Introduction

Most random amino acid mutations destabilize proteins and impair their function (1,2). Nevertheless, deleterious mutations can accumulate and persist within populations through multiple mechanisms that offset fitness costs. Epistatic interactions can mask fitness defects in specific genetic contexts; a mutation that is highly deleterious in one genetic background can become nearly neutral in the presence of another mutation, permitting the accumulation of otherwise deleterious variants (3–6). Proteins can also exhibit mutational robustness, maintaining structural stability and function despite acquiring destabilizing mutations (7, 8). At the organismal level, allelic redundancy can buffer fitness defects when deleterious mutations affect only a subset of gene copies (9, 10). Environmental pressures, such as antibiotic exposure, can render otherwise deleterious mutations selectively advantageous, promoting their persistence in populations (11).

For deleterious nucleotide substitutions, fitness can sometimes be restored through direct reversion to the original sequence. However, when multiple mutations accumulate, or when selective pressures favor retention of the deleterious mutations, additional compensatory mutations provide an alternative route to fitness recovery. These mutations can restore organismal fitness to near wild-type (WT) or even WT levels (Fig. 1 A), a phenomenon documented across diverse biological systems (6, 11–13). Notably, compensatory mutations can occur at higher frequencies than direct reversions (14), particularly following antibiotic-induced population bottlenecks (15–18).

Fig. 1.

Intragenic compensatory mutations broadly operate through two mechanistic modes: global structural stabilization and pairwise interaction repair. Due to their effect on protein stability, global compensatory mutations have weak spatial constraints relative to destabilizing mutations, provided they contribute to the same folded energetic unit. This spatial flexibility enables broad-spectrum functional rescue across multiple mutational positions (18). In contrast, pairwise compensatory mechanisms typically involve residues in three-dimensional contact, forming active-site interfaces, α-helix packing cores, or allosteric communication pathways (19). As a result, local compensation is often highly context specific (20) and can even become destabilizing in the absence of the partnered deleterious mutations (14).

Previous investigations of compensatory evolution have largely relied on small-scale experiments examining the effects of limited sets of genetic variants on individual macromolecules, including proteins (11, 12, 21, 22), mRNA (23, 24), and tRNA (25). These studies demonstrated that compensatory mutations can restore fitness, but also reported cases where recovery appeared to primarily rely on direct reversion (26, 27). Because only a small fraction of sequence space (< 100 genotypes per study) was explored, findings varied substantially across studies (13, 28, 29). This limited sampling has impeded the identification of robust predictors of compensatory potential and obscured broader principles governing the compensation of deleterious mutations.

To address this gap, we analyzed a large-scale deep mutational scanning (DMS) dataset comprising ∼400,000 amino acid variants of the yeast HIS3 gene, which encodes the histidine biosynthesis enzyme imidazoleglycerol-phosphate dehydratase (IGPD/His3p) (30). This dataset provides near-complete mutational coverage (>95% of residues) and includes ∼30,000 variants that impair growth in histidine-deficient media. At a scale more than three orders of magnitude larger than previous studies, this dataset enables a systematic exploration of the protein-level principles governing compensatory evolution. Our analysis shows that compensatory potential can be predicted, in part, from protein structural features. Most notably, we identify a distinct class of super compensator substitutions that enhance fitness across diverse genetic backgrounds and buffer the fitness effects of mutations, effectively flattening fitness landscapes. The prevalence of these super compensators is further supported by analyses of several additional DMS datasets. Together, these findings position compensatory substitutions, particularly super compensators, as significant contributors to the shaping of evolutionary trajectories and maintenance of genetic diversity in populations.

