for asymmetric functionalizing mutations. In addition, fused peptides would be able to explore destabilizing mutations more readily, as the effect of each mutation is not amplified by its occurrence in all protomers of an oligomer²⁷.

Although fully symmetric and fold-forming symmetrically oligomerizing peptides have been shown to be biochemically active, these activities are centered on binding of sugars (β-trefoil¹⁸, β-propeller⁶), binding to and potentially coacervating with nucleic acids (the (HhH)₂-fold¹⁰, the double-ψ β-barrel²³), and shuttling electrons (ferredoxin²²). Thus, while symmetric folds occupy diverse catalytic roles, symmetrization studies frequently select folds with primarily binding roles, or obviate known catalytic activities in the process.

Asymmetric folds are the frontier of fold simplification studies

While rotationally symmetric structures provide a blueprint for their own simplification, the primitive states for folds that lack rotational symmetry are significantly harder to infer. In these cases, the location of key functional elements, intuition about the length limits of folding, and fragment conservation are often taken as a guide. To our knowledge, the early evolution of just three asymmetric folds – P-Loop NTPases, the synthetic domain of class I aminoacyl-tRNA synthetases (AARSs) (a member of the larger HUP lineage of proteins²⁸), and the synthetic domain of class II AARSs – has been studied in detail.

Early attempts to localize the functional elements within two P-Loop NTPase families, adenylate kinase²⁹ and the ATP synthase β-subunit³⁰, provided the first clues about how this ancient and ubiquitous protein lineage may have emerged. The ~44-50 residue extracted peptides included the first βαβ motif of the P-loop NTPase fold and contained the highly conserved phosphate-binding loop. Biophysical characterization indicated that the peptides were partially structured and, remarkably, could bind ATP and ADP. Thus, a full P-loop NTPase domain is not required for simple nucleotide binding.

More recently, the Tawfik group has demonstrated that P-Loop NTPase prototypes based on a Rosetta-designed, idealized Rossmann fold³¹ could also bind to nucleotides³², as well as mediate helicase-like activity³³ and weakly catalyze phosphoryl transfer³⁴. The weak catalytic activity of P-Loop NTPase prototypes suggests that the primal role of the Walker A motif was phosphate binding, with robust catalysis emerging later, in the context of an elaborated fold. This view is consistent with the observation that the longest P-Loop NTPase construct (a βαβα peptide) was the most efficient at phosphoryl transfer³⁴ but shorter constructs could bind nucleic acid ligands with equivalent or higher affinities. In each case, the prototypes were found to adopt secondary structures in solution, and several constructs formed oligomers. Despite their encouraging properties, P-Loop NTPase prototypes are difficult to work with: First, they are prone to precipitation, with shorter constructs requiring high concentrations (~1 M) of imidazole or arginine to retain solubility in stock solutions^32–34. Second, some P-Loop NTPase prototypes were found to have different properties between purifications, suggesting that fairly subtle factors can influence their conformation and oligomerization state³².

Carter and coworkers hypothesized that the catalytic core of the class I and II AARSs may have been a primordial fragment around which the larger fold condensed^35,36. To test this hypothesis, the catalytic cores of TrpRS, HisRS, and LeuRS were extracted and characterized as independent peptides. The researchers envisioned two hypothetical stages: A fold-seeding protozyme³⁶ of ~50 amino acids based on TrpRS or HisRS and a more elaborated urzyme³⁵ of ~130 amino acids based on TrpRS, HisRS, or LeuRS. We will focus our discussion on class I AARS constructs, which were derived from TrpRS³⁵ and LeuRS³⁷, due to space limitations. Though the selection of TrpRS may seem an unlikely choice, as Trp is surely a later addition to the genetic code³⁸, phylogenetic trees of the class I AARS synthetic domain support this selection, as the clade comprising TrpRS and TyrRS synthetic domains are consistently the earliest split in the tree³⁹. In each case, the protozyme constructs are βαβ peptides containing the highly conserved HIGH motif (TIGN in TrpRS) at the tip of the α-helix³⁶, which is essential for amino acid activation³⁵. The larger urzyme constructs are βαβαβαβ proteins that also include the highly conserved KMSKS motif³⁵.

