Introduction

Most proteins produced by extant organisms are composed of combinations of 20 amino acids with diverse physicochemical properties. The possible combinations of amino acid sequences—and thus their physicochemical characteristics—are virtually infinite. Among them, sequences that are compatible with living organisms have been selectively retained through the course of evolution ¹. This compatibility is determined by a balance between harms and benefits to biological functions, where the harms include various factors that decrease fitness, such as cytotoxicity or excessive consumption of cellular resources ².

One effective way to understand the compatibility of amino acid sequences is to identify patterns commonly present in amino acid sequences (natural sequences) used by extant organisms. With the vast amount of protein sequence data now available, attempts have been made to extract features of natural sequences using artificial intelligence and to apply them to three-dimensional structure prediction and de novo protein design ^3–6. Although this approach has achieved success in extracting, predicting, and designing functional sequences, it does not directly reveal the compatibility of sequences that do not exist in nature, nor the characteristics of sequences that are not compatible due to harmful effects. “Nullpeptides,” which are peptides absent from the proteome of a given organism, may represent sequences that have been eliminated during evolution ^7,8. However, such inferences still depend on the limited sequence space of natural proteins, and thus have inherent limitations in capturing the full landscape of non-compatible sequences.

Another approach to investigating sequence compatibility is the synthetic exploration of compatible combinations among amino acid sequences. By constructing a comprehensive sequence library and performing functional screening, it is possible to obtain functional sequences that do not exist in nature ^9–11. Furthermore, the use of Deep Mutational Scanning (DMS) enables systematic and quantitative evaluation of sequences within the library, allowing the identification of both the diversity and the characteristics of compatible sequences ^12,13. The resulting data can also serve as training datasets for machine learning, and with sufficient data density, highly accurate predictive models can be developed ^14,15. Neme et al. demonstrated through DMS in Escherichia coli that, while many random sequences are harmful, a considerable number of beneficial sequences also exist ^16,17. However, this exploration still covers only a limited portion of the vast sequence space, and the specific features and physiological effects of harmful or beneficial sequences remain largely unknown.

Therefore, in this study, we aimed to extract harmful and beneficial features from extremely simple model sequences. Of particular interest are continuous stretches of identical amino acids, known as amino acid homorepeats ^18–24 or PolyX. PolyX sequences are also referred to as homopolymeric amino acid stretches/repeats ^25–29, amino acid runs ³⁰, or homopeptide repeats ³¹, and they are present in a wide range of natural proteins. PolyX motifs exhibit distinct physiological functions and have been exploited in various biotechnological applications ^19,25,32. For example, PolyP (∼10 residues) segments of formin act as nucleation sites for actin filament formation ³³, and the arginine-rich region of protamine containing PolyR tracts (∼6 residues) binds strongly to DNA, inducing chromatin condensation ³⁴. Translation of the poly(A) tail resulting from termination failure produces PolyK, which serves as a signal for protein degradation ^35,36. The metal-chelating property of PolyH (∼6 residues) is utilized as a purification tag ³⁷, while the membrane-penetrating ability of PolyR (∼8 residues) is exploited for molecular delivery ³⁸. Conversely, PolyX sequences can also exert harmful effects and are associated with diseases. Abnormal expansion of PolyQ beyond ∼40 residues leads to aggregate formation and neurodegenerative disorders ³⁹. Positively charged PolyR and PolyK (∼10 residues) interact with the negatively charged ribosomal tunnel, causing translational stalling ⁴⁰, and PolyW (∼10 residues) induces ribosome stalling that prevents CAT-tail addition ⁴¹. Negatively charged PolyD and PolyE (∼10 residues) destabilize ribosomes and induce translational arrest ^42,43.In natural protein sequences, both the frequency and the length of PolyX motifs show amino acid–specific biases ^18,29,31,44, which are thought to reflect a balance between harmful and beneficial effects. PolyX sequences are considered evolutionary hotspots, as they can readily arise through mutations ^29,45–47. When present in natural proteins, they may confer benefits, whereas their absence could indicate harms that led to their elimination. Although the cytotoxicity of certain PolyX sequences has been examined in mammalian cells ^26,27,48, general trends that transcend species and expression conditions remain unexplored.

In this study, we systematically evaluated the harmful and beneficial effects of peptide sequences composed of ten identical amino acids (Poly₁₀X) by individually overexpressing them in yeast. A length of ten residues was chosen as it is sufficiently long to manifest amino-acid-specific physicochemical effects, such as charge accumulation or hydrophobic interactions, while remaining short enough to minimize sequence context and secondary-structure complexity. By fixing sequence length and repeating a single residue, this approach allows direct comparison of how intrinsic physicochemical properties—such as charge and hydrophobicity—affect cellular tolerance. Moreover, comparison with PolyX occurrence in natural proteomes enables us to infer which sequence features are favored or disfavored by evolutionary selection, thereby testing whether their harmful or beneficial effects contribute to determining their abundance in extant proteomes.