Results

A DMS Dataset to Identify Intragenic Compensation

To systematically examine the full spectrum of compensatory mutations across an entire protein, we analyzed a large-scale deep mutational scanning (DMS) library of the Saccharomyces cerevisiae HIS3 gene (30). The HIS3 coding sequence was divided into twelve 20–30 residue segments, each subjected to localized mutagenesis. Fitness of all variants was quantified through pooled competitive growth assays and normalized such that genotypes carrying premature stop codons were assigned a fitness of zero and wild-type genotypes a value of one (Materials and Methods). After filtering the data for quality (Materials and Methods), the final dataset comprised 372,867 amino acid variant genotypes. Notably, many of these variant combinations correspond to naturally occurring alleles observed across 21 extant yeast species (Fig. S1A, B). Among them, 31,701 (8.5%) genotypes displayed a relative fitness of zero and served as the foundation for subsequent analyses (Fig. S1C).

In our data, we find that unfit (fitness = 0) genotypes can be rescued by additional substitutions (Fig. 1 B). To quantify the extent of rescue, we defined rescuability as the average fitness gained by unfit genotypes upon acquiring one or more additional substitutions (Fig. 1 C; Materials and Methods). Using this metric, we find that only a small subset of unfit genotypes exhibit appreciable rescuability (Fig. 1 D).

To control for errors in fitness estimation, we used nonsense genotypes containing an internal stop codon as a negative control. Functional rescue of such genotypes is implausible, and indeed we observed little to no rescue among them in our data. Using their rescuability distribution, we estimated a false discovery rate of 5% for rescuability values ≥ 0.02 and 1% for values ≥ 0.04 (Fig. 1 E). Based on these calibrations, we classified an unfit genotype as rescuable when its rescuability was ≥ 0.1, identifying 3,750 (11.8%) such genotypes. Notably, comparable rescuability were obtained across a range of fitness cutoffs for definitions of fit and unfit genotypes (Fig. S1D) as well as under alternative procedures for calculating rescuability (Fig. S1E).

Characterization of Structural Features Governing Compensation Potential

As nonfunctional variants span nearly the entire protein (Fig. S1F), we used the His3p library to evaluate structural features influencing intragenic compensation. To quantify compensatory potential structurally, we defined position-wise rescuability as the mean rescuability of all unfit genotypes carrying substitutions at a given residue (Fig. 2 A). Position-wise rescuability varied markedly across residues, with solvent-exposed, core-distal positions, which tend to be weakly evolutionarily conserved, exhibiting significantly higher rescuability (Fig. 2 B), consistent with previous observations (29).

Fig. 2.

To investigate the mechanistic relationship between rescuability and structural features, we examined predicted changes in Gibbs free energy change of folding (ΔΔG_fold) for 10,619 representative genotypes. Across all variants, rescuable genotypes exhibited lower ΔΔG_fold than non-rescuable ones, indicating that severe structural destabilization constrains compensatory potential (Fig. 2 C; Fig. S2A). Moreover, the addition of compensatory mutations partially mitigated destabilization, while restoration to WT stability was uncommon (Fig. 2 D). Compensatory stabilization was most pronounced for moderately destabilized variants (ΔΔG_fold < 40), suggesting that structural stabilization contributes significantly, but not exclusively, to intragenic compensation.

Because His3p functions as a homo-24-mer complex, we also assessed stability changes at subunit interfaces, where substitutions can influence both monomer folding and inter-subunit binding (31–33). For genotypes with more than 50% of substitution in interface regions, predicted changes in Gibbs free energy change of binding (ΔΔG_bind) mirror the folding analysis: rescuable genotypes had smaller destabilizing shifts, and compensatory substitutions further mitigated losses in binding affinity (Fig. 2 E and F; Fig. S2B). Together, these results suggest that intragenic compensation enhances protein robustness by alleviating destabilizing effects both on monomeric structure and multimeric assembly.

Super Compensatory Substitutions Increase the Fitness of Diverse Variants

Given the observed structural specificity of rescuability, we next sought to understand modes of compensation across different genetic backgrounds. Specifically, we asked whether compensatory substitutions act in a genotype specific manner or whether some substitutions can restore fitness across a broad range of deleterious genotypes.