Experimental characterization of AARS protozymes reveals that they are not well behaved. These peptides do not adopt a well-defined structure in solution and have poor solubility⁴⁰, often requiring maltose binding protein tags for experimental characterization. The longer urzyme constructs, however, have detectable α-helix secondary structure, may be monomeric⁴¹, and can interact with RNA⁴². While urzymes can weakly catalyze pyrophosphate exchange, protozyme activity – either pyrophosphate exchange³⁶ or pyrophosphate release – is closer to background levels. As such, there is debate as to the extent to which AARS protozyme constructs are meaningfully active⁴⁰.

Our intention in noting the limitations of current research is to emphasize the difficulty of studying the early evolution of asymmetric proteins, where folding and solubility is often compromised. Complicating matters further, these studies often focus on catalytic activities, which have proved significantly more difficult to achieve than ligand binding by minimized folds. We are optimistic that new design approaches such as generative machine learning models (Box 2) will drive substantial progress in the minimization of asymmetric protein folds.

Box 2. Can we learn about evolution from design?

Artificial intelligence (AI), such as large language models (LLMs), makes it possible to create thousands of peptide and protein sequences seeded from a common source. Thus, AI allows for a new perspective on protein engineering and evolution, in which researchers utilize artificially-generated designs to test hypotheses about the stability and function of fundamental historical motifs. In this way, we can learn about early fold evolution directly through the predictive design of novel sequences and structures, potentially unrelated to the protein families found on Earth. For example, André and colleagues used deep learning models to generate three stable folds from a restricted alphabet of only 10 amino acid types and lacking all basic and aromatic amino acids⁵⁴. They further demonstrated through simulation that the stability of these minimalist proteins was as robust to mutation as proteins containing a full amino acid alphabet.

Design methods can also be used to optimize other biophysical properties, including aggregation⁸⁵, solubility⁸⁶, and ligand/cofactor binding affinity⁸⁶. In fact, there has been immense effort in designing computational tools that strategically combine physics-based modeling, deep learning, and/or structural predictions to design structures with optimized properties^85,86. Designed sequences/structures from these tools can thus provide us key insights into how specific properties could have been selected for in small, primitive peptides, for example, through the analysis of large numbers of predicted peptides.

Finally, despite the prevalence of microproteins throughout the proteome⁸⁷, correspondingly little is known about their biology and structure. In particular, their smaller size and difficulty in traditional experimental characterization causes structure prediction programs to fall short in predicting accurate and stable structures⁸⁸. There have, however, been advances in this space, including PEP-FOLD2⁸⁹, which utilizes a coarse-grained energy force field to predict canonical folding of short peptides without relying solely on experimental structural information⁸⁹. Further, Veluri et al., conducted an extensive computational study on miniproteins, in particular noting the prevalence of structural motifs in miniproteins that are found in larger more well-studied protein scaffolds, supporting the idea that combining and iterating on miniprotein sequences could be a useful strategy for predicting tentative ancient structures based on conserved structural motifs⁹⁰. Thus, there has been substantial progress in computational design of miniproteins, and this provides an exciting frontier for the field.

Figure B2. Artificial intelligence-supported evolutionary hypothesis testing.

Artificial intelligence (AI) generated sequences can support the design of putative fold-founding fragments, allowing hypotheses about the origins of asymmetric protein folds to be studied.

Can a short peptide seed the emergence of an entire fold?

The asymmetric fold minimization studies above attempt to achieve some aspect of structural and functional continuity between a secondary structure motif-sized peptide and a complete fold. Could the first folds have been seeded by even smaller peptides? For ferredoxins (a rotationally symmetric fold), there is evidence to support this possibility. Leslie Dutton first demonstrated that a simple peptide “maquette” based on the iron-sulfur cluster-binding motif of ferredoxins (CxxCxxC) could accommodate a redox-active iron-sulfur cluster⁴³. Subsequent studies from the Nanda and Falkowski groups systematically simplified the ferredoxin fold, demonstrating a fully symmetric form²² and a homo-oligomerizing form⁷, as well as a heterochiral variant of the maquette peptide⁴⁴. The symmetric form could act as a functional electron shuttle in vivo²². Taken together, these studies present a plausible trajectory from a short, simple peptide to a ubiquitous, contemporary fold.