Results

The harmful effects of each Poly₁₀X are generally conserved across species

In this study, we began by investigating the expression limits of enhanced green fluorescent protein (EGFP) fused at its C-terminus with 20 different Poly₁₀X peptides (Figure 1A, S1, and S2). The fusion was made at the C-terminus rather than the N-terminus to minimize the influence of N-terminal sequences, which are known to strongly affect translation efficiency ⁴⁹. For harmfulness assessment in yeast, we expressed EGFP and EGFP-Poly₁₀X from the strong constitutive TDH3 promoter (TDH3_pro) and increased plasmid copy numbers using the genetic tug-of-war (gTOW) method ^50,51. During this process, we measured both growth inhibition and EGFP expression levels (cellular fluorescence). In the gTOW assay, plasmid copy number can be increased by culturing cells with or without leucine supplementation. Under both conditions, we evaluated cell growth and fluorescence to calculate the maximum growth rate (MGR) and maximum fluorescence intensity (MFI). To enable one-dimensional comparison of the harmfulness of each Poly₁₀X variant, we defined relative neutrality based on growth and expression data as follows: Relative neutrality = %(MGR_Poly₁₀X / MGR_Δ) × %(MFI_Poly₁₀X / MFI_Δ). Where, Δ refers to the control EGFP without any Poly₁₀X. Thus, the relative neutrality values were normalized such that the control (Δ) equals 10,000 (Figure 1A) .

Figure 1.

As a result, the growth rate and fluorescence intensity varied greatly depending on the type of expressed EGFP-Poly₁₀X (Figure 1B). Under leucine-depleted conditions, strong growth inhibition was observed for about half of the Poly₁₀X variants, and yeast cells showed almost no growth. The harmfulness of each Poly₁₀X was quantified as relative neutrality (Figure 1C), revealing that the degree of harmfulness differed substantially among amino acids. The addition of some Poly₁₀X sequences markedly reduced the neutrality of EGFP (increasing harmful effects), whereas the addition of other Poly₁₀X sequences resulted in neutrality comparable to, or even higher than, that of EGFP.

Next, to examine whether the neutrality of Poly₁₀X is conserved across species, we performed equivalent experiments in Escherichia coli. Since the gTOW method cannot be applied to E. coli, we constructed multicopy plasmids expressing EGFP-Poly₁₀X under the control of the lac promoter (lac_pro) and induced their expression with isopropyl β-D-1-thiogalactopyranoside (IPTG) (Figure S3). The resulting data were also converted into relative neutrality values (Figures 1D and 1E). In E. coli, although the expression levels of EGFP-Poly₁₀X varied considerably among the different Poly₁₀X variants, the effects on growth were limited (Figure 1D). This is likely due to the weaker selection pressure on plasmid maintenance compared with the yeast gTOW system—plasmids expressing highly harmful Poly₁₀X variants are thought to be lost from the population during cultivation (Figure S3E). Previously, the cytotoxicity of PolyX (∼30 residues) in mammalian cells was evaluated by Oma et al. ²⁶. They assessed PolyX cytotoxicity in African green monkey kidney–derived COS-7 cells by strongly expressing YFP-PolyX under the CMV promoter and measuring cell viability. Because numeric cell viability data were not provided, we relied on reported statistical significance and converted it into four ranks (1–4 corresponding to p < 0.001, p < 0.01, p < 0.05, and p > 0.05, respectively). We then compared the Poly₁₀X neutrality data obtained in this study for yeast and E. coli with those reported by Oma et al. for mammalian cells. The results showed a generally high correlation among the three systems (Spearman’s correlation ρ > 0.6; Figures 1F-H), indicating that Poly₁₀X harmfulness exhibits a conserved trend across species.

To further refine the analysis, we evaluated the neutrality of highly harmful Poly₁₀X variants by using the WTC₈₄₆ promoter, which allows tunable expression ⁵². By gradually modulating expression levels, we assessed the neutrality of each Poly₁₀X variant (Figure S4). The neutrality values (Figure 1I) showed a stronger correlation with mammalian cell cytotoxicity (ρ = 0.72, Figure 1J), thereby reinforcing the conclusion that Poly₁₀X harmful exhibits a conserved trend across species. Figure 1K compares the neutrality of EGFP-Poly₁₀X variants in yeast across a range of expression strengths. Some amino acids were consistently harmful across all expression levels, whereas others were more neutral than EGFP (i.e., beneficial) at specific expression ranges.

Poly₁₀X harmful and beneficial effects are associated with amino acid polarity and hydrophobicity

Next, to examine whether the harmful or beneficial effects of Poly₁₀X depend on the physicochemical properties of the fused protein, we fused Poly₁₀X to the C-terminus of six different fluorescent proteins and evaluated their neutrality in the same manner (Figure 2A, S5–S9). The obtained neutrality values showed similar trends, with correlations above 0.85 between datasets (Figure 2C). Based on this, we analyzed the relationship between amino acid biochemical and physicochemical parameters and neutrality. Correlation analysis with 28 amino acid indices ^2,53 (Figure S10A) revealed that biosynthetic energy cost showed the highest correlation (ρ = –0.76, p = 1.8e-23), followed by polar requirement (ρ = 0.66, p = 2.2e-16) and hydrophobicity (ρ = –0.60, p = 4.0e-13) (Figure 2D). Because these parameters mutually showed high intercorrelation (Figure S10B), neutrality may be largely explained by these characteristics.