To quantify these modes of compensation, we represented intragenic rescue relationships using a directed graph in which nodes correspond to amino-acid states and directed edges encode compensatory effects. An edge from amino acid state A to B indicates that introducing substitution B increases the fitness of genotypes carrying state A. Edge weights represent the fraction of such genotypes rescued beyond a defined fitness threshold (≥ 0.5; Fig. 3 A-B). The overall compensatory ability of a substitution was defined as the sum of its incoming edge weights, quantifying the breadth of rescue across heterogeneous genetic contexts. Thus, the network captures how well each substitution can rescue genotypes with distinct existing amino acid states.

Fig. 3.

This analysis revealed a small subset of super compensatory substitutions—specific amino acid changes (e.g., A110D) capable of increasing the fitness of genetically diverse variants (Fig. 3 C and D; Fig. S3; Fig. S4). The corresponding non-WT amino-acid states resulting from these substitutions (e.g., 110D) are hereafter referred to as super compensators.

To investigate why certain substitutions emerge as super compensatory, we employed a published neural-network model that assigns each amino acid state a fitness impact score (30), quantifying its contribution to a latent variable termed fitness potential. The model maps fitness potential to predicted fitness via a sigmoidal transformation (Fig. 3 E and F; Fig. S5). Within this framework, each genotype’s fitness potential equals the sum of the fitness impact scores of its constituent amino acid states, providing an interpretable representation linking amino acid state-specific effects to overall fitness. Remarkably, super compensators exhibited substantially higher fitness impact scores than their WT counterparts (Fig. 3 F and G). These elevated scores enable super compensatory substitutions to increase fitness potential—and thus predicted fitness—across diverse genetic backgrounds, offering an explanation for their broad compensatory action.

Super Compensators Buffer Substitution effects

To understand how super compensators achieve broad compensation, we examined their influence on the fitness effects of other substitutions. Specifically, we compared the fitness effects of identical substitutions introduced into wild-type (WT) versus super compensator backgrounds. Remarkably, replacing a WT amino acid with a super compensator substantially attenuated the fitness effect of both deleterious and beneficial substitutions at other positions, indicating a strong buffering effect (Fig. 4 A).

Fig. 4.

To quantify this buffering effect, we computed the fitness difference (Δf) of each substitution with the WT (Fig. 4 B, y-axis) and the non-WT amino acid state (Fig. 4 B, x-axis). If substitution effects were comparable in both contexts, data points would distribute symmetrically around the diagonal. Instead, for super compensators such as 110D, the distribution collapsed along the vertical axis (Fig. 4 E), demonstrating substantial dampening of substitution effects. We further quantified this behavior by calculating the absolute deviation between the x- and y-axes (Fig. 4 C, F), and assessing the skewness of this distribution. Substitutions in super compensator backgrounds exhibited both reduced absolute fitness effects and greater skewness relative to non-super compensator contexts (Fig. 4 D, G), confirming the enhanced buffering capacity of super compensators.

To identify the underlying cause, we revisited the sigmoidal relationship between fitness potential and predicted fitness defined by the neural-network model. Most genotypes in our library reside on the high-fitness plateau, where substantial variation in fitness potential produces only minor change in predicted fitness (Fig. 4 H). Beneficial substitutions that push genotypes further along this plateau have reduced sensitivity toward fitness deviations due to additional substitutions, yielding positively skewed fitness-effect distributions. Conversely, deleterious substitutions drive genotypes off the plateau, increasing sensitivity and generating negatively skewed distributions (Fig. 4 H). Across all genotypes, we found that amino acid states with higher fitness impact scores than their WT counterparts exhibited higher skewness in the distribution of absolute fitness-effect differences (Fig. 4 I). Similarly, substitutions with higher compensatory ability showed correspondingly increased skewness (Fig. 4 J). These observations indicate that super compensators buffer both deleterious and beneficial substitution effects by operating within a flattened region of the fitness landscape.