Similar studies demonstrated that a Walker A motif-like hexapeptide (derived from p21 ras) could bind phosphate⁴⁵. Although the structure was unknown, it was hypothesized to adopt the nest-like conformation observed in contemporary folds. To test this hypothesis, we performed a detailed structural characterization by molecular dynamics, but did not find evidence of nest-like conformations⁴⁶. Moreover, we found that shuffled sequences of the Walker A motif interacted with phosphate and nucleotides to equivalent degrees, suggesting that the Walker A motif may not have special properties as a free peptide in aqueous solution. Our results highlight the difficulty of interpreting profound sequence conservation, which may reflect eroded fossils of a more elaborate primitive state or true seeding peptides that were functional in an as-yet unknown environmental context⁴⁷. This concern similarly extends to the simplification of asymmetric folds by extracting core conserved fragments, generating constructs with hydrophobic faces that can promote oligomerization and precipitation that may or may not have been historically relevant.

Diverse folds can be built from restricted amino acid alphabets

A core question in the field of protein simplification has been whether restricted amino acid alphabets can encode foldable proteins. Primitive life may have used fewer amino acid types to build proteins than contemporary life (Box 1), such as an absence or scarcity of basic and aromatic amino acids. In addition, probing the limits of folding may report on the discoverability of various protein folds, perhaps providing insight into the nature of the first folds. In Box 3, we provide a discussion of what the first folds may have been. Two strains of alphabet simplification are common: The enrichment of a single type of amino acid (enrichment studies) and the restriction of the amino acid alphabet to 7-14 amino acid types (restriction studies).

Box 3. Which protein class came first?

There are four primary structure classes: all-α, all-β, α/β, and α+β. All-α and all-β proteins are composed primarily of α-helices or β-sheets, respectively. α/β and α+β proteins have both α-helices and β-sheets, with these two structural elements alternating along the chain in α/β proteins while being segregated in α+β. Which protein class, and which fold within that class, likely arose first?

Patterns of fragment reuse across the protein universe reveal that α/β proteins are highly interconnected^76,91, with ostensibly unrelated folds sharing subdomain-sized fragments. Notable examples of fragment sharing between α/β proteins include the TIM-barrel and flavodoxin⁷⁵ and the Rossmann fold and P-loop NTPases⁷⁷. α/β proteins dominate metabolism^92,93, likely due to their close association with nucleotide cofactors,⁹⁴ and protein fold evolution viewed from the perspective of metabolism favor their early emergence^95,96. Simplification studies have demonstrated that structural elements based on the catalytic core of P-loop NTPases can oligomerize and exert various biochemical functions^32,33, including weak phosphoryl transfer activity³⁴. Studies of simple ‘urzymes’ derived from the catalytic core of the tryptophanyl aminoacyl-tRNA synthetase catalytic domain – an α/β domain classified as part of the HUP evolutionary lineage²⁸ – report aminoacyl transfer activity³⁶. Short prototype folds with Rossmann- and P-loop NTPase-like topologies have also been successfully created^31,97, as well as sequence-simplified variants of PRTase-like that were shown to be catalytically active, even without C, H, I, M, N, Q, and W⁵¹.

An analysis of ribosome-associated proteins, however, provides a different perspective, favoring instead the early emergence of all-β structures^12,23,98. Amino acid simplified variants of small β-barrel proteins (i.e., SH3, OB-fold, and cradle loop barrel) suggest a low information demand and provided the first experimental demonstration that these folds can readily interconvert¹², validating earlier predictions⁷⁴ and supporting the metafold concept⁹⁹. The different conclusions reached by ribosome-centric and metabolism-centric perspectives may be the result of different emergence filters, with metabolic folds shaped by the requirement for independent foldability and ribosome-associated-folds shaped by the need to bind to rRNA. Finally, binding to metals or metal clusters can induce structure in short and simple peptides⁴³, serving as a stable core around which further elaborations can accrete¹⁰⁰. The early emergence of ferredoxins, which adopt an α+β fold, may have been seeded by such interactions⁷.

In our view, multiple distinct forms (Figure B2) likely served as entry points to folded protein structures.

Figure B3. Evolutionary seeds of the contemporary protein universe.

Fold emergence was likely seeded by interactions with RNA, cofactors, and metals, with each binding partner potentially favoring the emergence of different fold classes.

Enrichment studies seek to reveal the determinants of protein structure and folding. For example, mutation of 34% of core residues to Met in T4 lysozyme⁴⁸ or 47% of all residues to Ala in a homeodomain⁴⁹ were achieved without a loss of foldability, highlighting the surprisingly plastic nature of the protein core and confirming that only a subset of positions are essential for defining a fold. More recently, this approach was applied to a simple, Rosetta-designed Rossmann-like fold³¹ by the enrichment of Val residues⁴. Despite having 30 Val residues (>32% of the sequence), the resulting protein was both foldable and thermostable, with a T_m of 106.1 °C. How did this protein retain fold-defining features in the absence of specific packing interactions? Loop lengths turned out to be a key design feature that constrains and defines fold topology.