Figure 2.

To further characterize the nature of Poly₁₀X, we performed cluster analysis and classified amino acids into three groups (Figure 2E). Cluster 1 included amino acids that increased the relative neutrality above 10,000 for most fluorescent proteins when added to the C-terminus, indicating beneficial effects that reduced protein cytotoxicity. Cluster 2 consisted of amino acids that exhibited harmful effects regardless of the fluorescent protein to which they were attached. Cluster 3 contained amino acids whose effects varied depending on the fused fluorescent protein. Notably, we did not detect any beneficial amino acids in Gamillus, probably because the expression level was insufficient to cause overexpression-dependent growth inhibition, even at maximum induction (Figure S7). Figure 2F shows the representative biochemical and physicochemical properties of amino acids in each cluster. Cluster 1 included E, S, N, and Q, which are low-cost, have high polar requirements, and low hydrophobicity. Cluster 2 mainly contains amino acids with high cost, low polarity, and high hydrophobicity (C, P, M, V, F, I, W, L, Y), but also includes positively charged ones (K, R). Cluster 3 contains G, T, H, and A, which exhibit intermediate cost, polar requirement, and hydrophobicity, while D is somewhat exceptional.

To examine whether the strong harmfulness of the four Poly₁₀X variants (F, I, W, and Y) was caused by the high amino acid synthesis cost, we re-evaluated the harmfulness of EGFP (Δ) and four EGFP–Poly₁₀X constructs after supplementing the medium with excess corresponding amino acids. As a result, in all cases, the harmfulness was not alleviated (Figure 2G and S11), suggesting that amino acid depletion due to high biosynthetic cost was not the cause of harmfulness. As mentioned above, since the biosynthetic cost correlates well with polarity and hydrophobicity, the apparent correlation with synthesis cost likely reflects a pseudo-correlation arising from these physicochemical properties.

Structural context modulates the effect of Poly₁₀X, while its overall neutrality trend is conserved

So far, we have examined the effects of C-terminally fused Poly₁₀X. Next, to investigate the effects of Poly₁₀X in different structural contexts, we constructed three types of variants (Figures 3A and S12–14): Poly₁₀X was inserted between two fluorescent proteins (EGFP and mCherry); Poly₁₀X was inserted into an internal loop of EGFP; and the C-terminally fused Poly₁₀X was detached from the fluorescent protein by a P2A sequence ⁵⁴. These designs allowed us to test whether the effects of Poly₁₀X depend on the presence of a downstream protein domain, its position within a protein, or its physical coupling to the fluorescent protein. As a result, all constructs showed distinct growth rates, fluorescence levels, and neutrality indices depending on the Poly₁₀X sequence (Figures 3B–G and S12–14). It should be noted that in the detached construct, the fluorescence intensity reflects the amount of free EGFP released by the P2A cleavage, but not the amount of the Poly₁₀X peptide itself. In contrast, internal Poly₁₀X insertions may influence not only protein neutrality but also the fluorescence properties of EGFP, meaning that fluorescence does not necessarily reflect expression levels directly. The neutrality indices among the three structural contexts and the C-terminal EGFP–Poly₁₀X construct showed rank correlations above 0.74 (p = 2.2e-4) (Figure 3H), indicating that Poly₁₀X exhibits a generally conserved trend in neutrality regardless of structural context. Among them, K and P showed markedly reduced harmfulness when placed between two fluorescent proteins compared with when fused to the C-terminus (Figure 3I, upper left). I, V, and W maintained strong harmfulness even in the detached Poly₁₀X construct (Figure 3I, lower right), suggesting that the Poly₁₀X peptides themselves can exert harmfulness independently. Conversely, the significant harmfulness of C and H (Figure 1C) were lost in the detached Poly₁₀X construct (Figure 3G), indicating that these Poly₁₀X sequences exhibit harmfulness only when fused to EGFP. No beneficial effect of Poly₁₀X was observed when the Poly₁₀X were detached (Figure 3G). Therefore, the beneficial effects might arise from mitigating the overexpression-induced harmfulness of fluorescent proteins.

Figure 3.

Poly₁₀X induces protein relocalization and aggregate formation

Next, to investigate the effects of Poly₁₀X on cellular functions, we observed the intracellular localization of EGFP–Poly₁₀X using fluorescence microscopy (Figures 4A and S15–S16). As a result, highly harmful and hydrophobic amino acids—W, I, Y, V, L, and F, except for M—were observed as intracellular aggregates. The appearance of these aggregates varied among amino acids, suggesting that they form distinct types of intracellular assemblies. D and C were enriched in the nucleus, whereas R formed a single bright punctum. Other amino acids showed diffuse cytoplasmic localization similar to the control. These results suggest that highly harmful Poly₁₀X sequences exert their harmfulness through aggregation propensity and interactions with cellular structures. Notably, these localization patterns were largely consistent with the subcellular localization of YFP–PolyX observed in COS-7 cells by Oma et al. ²⁷, indicating that the intracellular behavior of PolyX is conserved among eukaryotic cells.