Whereas prior studies have emphasized the ruggedness of protein fitness landscapes, characterized by numerous local peaks and valleys (34–37), our results show that super compensators can locally smooth these landscapes. The super compensator 93V provides a clear example: in variants containing substitutions at positions 66 and 86, 93V substantially reduces fitness variability among neighboring variants, effectively flattening the local fitness surface (Fig. 4K).

Super Compensators Stabilize Inter-residue Interactions

Building on the role of stability restoration in compensation, we next evaluated structural features of super compensatory substitutions. In general, super compensators significantly increased overall protein stability (Fig. 4 L; Table S1). For instance, 110D forms a salt bridge with R112, effectively stabilizing a neighboring loop region. In contrast, substitutions exhibiting lower fitness impact scores than the WT, such as I19D, were predicted to disrupt a β-sheet and its adjacent helix (Table S1).

Notably, most super compensators are spatially distant from the catalytic sites (> 8 Å), with none located within 4 Å of the substrate (Fig. S6). This spatial separation supports a mechanism of global stabilization rather than direct modulation of active-site interactions.

Super Compensators Enhance Organisms’ tolerance to Deleterious Mutations

To experimentally verify the functional consequences of super compensation, we performed deep mutational scanning to determine whether super compensators confer mutational robustness by buffering deleterious substitutions. We designed the library to contain two fully randomized flanking positions (NNK codons) surrounding a central position encoding either the WT or the super compensator S189A amino acid (Fig. 5 A). Under histidine-deficient conditions, variants carrying premature stop codons displayed severely reduced fitness, which returned to WT levels upon histidine supplementation (Fig. 5 B), confirming the reliability of our assay.

Fig. 5.

Introduction of the super compensator 189A resulted in a marked fitness increase for 77% of all variants in the library (Fig. 5 C). The rescue effect was strongest among genotypes harboring mildly deleterious substitutions (Fig. 5 D). Together, these results provide direct experimental evidence that super compensators can increase mutational tolerance by buffering fitness defects caused by otherwise deleterious mutations.

Super Compensators are Prevalent across Diverse Proteins

To determine whether super compensators represent a general phenomenon beyond His3p, we applied the same analytical framework to published DMS datasets from other randomly mutagenized proteins assayed using diverse fitness readouts (38). This analysis revealed that super compensators are broadly distributed across proteins and similarly buffer the fitness effects of mutation in independent systems (Fig. S7).

Notably, fitness was defined differently across datasets, yet our findings were robust to these differences. In His3p fitness was quantified by yeast growth rate, whereas in another dataset it was protein stability in a cell free system (39) and in the case of GFP, using fluorescence intensity as a proxy for protein folding and activity (40). Together, these results support a general model in which super compensators exert their effects, at least in part, by enhancing protein stability.

Discussion

Our analysis reveals the principles governing intragenic compensation, delineating both the characteristics of rescuable genotypes and compensatory substitutions. We found that the compensatory potential of non-functional variants is strongly correlated with protein structural context: substitutions at solvent-exposed, weakly spatially constrained positions are more easily rescued. This finding clarifies why Tomala et al. (13) identified a limited set of compensatory substitutions for temperature-sensitive mutations, which often reside in structurally critical regions, consistent with independent evidence that some deleterious mutations are fundamentally uncompensatable (27). Consistent with Tomala et al.’s results, we observe that deleterious mutations causing modest structural destabilization can be rescued, reinforcing our observation that the destabilizing effect of substitution in folded and conserved positions is stronger than those in the solvent-exposed positions (31, 41). Beyond structural context, we also demonstrate that amino acid physicochemical properties can serve as predictors for fitness. Specifically, when the AAindex values—a set of quantitative descriptors of amino acid attributes—in mutant variants approximates those of the WT, fitness recovery is observed (Fig. S8). That is to say, compensation can occur when mutations bring the AAindex values closer to those of the WT.