Alphabet restriction studies have been applied to diverse folds (Figure 1C), including the SH3 (14 AA types⁵⁰), PRTase-like (13 AA types⁵¹), chorismate mutase II (9 AA types⁵²), β-trefoil (12 AA types⁸), α-β plait (10 AA types⁵³), (HhH)₂-fold (9⁹ or 10^10,21 AA types), double-ψ β-barrel (7 AA types²³), HTH (10 AA types^11,54), ferredoxin (10 AA types⁵⁴), R3H domain-like (13 types⁵⁵), and ribonuclease H-like folds (10 types⁵⁴). While earlier studies focused more on reduction of the alphabet size, later studies would emphasize prebiotic availability when selecting an alphabet in an effort to better reflect primitive biological constraints. Very recently, a deep learning-based protein sequence design approach was applied to the question of sequence simplification⁵⁴. In this work, three folds – HTH, ribonuclease H-like, and ferredoxin – were designed using an amino acid alphabet entirely devoid of non-prebiotic amino acids (Box 1).

Taken together, these studies demonstrate that the interactions mediated by aromatic amino acids (π-π/cation-π interactions), basic amino acids (salt bridges with acidic amino acids), and sulfur-containing amino acids (disulfide bonds) can be excluded from a protein, often in combination, without a total loss of foldability. Even proteins that are devoid of charged amino acid side-chains can retain foldability⁵⁶. These results highlight the primary role of secondary structure formation⁵⁷ and hydrophobic-polar patterning⁵⁸ in defining protein structure. As we have noted previously³, the prebiotic amino acid alphabet (Box 1) includes amino acids with a high propensity to form turns (Gly, Pro), α-helices (Ala, Glu, Leu), and β-sheets (Ile, Thr, Val). This set of amino acids also has both hydrophobic (Ala, Ile, Leu, Val) and polar/charged (Asp, Glu, Ser, Thr) amino acid types.

To retain stability even with a restricted amino acid alphabet, two common approaches have been reported: Increasing the net proportion of hydrophobic residues⁵⁴ and/or optimizing for secondary structure propensities^5,8. We also note that non-canonical basic amino acids such as ornithine, which may have been part of a primitive genetic code (Box 1), can be used in place of Arg or Lys, thereby recovering the ability to form salt bridge interactions as well as achieve functions that are frequently mediated by positively charged moieties, such as nucleic acid-binding²¹.

Like the minimization studies of symmetric and asymmetric proteins, alphabet restriction studies often lose their native activities or retain binding activities, as in the (HhH)₂-fold²¹ and the HTH fold¹¹. Notable exceptions are the work of Yokoyama and Hilvert, who selected for enzymatically active sequences from libraries of simplified or partially simplified proteins with cell-survival assays^51,52 or made modifications to inferred ancestral sequences⁵³. It should be noted, however, that the simplified folds generated from these approaches often include aromatic and basic amino acids that are apparently required for stability and/or enzymatic activity.

The successful alphabet restriction of various folds underscores the fundamental designability of contemporary folds and the redundancy of the amino acid alphabet. Given the number of folds that have been simplified by amino acid restriction to date, it is clear that primitive amino acid alphabets of approximately 10 amino acid types could have encoded foldable proteins, particularly when environmental buffering is considered. However, the fact that a fold can be designed with a restricted alphabet does not imply that it is ancient; for example, β-trefoils are likely a comparatively recent innovation. Nevertheless, we expect that the most ancient folds can be encoded by restricted alphabets. While we support the continued alphabet restriction of diverse folds, we anticipate studies of environmental buffering, fold traversability (both discussed below), and enzymatic activity will be the most fruitful topics moving forward.