Figure 4.

We also examined the localization of constructs in which Poly₁₀X was inserted between EGFP and mCherry (Figures S17–S19). For most amino acids, EGFP and mCherry showed identical localization patterns, consistent with those observed for EGFP–Poly₁₀X. In contrast, in the Poly₁₀K and Poly₁₀P constructs, EGFP localized to the cell surface, while mCherry fluorescence was barely detectable (Figure 4B). Because such behavior was not observed in EGFP–Poly₁₀K or EGFP–Poly₁₀P (Figure 4A), some Poly₁₀X can exhibit distinct behaviors depending on the structural context.

Poly₁₀E reduces the harmful effects of protein overexpression through aggregation suppression

As described above, beneficial Poly₁₀X sequences are thought to alleviate the harmful effects caused by the overexpression of fluorescent proteins. Overexpressed fluorescent proteins can exert toxicity through multiple mechanisms ⁵⁵. In the case of EGFP, overexpression leads to aggregate formation due to misfolding and imposes a burden on proteostasis ⁵⁶. Since these adverse effects become more pronounced at high temperatures, we evaluated the neutrality of EGFP-Poly₁₀X at 38 °C. As a result, the growth of most Poly₁₀X-overexpressing strains, including EGFP (Δ), which had been able to grow at 30 °C, was markedly reduced (Figure S20). In contrast, appending Poly₁₀E, Poly₁₀G, Poly₁₀Q, or Poly₁₀S maintained growth even under this condition, resulting in a substantial increase in relative neutrality (Figure 5A and Figure S20), with Poly₁₀E showing the most pronounced increase in relative neutrality (Figure 5B).

Figure 5.

The adverse effects on proteostasis caused by overexpression are limited in the case of moxGFP, whose folding properties have been improved compared to EGFP ⁵⁶. Indeed, all moxGFP-Poly₁₀X overexpression strains that were able to grow at 30 °C (Figure S5) also grew at 38 °C, and their relative neutrality did not change significantly (Figure 5C, 5D, and S21). In contrast, although Poly₁₀E did not show any beneficial effect in the 30 °C experiments using moxGFP (Figure S6), it became beneficial at 38 °C (Figure 5C and 5D). It is likely that moxGFP, which has stable folding at 30 °C, does not benefit from Poly₁₀E under that condition, but as the structure becomes less stable at 38 °C, the beneficial effect of Poly₁₀E emerges.

Next, we examined whether the addition of Poly₁₀E mitigates the formation of protein aggregates, Hsp70 aggregates, morphological abnormalities in mutant strains, and the induction of the heat shock response that occur upon EGFP overexpression ⁵⁶. The amount of insoluble EGFP-Poly₁₀E in overexpressing cells was significantly lower than that of insoluble EGFP (Figure 5E, 5F, and S22). The number of Hsp70 aggregates in EGFP-Poly₁₀E–overexpressing cells showed a decreasing trend, though not statistically significant, compared to EGFP-overexpressing cells (Figure 5G), and images showing more dispersed aggregates were frequently observed (Figure 5H and S23). Furthermore, the morphological abnormalities of the cdc24-5 and rpl19aΔ strains induced by EGFP overexpression were not observed in EGFP-Poly₁₀E–overexpressing cells (Figure 5I, S24, and S25). Finally, transcriptome analysis by RNA-seq revealed that the heat shock response and the induction of the proteasome-related transcription factor RPN4, both triggered by EGFP overexpression, were markedly reduced in EGFP-Poly₁₀E (Figure 5J and S26A). In contrast, EGFP-Poly₁₀D, which does not show a growth advantage over EGFP (Figure 1C), exhibited only limited suppression of the heat shock response and RPN4 induction (Figure 5K and S26A). Collectively, all these results support the conclusion that the addition of Poly₁₀E to EGFP alleviates the proteostatic burden caused by misfolding during EGFP overexpression.

Notably, when one of the most harmful Poly₁₀X variants, Poly₁₀I, was appended to EGFP, it induced a strong heat shock response and RPN4 activation, opposite to the effect observed with Poly₁₀E (Figure 5L and S26B). Therefore, Poly₁₀I is suggested to either promote misfolding of EGFP more strongly or aggregate on its own, thereby disrupting intracellular proteostasis. These contrasting effects suggest that the physicochemical properties of the appended poly-amino acid strongly influence protein folding and proteostasis.