Focusing on the properties of compensatory substitutions themselves, we identified a distinct class of super compensatory substitutions capable of increasing fitness of diverse genetic backgrounds. This phenomenon parallels the concept of cryptic mutations (42), which facilitate the accumulation of otherwise deleterious substitutions. Our experimental validation directly demonstrates this capacity: introducing super compensators reduced deleterious effects of mutations, thereby increasing mutational robustness. The existence of super compensators can thus provide an alternative explanation for the mutational robustness of proteins. Previous work by Guo et al (43) and György Abrusán et al (44) found mutational robustness to vary across protein secondary structures. This aligns with our observation of increased compensation at solvent exposed positions and indicates that structure context shapes both the emergence and effectiveness of compensatory mechanisms.

The evolutionary implications of super compensators may extend beyond laboratory assays. If super compensators can increase the fitness of diverse variants, they should leave a detectable signature in natural sequences by enabling the retention of mildly deleterious alleles. Consistent with this expectation, comparative analysis of 335 His3p orthologs revealed that super compensators frequently co-occur with mildly deleterious amino acid states across extant yeast species (Fig. S9A). Moreover, the strength of the super compensator (its fitness impact score) was positively correlated with the severity of its partner deleterious amino acid states (Fig. S9B). These observations provide a plausible explanation for the sequence divergence observed among His3p orthologs and suggest that super compensators may have contributed to natural genetic variation.

These insights have practical implications for protein engineering. The ability of super compensators to confer mutational robustness suggests that predicting and incorporating them could serve as an effective strategy for designing proteins with enhanced stability and expanded functions. Accordingly, prioritizing stabilizing substitutions represents a valuable component for protein design, consistent with prevailing approaches that emphasize predicting stability-increasing mutations (45–48).

While the His3p DMS dataset is invaluable, its limitations warrant consideration. First, its basis in extant sequence backgrounds may underrepresent severely deleterious substitutions. Second, segmenting the gene precluded the detection of inter-segment compensatory effects. Nevertheless, to our knowledge, the His3p library remains one of the largest and most comprehensive resources that contains multi-site mutants, making it ideal for probing structure-compensation relationships. Furthermore, we are able to recapitulate our results regarding the stabilizing role of super compensator across other large-scale DMS datasets, including those based on GFP fluorescence (40) and other stability-directed assays (39). This cross-validation underscores the generalizability of our conclusions beyond His3p.

In summary, this work deepens our understanding of how compensatory evolution is constrained by protein structure, reveals how super compensators promote mutational robustness and sequence diversity, and highlights how these principles can inform rational strategies for protein engineering.

Materials and Methods

Experimental Design

The experimental methodology and analytical procedures were already fully described in the previous work (30). We briefly summarize the key steps as follows: first, plasmids containing each HIS3 gene variant encoding IGPD with distinct amino acid substitutions were constructed and transformed into ΔHIS3 S. cerevisiae. Mixed populations of yeast strains carrying different HIS3 variants were then co-cultured in bulk competition assays. Deleterious substitutions in IGPD resulted in slow growth and decline of variants from the population over time, neutral or beneficial substitutions enabled rapid growth and eventual dominance.

Through competitive growth analysis, variants with missense substitutions exhibiting growth rates comparable to or lower than premature stop codon-containing variants were assigned a fitness value of zero. In contrast, WT genotypes and the majority of single-substitution variants demonstrated fitness values of 1. The experimental design included two biological replicates with two technical replicates per biological replicate. Following rigorous quality control, which excluded data with a replicate correlation of < 0.85, a further filter was applied to ensure robust fitness assessment. Segment 9 exhibited low correlation between biological replicates and was therefore excluded from this work (30). Specifically, amino acid genotypes represented by fewer than three synonymous nucleotide variants were discarded. The resulting dataset consisted of genotypes with an average of seven amino acid substitutions from the reference sequence. The fitness of each genotype was assessed by averaging nine synonymous variants. As these synonymous mutants exhibited little to no fitness difference (30), this internal replication provided high confidence.