Diverse environments can support the emergence of folds

Various environments have been proposed to have supported or buffered the early emergence and evolution of folded proteins^8–10,12,23, but is such compensation needed? Akanuma et al. have argued that stability can be readily achieved by restricted alphabets⁵³ (T_m = 85 °C using 10 types of amino acids including, Arg and Phe), and recent work using machine learning approaches support this view⁵⁴ (T_m > 90 °C using the 10 prebiotic amino acids). We note, however, the ability to generate stable protein forms using restricted amino acid alphabets does not imply that the first protein forms were themselves highly stable or foldable. Modern proteins (even including those from hyperthermophiles⁵⁹) are only marginally stable⁶⁰ and, depending on biological context, exist in varying degrees of disorder, ranging from fully disordered to partially disordered (folded with intrinsically disordered regions) through to fully folded scaffolds^61,62. The marginal stability of protein scaffolds is a consequence of the weak interactions that hold the folded state together and evolutionary pressure for biological function, which often requires a balance between conformational flexibility and stable folding⁶³. Furthermore, the interplay between regions with well-defined structures and regions associated with greater flexibility may be a key component to protein evolvability⁶⁴. Is it likely that, similarly to contemporary proteins⁶⁰, emergent folds were only marginally stable (e.g., ref. ⁶³) and, while stable folds can be created from simplified alphabets, supportive environments and chemical chaperones are of great interest to the fold emergence process (Figure 2), particularly if unstructured or marginally foldable peptides were a starting point for folded proteins.

Figure 2. Supportive environments for protein fold emergence.

Various environments and compounds have been proposed to have supported fold emergence, including high concentrations of salt (halophile conditions), polyamines, dications, and coacervates.

The kinds of environmental compensations needed for fold emergence depend in part on the properties of the proposed restricted amino acid alphabet. The first supportive environment identified from an alphabet restriction study was high salt (halophile) conditions⁸. In this case, a simplified β-trefoil protein comprised of just 12 amino acid types had a T_m = 34.2 °C and was only 86% folded at the temperature of maximum stability if a buffer with 0.1 M NaCl⁸. In a buffer with 2.0 M NaCl, however, the melting temperature increased dramatically, to T_m = 49.1 °C⁸. Salts are an attractive environmental buffer because they are readily available in the surrounding environment and can be concentrated during freeze-thaw and wet-dry cycles⁶⁵. Salts exert general effects on protein stability, such as charge screening and excluded volume effects (particularly at high concentrations), that can support the folding of simple proteins. We have demonstrated that high concentrations of monocations⁸ and seawater concentrations of dications⁹ can also promote the folding of acidic and hyper-acidic proteins. The potential role of dications was expanded further by the Hlouchová group, who demonstrated that they can mediate interactions between an acidic protein and the negatively charged backbone of nucleic acids¹¹, enabling nucleic-acid binding with an alphabet that lacks basic amino acids.

A second proposed compensation for a lack of basic amino acids is polyamines. Polyamines are ubiquitous in biology, play key roles in promoting the folding of RNA⁶⁶, and can promote peptide aggregation into amyloids⁶⁷. We have shown that polyamines can also promote folding of a hyper-acidic α-helical protein⁹, compensating for the lack of basic amino acids. Thus, polyamines may have served as essential chemical chaperones in primitive biology, supporting the folding of both acidic proteins and RNA, as well as promoting associations between simple peptides. Although the mechanism of polyamine-mediated stabilization is not fully understood, we hypothesize that it relates to both the positively charged amino groups – which are able link carboxylic acids on the protein surface through salt bridges – and the intervening aliphatic moieties, which can nestle into small hydrophobic pockets.

Liquid-liquid phase separation is proposed to have played a key role in the origin of life⁶⁸, serving as a mechanism to concentrate proto-biological components into membraneless compartments. The unique environment within a coacervate can also promote peptide oligomerization and folding^10,69, suggesting that the coacervate context may have supported the emergence of folded structures in general and nucleic acid-binding domains in particular^21,70. Here, too, the mechanisms of stabilization are not fully understood but likely relate to the high peptide concentration (allowing access to states with a relatively high K_D), excluded volume effects (which favor compact states), reduced water activity (which strengthens the hydrophobic effect), and conformational biases induced by interaction with RNA. Intriguingly, the coacervate context is itself a consequence of the properties of the participating peptides. Thus, peptides that are able to form coacervates, even if they are not well-folded in aqueous solution, can contribute to the creation of an environment that supports the exploration of folded states. As a consequence, coacervate-forming peptides may be primed for elaboration into more complex protein structures as the properties of the coacervate are tuned⁷¹, a scenario we refer to as “molecular self-domestication.”