The neutrality of Poly₁₀X mirrors its evolutionary usage in proteomes

To explore whether the experimentally observed neutrality of Poly₁₀X is reflected in natural proteomes, we first analyzed the maximum number of consecutive residues for each of the 20 amino acids (PolyX_max) and the number of sequences containing ten or more consecutive identical residues (Num-Poly₁₀X) in the S. cerevisiae S288C reference ORFs. We found that both PolyX_max and Num-Poly₁₀X showed clear biases (Figure 6A). About half of the amino acids never appeared as stretches longer than ten residues. To test whether the observed PolyX_max values are maintained by natural selection, we compared them with those obtained from a simulation using randomized ORF sequences (Shuffled S288C ORFs). This analysis revealed that D, Q, N, S, E, R, H, P, K, A, and V displayed significantly longer consecutive runs than expected (q < 0.001, Monte Carlo method with FDR correction). The other amino acids fell within the expected range, and no clear bias toward increasing or decreasing repeat length was detected. Next, using ORFs from 1,392 S. cerevisiae isolates (pan Sc ORFs) ⁵⁷, we calculated PolyX_max and Num-Poly₁₀X with higher resolution. In the S. cerevisiae S288C reference strain, the amino acids A, V, T, G, L, F, I, Y, and M, which showed zero Num-Poly₁₀X values, appeared as Poly₁₀X sequences at very low frequencies (≤ 1); Poly₁₀X of C and W were never observed (Figure 6A).

Figure 6.

To examine whether the PolyX occurrence patterns observed in S. cerevisiae are shared across other organisms, we analyzed nine species (C. elegans, D. melanogaster, sea urchin, amphioxus, Ciona, zebrafish, human, A. thaliana, and rice) for PolyX_max and Num-Poly₁₀X normalized by gene number, and compared their frequencies among species. A comparison of PolyX_max between S. cerevisiae and the other species showed a high correlation coefficient of 0.57 or greater, suggesting that the overall occurrence pattern of PolyX_max is largely conserved across species (Figure 6B). In contrast, a similar comparison using Num-Poly₁₀X yielded a lower correlation of 0.38 or higher (Figure 6C). Whether this lower correlation reflects species-specific characteristics or results from annotation errors in ORF datasets remains to be determined.

Finally, we compared the occurrence patterns of PolyX observed in ORFs with the neutrality of Poly₁₀X determined in our experiments, in which each Poly₁₀X was fused to the C-terminus of a fluorescent protein (Figure 2). As shown in Figure 6D and 6E, both PolyX_max and Num-Poly₁₀X in the pan-ORF dataset showed strong correlations with neutrality. Five amino acids—E, S, N, Q, and D—were clearly separated from the others, exhibiting longer PolyX_max values and higher occurrence numbers in Poly₁₀X. These amino acids also displayed beneficial effects when fused to the C-terminus of certain fluorescent proteins (neutrality >10,000). In contrast, hydrophobic amino acids with low occurrence frequencies generally showed low neutrality and high harmfulness. Exceptions were A, G, and T, which exhibited high neutrality despite having an average Num-Poly₁₀X of less than one in the pan Sc ORF dataset. The probability that these amino acids (A, G, and T) would randomly appear as ten consecutive residues is low (Figure 6A). Therefore, their behavior may represent that of amino acids that are neutral and lack inherent beneficial effects. Taken together, these findings indicate that Poly₁₀X neutrality, as determined experimentally, closely explains the occurrence patterns of amino acid homorepeats in natural proteomes.

Discussion

In this study, we systematically evaluated the neutrality (harmful and beneficial properties) of simple amino acid sequences in which a single residue is repeated ten times (Poly₁₀X). We found that neutrality varied markedly among amino acids, with hydrophobic residues exhibiting pronounced harmfulness (Figures 1 and 2). In contrast, several hydrophilic or negatively charged amino acids (e.g., E, S, N, Q) showed beneficial effects on cellular tolerance to protein overexpression, which was particularly intriguing. Notably, the harmfulness of Poly₁₀I, Poly₁₀V, and Poly₁₀W persisted even when detached from EGFP (Figure 3), indicating that these repeats possess intrinsic harmfulness. Conversely, the beneficial effects of Poly₁₀E, Poly₁₀Q, Poly₁₀N, and Poly₁₀S disappeared upon detachment from EGFP (Figure 3), indicating that these repeats reduce cellular burden by mitigating the toxicity of the host protein. Mechanistically, the harmful effect of Poly₁₀I appears to stem from its intrinsic aggregation propensity, whereas the beneficial effect of Poly₁₀E is likely due to its ability to prevent aggregation (Figures 4 and 5).

Furthermore, when we compared the occurrence frequencies of PolyX across the proteomes of diverse organisms, we found that their cross-species occurrence profiles were remarkably similar across species and showed strong correlation with the experimentally determined neutrality values (Figure 6). Notably, hydrophobic residues, which rarely form homorepeats in natural proteins, exhibited high harmfulness in our assays, suggesting that such sequences are likely eliminated immediately upon emergence during evolution. In contrast, amino acids such as E, N, Q, and S—whose homorepeats are widely observed in natural proteins—showed low harmfulness and, in some cases, even beneficial effects, implying that their appearance may confer an advantage to cells from the outset. PolyQ, for example, can contribute to normal protein function by modulating protein–protein interactions when maintained at moderate lengths, but become pathogenic upon excessive expansion ³⁹. Thus, the evolutionarily permissible range of PolyQ length is likely determined by a balance between harm and benefit.