Calculation of Rescuability

For each unfit genotype, we identified all genotypes containing the same set of substitutions plus one or more additional substitutions (hereafter referred to as the superset, total count N). Fitness for each amino acid genotype was estimated using synonymous nucleotide variants, yielding a mean fitness (S) and standard deviation (S.D.). The 95% confidence interval for fitness was defined as S ± 1.96×S.D. A superset genotype A was considered to compensate for genotype B when its lower confidence bound exceeded the upper confidence bound of B, i.e., S_A-1.96S.D._A ≥ S_B +1.96S.D._B. For each unfit genotype, we calculated fitness difference (Δf = S_A - S_B) for all compensating superset genotypes (N1 out of N). Rescuability was then defined as the sum of these Δf values normalized by the total superset count :

As an independent validation, we used a non-parametric approach using synonymous variants. For each candidate compensating genotype A and unfit genotype B, we compared the distribution of fitness values of their respective synonymous variant sets (G_A and G_B) using a two sample Kolmogorov-Smirnov test (α=0.05). Genotype A was considered compensatory if the fitness distribution of G_A was significantly higher than G_B. Fitness difference (Δf) and rescuability values were then computed analogously to the primary method.

Calculation of Structural and Evolutionary Metrics

WCN Score

The WCN calculation was performed using established methodology from Lin CP et al.(49), with the formula:

where r_ij denotes the distance between the C_α geometric center of residues i and residues j in IGPD (N total residues). The WCN value for each position was calculated by averaging measurements across 24 monomeric structures.

RSA Score

The RSA score for each position of IGPD was derived from the DSSP file (generated using IGPD’s PDB file) (50). The RSA values were normalized using the amino acid-specific surface area reference established by Matthew et al. (51).

ΔΔG Prediction

The ΔΔG associated with structural stability or interface binding following substitutions was calculated using the standard Rosetta (53) protocol with default parameter settings.

Conservation Score

Conservation scores were derived using sequence entropy analysis of IGPD homologs. Monomeric IGPD sequences from 21 extant species were aligned using T-coffee Expresso (52). For each alignment position X, sequence entropy was calculated as:

Where P(X) represents the probability of each amino acid at position X, estimated from the multiple sequence alignment. The conservation score for position X is defined as:

Where gap_fr quantifies the gap occurrence rate at position X.

Association of Physicochemical Features with Fitness

For each genotype, we quantified physicochemical distance from the wild-type sequence as the difference in AAindex values (AAindex distance) relative to WT across 553 AAindex descriptors. Analyses were performed within mutation sets, defined as genotypes carrying substitutions at the same positions but differing in amino-acid identity. For each mutation set and AAindex descriptor, we computed spearman correlations (ρ) between AAindex distance and corresponding fitness across genotypes. Mutation sets were required to contain at least 82 genotypes, corresponding to the minimum sample size needed to detect modest correlation (ρ=0.3) with α = 0.05 as determined by an a priori power analysis using G*Power 3.1 (54). Correlation P values were adjusted for multiple testing across all descriptors and mutation sets using the Benjamini–Hochberg false discovery rate (FDR), correlations with FDR < 0.05 were considered significant. For visualization, non-significant correlations were set to zero.

Directed-Graph Representation of Compensatory Interactions

For each substitution at position i (S_i) and each non-WT amino acid state at distinct position j (N_j), we identified pairs of genotypes that only differed at position i: G₁ = (S_i, N_j) and G₂ = (WT_i, N_j). The compensatory ability of S_i with respect to N_j was quantified as the fraction of such genotype pairs for which Δf = f(G₁) -f(G₂) > 0.5. These relationships were represented as directed edges from S_i to N_j in a network graph, with edge weights corresponding to the fraction of rescued genotype pairs. The visualizations produced for Figure 3B and Figure S3A were produced using NetworkX (55) in Python.