Simplification studies help reveal transitions between folds

The question of fold traversability – whether a fold can, through substitutions and structural rearrangements – turn into another fold has significant implications for fold evolution. If protein sequence-structure space is highly connected, it suggests a more facile evolutionary exploration problem, where the first fold discovery events increase the chances for yet other folds to emerge⁷².

Two regions of protein sequence structure space have been noted for their apparent traversability, the small β-barrels^73,74 and α/β proteins^75–77, with the former being the quintessential example of fold traversability in evolution (Figure 3). Although sequence comparisons and fragment sharing analyses long identified these folds as being potentially interconvertible – a feature also implied by their close clustering in the embedding space of protein language models⁷⁸ (Figure 3) – an experimental demonstration of this fact was realized only recently by the Tagami group¹². They demonstrated that simple evolutionary processes, such as substitutions and circular permutations, can drive fold transitions – particularly those that involve the rearrangement of β-strands. In each case, the amino acid alphabet of the proteins under study were restricted, ranging from 7-13 amino acid types.

Figure 3. Topological transitions within the small β-barrel metafold.

t-SNE projection of embeddings from contrastive-learning sequence structure (CLSS)⁷⁸ , a contrastive learning model that co-embeds sequence and structure information. Note the close connection between the small β-barrel folds, and that their relative organization is consistent with an experimentally demonstrated trajectory. Oligomerization state, alphabet size, and the presence of basic amino acids are indicated as in Figure 1. β-strands in the topology diagrams are labeled according to SH3 to emphasize homology. PDB accession codes from left to right are 8JVZ (Src homology 3, SH3)¹², 8JVS (oligonucleotide-binding fold, OB-fold)¹², 8JVO (double-zeta β-barrel, DZBB)¹², and 7DXZ (cradle loop barrel, CLB)²³.

The rules that underlie fold transitions are not yet fully understood, as both the extent to which this process occurred in primitive biology and the degree to which this process relies on fold-specific idiosyncrasies remain unknown. Transitions among β-barrels¹² and topological diversity within P-Loop NTPases⁷⁷, however, suggest an inherent ability for β-sheets to rearrange, particularly when other fold-discriminating features such as loop lengths⁴ are relaxed. The propensity for β-sheets to form amyloids may be another reflection of this property. We hypothesize that amino acid restriction can promote fold switching by loosening the specificity for any given fold. Indeed, as alphabet size decreases, interaction modes and packing arrangements are lost, meaning that the repertoire of fold-specifying features must also decrease – potentially opening up pathways between folds through decreased energetic discrimination. If correct, this hypothesis implies that primitive biological systems were capable of a greater degree of fold exploration via transitions than contemporary ones, where greater specification is possible.

Concluding Remarks

Protein simplification studies provide evidence-based demonstrations of the limits of protein folding, viable structural and functional transitions, and folding-supportive environments. Taken together, these studies have yielded key insights into how the universe of folded proteins could have emerged despite constraints on amino acid availability and peptide length. Nevertheless, significant open questions remain (Box 4: Outstanding Questions). These include defining the principles that dictate the early evolution of asymmetric proteins, understanding the limits of evolutionary continuity and functional innovation, understanding how interconnected protein folds are in sequence space, understanding the role of environment in supporting evolutionary innovation, and understanding how much simplicity can still support a rudimentary metabolic or signaling pathway. We expect that progress on many of these challenges – including the simplification of asymmetric protein folds, the incorporation of complex biochemical function within simplified scaffolds, and the bridging of diverse fold types – will be accelerated by developments in machine learning methods, enriching our understanding of the chemical drivers of early protein evolution. However, as these methods become more sophisticated, the historical relevance of their results may decrease, with low-probability transitions become increasingly discoverable by advanced design methodologies, making it harder to infer historical reality. Overall, it is an exciting time to study early protein evolution as the answers to these questions shift into focus.

Box 4. Outstanding Questions Box.

• What principles dictate the early structural evolution of asymmetric proteins?

• What are the limits of evolutionary continuity? To what extent were modern proteins shaped by the early success of short (5-10 residue), functional peptides?

• Primitive proteins excel at binding – to metals, FeS clusters, nucleotides, and nucleic acids. What are the limits of enzyme catalysis, and can we construct robustly active simple catalysts?

• How traversable is the protein universe? What folds are connected to each other and what is the directionality of their evolutionary relationship? Which of these connections have likely been traversed by life on Earth?

• As design approaches rapidly advance, how do we assess what is historical versus what is possible but did not happen on Earth, even if we can realize it experimentally?