Taken together, the occurrence frequencies of amino acid sequences in natural proteomes are well explained by experimentally determined neutrality, reflecting a balance between harmful and beneficial effects. Our assays were performed by placing simple homorepeat sequences at the C-terminus or within internal loops of fluorescent proteins and expressing them in the cytosol. Because homorepeats found in natural proteins predominantly occur in intrinsically disordered regions (Figure S29), our experimental design provides an appropriate framework for evaluating their harmful and beneficial effects. In contrast, whether similar principles apply to more complex sequences buried within membranes or folded protein cores remains an open question. For structured or interaction-dependent motifs, intrinsic harmfulness might be masked—or even neutralized—by specific binding partners. This idea raises the possibility that “eliminating harmfulness” itself acts as an evolutionary pressure driving the formation of higher-order structures or interaction modules. Interestingly, hydrophobic homorepeats—among the most harmful sequences in our assays—are also utilized as secretion signals ⁵⁸. It is tempting to speculate that what was originally a “dangerous” sequence requiring removal from the cytosol may have been evolutionarily co-opted into the secretory pathway.

The design of artificial proteins is rapidly advancing. However, it remains difficult to determine whether sequences absent in nature are missing because they offer no benefit or because they are eliminated due to harmfulness. The parameter we identified in this study—neutrality, which integrates both harmful and beneficial effects—may represent a fundamental evolutionary design principle that shapes the compatible amino acid sequence space in living organisms. A deeper understanding of this parameter will aid in rational protein design and may also shed light on the molecular mechanisms underlying neurodegenerative diseases and aging, in which mutations elevate protein harmfulness.

Materials and Methods

Strains, plasmids, growth conditions

S. cerevisiae and E. coli strains and plasmids used in this study are listed in the Key Resources Table. Synthetic complete (SC) medium lacking uracil (U) or leucine (L) was used for yeast cultures. LB medium was used for E. coli cultures.

Genetic tug-of-war (gTOW) method

The gTOW method ^50,51 enables artificial elevation of plasmid copy number for a target protein by exploiting the combination of a multi-copy plasmid replication origin (2μ ORI) and the auxotrophic selection marker leu2-89. In SC–U medium, where no copy-number–increasing selection pressure is present, the plasmid remains at a low copy number (approximately 30 copies for an empty vector). In contrast, in SC–LU medium, strong selection pressure increases the plasmid copy number to approximately 150 copies. As plasmid copy number increases, the expression level of the target protein also rises. If expression becomes high enough to inhibit cell growth before the plasmid reaches ∼150 copies, this expression level is defined as the protein’s maximum tolerable expression level. Thus, gTOW allows quantitative determination of how much of a given protein a yeast cell can tolerate before growth inhibition occurs. By combining the measured expression level with the degree of growth inhibition, protein harmfulness can be represented as a one-dimensional parameter, relative neutrality, enabling direct comparison of harmfulness across different proteins ⁵⁵. In this study, modified fluorescent proteins were expressed under either the strong TDH3 promoter or the WTC₈₄₆ (P_7.tet1) promoter, a TDH3-derived variant whose expression can be induced by anhydrotetracycline (aTc) ⁵².

Measurement of growth rate, fluorescence intensity, and calculation of the relative neutrality in yeast

Cells of the BY4741 or BYW2 strain carrying the expression plasmids were pre-cultured in SC–U and then transferred to either SC–LU (for constructs expressed under TDH3_pro in BY4741) or SC–LU supplemented with aTc (for constructs expressed under the WTC₈₄₆ promoter in BYW2). Cells were cultured under these conditions while OD₅₉₅ and fluorescence signals were monitored every 10 or 30 minutes using an Infinite F Nano+ microplate reader. Fluorescence of EGFP, moxGFP, mNeonGreen, and Gamillus was measured using an excitation/emission filter set of 485/535 nm, whereas fluorescence of mScarlet-I and mCherry was measured using a 535/590 nm filter set. From these time-course measurements, the maximum growth rate (MGR) and maximum fluorescence intensity (MFI) were determined using a custom Python script. The relative neutrality index was calculated by multiplying the percentage of MGR relative to the control vector (Δ) by the percentage of MFI relative to the control fluorescent protein as follows: Relative neutrality = %(MGR_Poly₁₀X / MGR_Δ) × %(MFI_Poly₁₀X / MFI_Δ).

Measurement of growth rate, fluorescence intensity, and calculation of the relative neutrality in E. coli

Cells of the BW25113 strain carrying the expression plasmids were pre-cultured in LB + ampicillin medium and then transferred to fresh LB + ampicillin medium. Cells were cultured under these conditions while OD₅₉₅ and fluorescence (485/535 nm) signals were monitored every 5 minutes using an Infinite F Nano+ microplate reader. From these time-course measurements, the maximum growth rate (MGR) and maximum fluorescence intensity (MFI) were determined using a custom Python script. The relative neutrality index was calculated by multiplying the percentage of MGR relative to the control vector (Δ) by the percentage of MFI relative to the control fluorescent protein as follows: Relative neutrality = %(MGR_Poly₁₀X / MGR_Δ) × %(MFI_Poly₁₀X / MFI_Δ).