Quantification of Buffering of Substitutions Fitness Effects

To assess how super compensators modulate the magnitude of fitness effect of other substitutions, we considered only substitutions that produced a nonzero fitness effect. The fitness effect of a substitution was defined as the difference in fitness between two genotypes that differ only at a single position (for example, G₁=00000, G₂=00100, where 0 denotes WT amino acid and 1 denotes a non-WT amino acid). For genotype quadruplet (G1=00000, G2=00100, G3=00001, G4=00101) with fitness values f1, f2, f3, f4, we calculated |Δf1|=|f2 -f1| and |Δf2|=|f4 -f3|. Buffering of fitness effects by a non-WT amino acid state was quantified as the skewness of the distribution of |Δf1|-|Δf2|, which captures the tendency to reduce the magnitude of fitness effects of substitutions.

DMS Library Construction

To experimentally validate the predictions that super compensators can increase mutational robustness, one representative super compensator was selected, that is S189A. The sites for random mutagenesis were selected randomly, which are positions 181 and 202.

Plasmid Construction

To construct the libraries, we first identified naturally occurring, or, when necessary, inserted restriction sites that were in close proximity to the desired super compensators sites and mutagenesis sites within the S. cerevisiae HIS3 gene. Single-stranded oligonucleotides that coded for the desired amino acid substitutions (including all synonymous nucleotide variants) were synthesized with 15-25 bp homology arms flanking restriction sites.

Genomic DNA was extracted (Multi Sciences, MK010001) from the prototrophic S288C derived FY4 strain. The HIS3 locus (ATG – 250 bp, TAG + 300 bp) was then amplified using 500 ng of the extracted DNA, primers 376, 379, and the Golden Star T6 DNA polymerase (Tsingke Biotech, TSE101). The amplicon was PCR purified (Qiagen MinElute, 28004) and inserted between the SacI and KpnI sites (destroying them) within the pGREG506 ADH1pr-MCM-mCherry vectors (a gift from the Chao Tang’s lab) via Gibson Assembly (Tiangen Easygeno, VI201) to yield plasmid p142. After PCR, both amplicons were verified, purified, and assembled into the vector backbone to yield plasmid p143.

Library Construction

To construct the final libraries, 7.5 μg of p142 and p143 were digested for 4 hours at 37°C with 9 μL of KpnI, NheI (NEB, R3131S), and BstXI (NEB, R01131S) and EagI (NEB, R3505S) respectively. Each reaction was then loaded on a 0.7% agarose gel, with 0.5X TAE running buffer, gel purified (Magen, D2111), and concentrated (Tianmo Biotech, TD413). Subsequently, the library ssOligos (primers 337, 338) were then assembled into 500 ng of the digested vectors to a final concentration of 0.25 μM in 15 μL Gibson Assembly reactions (2 step Gibson Assembly, 3 minutes at 37°C, 60 minutes at 50°C). Finally, the assembly reactions were split into 5 transformations in chemically competent E. Coli (Transgen Biotech, CD101) and spread onto LB+Amp plates, and incubated overnight (∼14 hours) at 37°C. After incubation, the colonies were verified (colony PCR, Sanger sequencing) and scraped off the plates (∼55,000 colonies/library), mixed and the library plasmids were extracted (Tiangen Biotech, DP106), quantified, and stored.

Yeast Transformation and Bulk Competition

Competent yeast cells were prepared (Zymo Research, T2001) using yeast strain Y120 (derived from UCC8363; MKT1-RM SAL1-RM CAT5-RM MIP1-BY ura3Δ0 his3:KanMX). Subsequently, 5 μg of each plasmid was transformed into ∼52.5 × 10⁶ cells and incubated at 30°C, 220 RPM for 50 hrs in 1L synthetic complete dropout medium (SCD) lacking Uracil (-Ura).

After transformation (yielding ∼750,000 colonies/library), ∼3 × 10⁸ cells per library were inoculated into 1 L of SCD-Ura and grown at 30°C, 220 RPM for 6 hrs. After that, ∼3 × 10⁷ cells per library were inoculated into 1 L SCD-His (Histidine) and ∼2.1 × 10⁷ cells into 700 mL SCD-Ura and incubated at 30°C, 220 RPM. The competition was carried out for 120 hours with 12 hours between bottlenecks (2.1 × 10⁷ cells per library transferred into fresh media), two samples of 15 mL (technical replicates) were taken from each culture every 24 hours, centrifuged (3000 g, 5 mins) and stored (−20°C), the experiment from transformation to completion was carried out twice (biological replicates).