• What other environments could have supported the emergence of primitive folds? Specifically, what role might mineral surfaces or lipids have played?

• Can a simple biochemical pathway or system be realized with simplified proteins?

Highlights.

• Protein simplification studies identify folding-supportive environments, reveal connections between folds, and set evidence-based limits for fold emergence.

• Symmetric folds can be formed by the homo-oligomerization of a 28-74-residue peptide. Simplification of ancient folds that lack rotational symmetry is still an unmet challenge.

• Diverse folds can be constructed with alphabets of just 7-14 amino acid types, including alphabets that lack aromatic amino acids and/or basic amino acids.

• Environment buffering – for example, by salts, chemical chaperones such as polyamines, or coacervates – could have supported the emergence of complex folds.

• Fold space is traversable, particularly among the small β-barrels and α/β folds.

Glossary

Accretion

A process of protein growth by gradual buildup, for example, by lengthening of secondary structure elements, insertion of loops, and subdomain duplications

Class

A structural classification based on the composition and organization of secondary structure elements within a protein. α-proteins are comprised primarily of α-helices. β-proteins are comprised primarily of β-strands. α+β proteins are organized into segregated regions of all-α and all-β structure elements. α/β proteins are comprised of alternating α and β elements

Circular permutation

A type of protein sequence re-arrangement in which the N- and C-termini are moved to a new position within the sequence, and the former termini are joined

Disorder

Full or partial regions of a protein that lack a stable three-dimensional conformation, existing instead as dynamic ensembles of extended conformations. Note that the conformations sampled by disordered regions can be biased and that these regions can be key contributors to biochemical function. Under specific conditions (e.g., binding to a ligand) disordered regions can adopt a well-defined conformation

Domain

A protein structural unit, often between 75-150 amino acids in length. Although the definition of a domain is not fully agreed upon, and sources can differ in their assignments, domains are usually associated with three key properties: independently foldability, or the potential to do so; reuse within multiple structural and evolutionary contexts; and a key functional role, for example, cofactor binding. Proteins can be composed of one or more domains

Embedding space

A learned latent space in which inputs (e.g., protein sequences or structures) are transformed into high-dimensional vectors. Within a well-trained embedding space, similar inputs are positioned close to one another, while dissimilar inputs are far apart, enabling efficient representation of semantic relationships. For visualization, embedding spaces, which are usually high dimensional, can be projected into two dimensions, a process that involves some distortion of the high-dimensional relationships

Family

A group of evolutionarily related proteins with similar sequences, structures, and biochemical functions. Protein families are often associated with gene families. Note that protein families can be related to each other; for example, polyphosphate kinase 2 (PPK2) is a protein family within the larger P-Loop NTPase protein lineage

Foldability

The ability of a specific linear chain of amino acids to assume a well-defined ensemble of compact three-dimensional conformations characterized by significant secondary structure formation. When referring to an amino acid alphabet, foldability is the ability of that alphabet to encode foldable proteins

Metafold

A group of related but topologically distinct protein folds. The small β-barrel folds are the archetypical example of a metafold

Microprotein (miP)

A small protein, usually between 10-100 amino acids in length, translated from a small open reading frame (smORF), typically located within intergenic regions. Microproteins currently form a subset of the largely unexplored “dark proteome”. This term is typically used in the fields of genomics and proteomics, and is intended to avoid confusion with fragments broken off from larger proteins

Miniprotein

A small protein, usually between 50-100 amino acids in length, that folds into a single domain. Miniproteins can either occur naturally (e.g., scorpion toxins) or be designed. This term is typically used in the fields of synthetic biology and protein evolution, and is intended to avoid confusion with fragments broken off from larger proteins

Oligomerization

The process by which two or more protein molecules associate to form a higher-order complex with a well-defined conformation. Oligomerization is typically a reversible process mediated by weak, non-covalent interactions, and the resulting complexes usually retain solubility

Prebiotic

A chemical compound or environmental condition that could have plausibly existed on Earth prior to the origin of life

Protein fold

A characteristic three-dimensional conformation adopted by a linear chain of amino acids. A protein fold is a structural classification; although proteins with similar sequences usually adopt the same protein fold, unrelated proteins may also adopt the same protein fold through convergent evolution

Traversibility

The ability of two folds to be connected by intermediate sequences that retain foldability via known evolutionary mechanisms, such as amino acid substitution, circular permutation, and fragment duplication