Clustering analysis

Clustering analysis was performed using a custom Python script (pandas, seaborn, and matplotlib) after organizing the relative neutrality data in an Excel file. Hierarchical clustering and heatmap visualization were conducted using the clustermap function in seaborn. Euclidean distance (metric = “euclidean”) was used as the distance metric, and average linkage (UPGMA; method = “average”) was applied for clustering. The resulting clustering patterns were visualized as row- and column-wise heatmaps, with dendrograms shown or hidden depending on the purpose of the analysis. A custom continuous colormap defined by three reference points was used, and the display range was fixed from 0 to 20,000. This scaling was chosen because relative neutrality values were normalized such that the control (Δ) corresponded to 10,000, representing the non-toxic state; accordingly, the color scale was centered at 10,000.

Amino acid parameters

For the correlation analysis in Figure S10, a total of 28 amino acid parameters were used, consisting of 18 nonredundant indices from the AAindex database (https://www.genome.jp/aaindex/; see https://doi.org/10.1002/pro.5239) and 10 additional parameters: hydrophobicity ⁵⁹, hydropathy index ⁶⁰, isoelectric point, side-chain molecular weight, amino acid usage (%) (this study), metabolic cost ⁶¹, biosynthetic cost (ATPs) ⁶², biosynthetic steps (enzymes) ⁶², and total energy costs of amino acids and nucleotide precursors under fermentative or respiratory conditions ⁶³.

Microscopic observation

Yeast cells overexpressing the target protein were cultured in either SC–U or SC–LU medium, and imaged using a DMI6000B microscope (Leica Microsystems), and images were processed using the Leica Application Suite X software. GFP and RFP fluorescence signals were acquired using a GFP filter cube and RFP filter cube, respectively.

For the morphology analysis in Figure S24 and S25, microscopic images were processed using Cellpose ⁶⁴ for cell segmentation. The segmented images were subsequently analyzed using the MeasureObjectSizeShape module in CellProfiler ⁶⁵ to quantify the major and minor axes of individual cells. Elongation ratio values were calculated for each cell using the major and minor axis lengths obtained from the MeasureObjectSizeShape module, either with a custom Python script or in Excel, according to the following formula: Elongation Ratio = MajorAxisLength / MinorAxisLength. Cells with an elongation ratio ≥ 1.5 were defined as morphologically abnormal.

Protein analysis

BY4741 strains expressing the target proteins were cultured overnight in 6 mL or 25 mL of SC–LU medium. Total protein was extracted from cells in the logarithmic growth phase (OD₆₆₀ = 0.9–1.0) using 0.2 mol/L NaOH, followed by solubilization in NuPAGE sample buffer (Thermo Fisher Scientific). For each analysis, total protein was extracted from a cell mass corresponding to 1 OD unit at OD₆₆₀.

For the cell lysate analysis, cells were washed with PBST [10 mM phosphate-buffered saline (pH 7.4), 0.001% Tween 20] supplemented with Halt protease inhibitor cocktail (Thermo Fisher Scientific), then disrupted with glass beads using a bead-beating homogenizer (Micro Smash MS-100, TOMY) at 5000 rpm for 30 s, repeated five times. Cell lysates were centrifuged at 15,000 rpm for 10 min to separate soluble and insoluble fractions. The insoluble fraction was washed with 1 mL PBST and centrifuged again at 15,000 rpm for 10 min.

Extracted proteins were fluorescently labeled using EZLabel Fluoro NEO (ATTO) according to the manufacturer’s instructions and separated by SDS–PAGE on 4–12% gradient gels. Fluorescent signals were detected using the SYBR Green fluorescence mode of a LAS-4000 image analyzer (GE Healthcare). Proteins separated by SDS–PAGE were transferred onto a PVDF membrane (Thermo Fisher Scientific). GFP signals were detected using an anti-GFP primary antibody (Roche), an peroxidase-conjugated secondary antibody (Nichirei Bioscience), and chemiluminescent substrate (Thermo Fisher Scientific). Chemiluminescent images were acquired with the chemiluminescence mode of the LAS-4000 image analyzer.

RNA sequencing analysis

S cerevisiae strain BY4741 with the control vector or overexpressing EGFP, EGFP–Poly₁₀D, or EGFP–Poly₁₀E was cultured in SC–LU medium, and strain BY4743 with the vector or overexpressing EGFP, or EGFP–Poly₁₀I was cultured in SC–U medium. Cells were collected at the logarithmic growth phase (OD₆₆₀ ≈ 1.0). Total RNA was extracted according to a previously described protocol ⁶⁶. The concentration of purified RNA was initially measured using a Qubit fluorometer (Thermo Fisher Scientific), and samples were stored at −80 °C until further use. Before library preparation, RNA concentrations were re-measured using the Quant-iT RiboGreen RNA Assay Kit (Invitrogen) on an ARVO Multimode Microplate Reader (PerkinElmer), and RNA quality was assessed with a 2100 Bioanalyzer (Agilent Technologies). cDNA libraries were prepared using the TruSeq Stranded mRNA Library Prep Kit (Illumina) and sequenced on a NovaSeq X Plus platform (Illumina) with 100-bp paired-end reads. Raw sequencing data were quality-checked using FastP ⁶⁷ and aligned to the reference genome with Hisat2 ⁶⁸ . Aligned data were converted to BAM format with Samtools ⁶⁹ and quantified using StringTie ⁷⁰. Differential gene expression analysis was performed using EdgeR ⁷¹. Gene Ontology (GO) enrichment analysis was conducted with the clusterProfiler ⁷² and org.Sc.sgd.db packages (DOI: http://doi.org/10.18129/B9.bioc.org.Sc.sgd.db) in R to identify significantly enriched GO terms. All analyses were performed using three biological replicates for each strain. The raw data were deposited into DDBJ (accession number: PRJDB39951).