Plasmid Extraction and Illumina Sequencing

After competition, cell pellets from each time point were thawed and plasmids extracted using the Tiangen Midiprep kit (DP106). Extracted DNA (∼10 ng) was labeled and amplified using staggered primers (primers 345-356) and Q5 DNA polymerase (NEB, M0492S) in 50 μL reactions (98°C for the 30 s; 98°C for 10 s, 55°C for 30 s and 72°C for 30 s for 17 cycles; 72°C for 2 min). The amplicons were purified (Qiagen MinElute, 28004) and pooled samples were submitted to Genewiz company for Illumina library construction and sequencing.

Data Processing

Raw paired-end FASTQ files were merged using FLASH (56), demultiplexed according to the unique primer pairs used to amplify each sample at specific timepoint, and reads for each genotype were counted using starcode (57). Reads for each genotype were the sum of all synonymous nucleotide genotypes and the sum across two technical replicates. Fitness for each genotype was calculated using PyFitSeq (58), and the average fitness across both biological replicates was used for further analysis.

Prevalence Analysis of Super Compensators

DMS datasets for 217 distinct proteins were retrieved from the ProteinGym database (38). The datasets contain quantitative fitness measurement derived from diverse biochemical and cellular readouts, including protein stability, enzymatic activity, and bacterial growth rate. The established computational pipeline for identifying super compensators was then applied to these datasets to detect potential super compensators and quantify the buffering ability against substitutions.

Supplementary Figures and Tables

Fig. S1.

Fig. S2.

Fig. S3.

Fig. S4.

Fig. S5.

Fig. S6.

Fig. S7.

Fig. S8.

Fig. S9.

Table S1.

Data availability

All data and codes used for analysis and creating figures are available on Zenodo at https://doi.org/10.5281/zenodo.17769277

Acknowledgements

We thank Dr. Chu Wang and Dr. Luhua Lai at Peking University for helpful discussions and comments on the manuscript. This work is funded by China Postdoctoral Science Foundation (2025M782617 to W.H.J.), the Natural Science Foundation of Hubei Province (JCZRQNB202600450 to W.H.J), Open Project Funding of the Key Laboratory of Fermentation Engineering (Ministry of Education) (202509FE11 to W.H.J.), the Natural Science Foundation of Xinjiang Uygur Autonomous Region (2025D01A136 to H.C.Y.). Additional funds to L.B.C. are provided by Peking-Tsinghua Center for Life Sciences.

Additional information

Author contributions

L.B.C. and W.H.J. designed the study; W.H.J., S.S.K., H.C.Y., Yiparemu and L.B.C. carried out the experiments and analyzed data; W.H.J., H.C.Y., S.S.K. and L.B.C wrote the manuscript. All authors read and approved the final version of the manuscript.

Funding

China Postdoctoral Science Foundation (中国博士后科学基金会职) (2025M782617)

• Wenhong Jiang

湖北省科技厅 | Natural Science Foundation of Hubei Province (湖北省自然科学基金) (JCZRQNB202600450)

• Wenhong Jiang

湖北省科技厅 | Natural Science Foundation of Hubei Province (湖北省自然科学基金) (202509FE11)

• Wenhong Jiang

Natural Science Foundation of Xinjiang Uygur Autonomous Region 新疆自然科学基金会维吾尔自治区（新疆自然科学基金会）[中国] (2025D01A136)

• Hengchuang Yin

新疆维吾尔自治区科学技术厅 | Natural Science Foundation of Xinjiang Uygur Autonomous Region (Xinjiang Natural Science Foundation) (XJRC-2025-KJ-YJCXPT-103)

• Hengchuang Yin