PolyX analysis in the S. cerevisiae proteome and other species

Proteome-wide PolyX analysis was performed using all verified ORFs registered in the Saccharomyces Genome Database (SGD), covering all 20 amino acids. As of October 2024, 6,585 ORFs are annotated in the S. cerevisiae (S288C) genome. From this set, 5,911 verified ORFs (Sc ORFs) were retained for analysis after removing those annotated as “Dubious.” A custom Python script first loaded the ORF list and excluded all entries whose Qualifier field was labeled “Dubious.” Regular expressions were then used to identify, for each ORF, the longest homorepeat for each amino acid. The resulting maximum repeat lengths were compiled into a dataframe, and two metrics were calculated as follows: PolyX_max – the maximum length of consecutive identical amino acids for each amino acid across all ORFs, and Num_Poly₁₀X – the number of ORFs containing a homorepeat of length ≥ 10 for each amino acid. For pan Sc ORFs, the same analysis was conducted using FASTA assemblies downloaded from Zenodo (https://zenodo.org/records/3407352), which include 1,392 S. cerevisiae genome assemblies ⁵⁷.

Simulated (shuffled) S288C ORFs were generated as follows using a custom Python script. Each sequence of above 5,911 ORFs was randomly shuffled while preserving amino acid composition. This produced randomized sequences with identical amino acid compositions but altered residue order. PolyX_max and Num_Poly₁₀X values were computed for each shuffled ORF set using the same regular expression pipeline described above. This simulation was repeated 10,000 times, using seed values from 0 to 9,999. For each iteration, PolyX_max and Num_Poly₁₀X results were stored in a dataframe. For every amino acid, the minimum, mode, mean, maximum, and standard deviation across the 10,000 trials were calculated and summarized.

For cross-species comparisons, proteomes of nine eukaryotic species (Caenorhabditis elegans, Drosophila melanogaster, Strongylocentrotus purpuratus, Branchiostoma lanceolatum, Ciona intestinalis, Danio rerio, Homo sapiens, Arabidopsis thaliana, and Oryza sativa) were downloaded from Ensembl databases (https://asia.ensembl.org). When multiple isoforms were annotated for a given gene, the isoform with the longest amino acid sequence was selected as the representative sequence for analysis. For each amino acid sequence, PolyX of identical amino acids were detected based on regular-expression matching. Two metrics were calculated for each of the 20 amino acids: (i) PolyXmax, defined as the maximum length of consecutive identical residues observed across all proteins, and (ii) Num–Poly₁₀X, defined as the number of proteins containing a homorepeat of length ≥10 residues. For interspecies comparisons, Num–Poly₁₀X values were normalized by the total number of genes analyzed for each species. Correlations between PolyX occurrence metrics (PolyXmax and Num–Poly₁₀X) and experimentally determined Poly₁₀X neutrality values were evaluated using Spearman’s rank correlation coefficients.

Supplementary Figures

Figure S1.

Figure S2.

Figure S3.

Figure S4.

Figure S5.

Figure S6.

Figure S7.

Figure S8.

Figure S9.

Figure S10.

Figure S11.

Figure S12.

Figure S13.

Figure S14.

Figure S15.

Figure S16.

Figure S17.

Figure S18.

Figure S19.

Figure S20.

Figure S21.

Figure S22.

Figure S23.

Figure S24.

Figure S25.

Figure S26.

Figure S27.

Figure S28.

Figure S29.

Data availability

All source data used for figure generation, and all analysis scripts have been deposited on GitHub (https://github.com/hisaomlab/Murase_PolyX). RNA-seq data have been deposited in a public repository (DDBJ accession number: PRJDB39951). All plasmids used in this study are available from the National BioResource Project (NBRP)-yeast (https://yeast.nig.ac.jp/yeast/).

Acknowledgements

We thank the members of the Moriya laboratory (Okayama University) for helpful discussions, Yuhei Chadani (Okayama University) for providing the E. coli strain, and Christian Landry (Université Laval) for valuable comments and suggestions on the manuscript.

Additional information

Funding

Nagase Science Technology Foundation

• Hisao Moriya

Institute for Fermentation, Osaka (IFO)

• Hisao Moriya

Japan Society for the Promotion of Science (25K22437)

• Hisao Moriya

MEXT | Japan Society for the Promotion of Science (JSPS) (24K02313)

• Hisao Moriya

New Energy and Industrial Technology Development Organization (NEDO) (P20011)

• Hisao Moriya