Has Adaptive Evolution of SARS-CoV-2 Hit a Ceiling?

1. Accumulation of Fixed Mutations Is Slowing Down

While the majority (over 60%) of the ~30,000 nucleotides of the SARS-CoV-2 genome have mutated at least once [4,7], only a small proportion of the mutations have been selected and fixed [8]. Most synonymous mutations are near neutral, while most nonsynonymous mutations are deleterious [6]. Most of the fixed mutations are in the Spike protein. By the beginning of 2025, SARS-CoV-2 had accumulated over 60 mutations in the 1273-residue-long Spike protein, 30 of which in the 223-residue-long RBD [9,10]. On the other hand, there are altogether about 100 fixed nonsynonymous mutations in the SARS-CoV-2 genome coding for 29 proteins totaling over 9000 amino acid residues [10]. In other words, the overall rate of fixed amino acid mutations is ~1% while there are ~5% fixed mutations in Spike and over 13% in the RBD.

1.1. Cooperative Emergence of Multiple Affinity-Enhancing Mutations Followed by Immune Escape Mutations

Buildup of missense mutations in the Spike has been gradual with two brief periods of acceleration, one at the emergence of Omicron BA.1, and the other at the emergence of Omicron BA.2.86 (Figure 1C) [10]. The sudden increase in missense mutations in BA.1 also manifested as an increase in the ratio between nonsynonymous mutations and synonymous mutations (dn/ds or ω, Figure 1D).

Although many missense mutations in the Spike protein experienced positive selection [8], only about a dozen of them enhanced receptor binding affinity [11,12,13,14,15]. Synergism and epistasis among affinity-enhancing mutations played a major role in the mutation count jumps, meaning certain mutations facilitated the fixation of other mutations, generating new modes of interactions between the RBD and ACE2 [2,17,26]. For example, the N501Y mutation turned the affinity-reducing Q498R mutation to an affinity-enhancing mutation. Synergism between N501Y and Q498R reconfigured the RBD-ACE2 interaction in the Omicron variant [18]. Synergistic affinity-enhancing mutations were often followed by immune escape mutations which provided more prominent selective advantages than affinity enhancement [12,19,20,21]. The affinity-enhancing mutations in BA.1 (S477N, Q493R, Q498R, N501Y) first appeared in the original Omicron variant (B.1.1.529) and the affinity-reducing mutations (including three deletions) were added subsequently for immune escape, enabling BA.1 to spread quickly, displacing the Delta variant in a few months (Figure 1A) [10,12].

Likewise, R403K did not increase ACE2-binding affinity in the earlier B.1.1 lineage but did in the BA.2 lineage background [22]. It first appeared in BA.2.86. Although BA2.86, with its high receptor-binding affinity, was recognized for its increase in frequency, it never rose to dominant status (with a peak frequency of ~9% in the US [23]). However, one more mutation, L455S, turned BA2.86 into JN.1, which quickly wiped out all other subvariants to dominate the variant landscape (Figure 1 A and B). L455S significantly reduced the ACE2-binding affinity of the RBD but provided JN.1 with strong immune-evading ability along with enhanced fusogenicity and improved cell entry [20,21,24].

Similarly, Q493E enhanced ACE2 binding only in the presence of L455S and F456L, found in some JN.1 sublineages. Q493E emerged independently in KP.3 and LP.8 [17]. Subsequently, a deletion, S31del, developed in KP.3.1 to produce KP.3.1.1. Although the deletion itself reduced the ability of the virus to infect cells by membrane fusion, it granted KP.3.1.1 a significant competitive edge over its parent by immune evasion [25,26]. S31del also independently emerged in other JN.1 sublineages such as KP.2.3 and LP.8, showing that S31 was a prominent immune target in the human population during the past year.

1.2. Convergence, Flip, and Reversion

Even though the Spike protein experienced strong positive selection [8], natural mutations in the Spike during the pandemic indicated that there are limited modes of receptor binding that the virus can evolve into. For this reason, mutations in the S gene tend to be convergent, recurrent, and cyclical. Table 1 lists some of the recurrent Spike mutations. Multiple variants resorted to the same tricks again and again.

The FLip mutants showed that the virus can swap positions of two adjacent amino acids in the Spike protein to evade neutralizing antibodies and to improve receptor binding. Multiple XBB and JN.1 sublineages swapped the positions of L455 and F456 via the FLip mutations (L455F + F456L) [16,27]. Both mutations, especially L455F, dampened receptor binding when acting alone [16], and F456L also reduced viral infectivity [15], which is probably why they did not develop early in the pandemic. However, more than three years later, F456L showed up, presumably because it enabled the virus to evade antibodies developed against earlier lineages. Subsequent L455F mutation effectively restored receptor binding affinity and further helped the virus to penetrate herd immunity [16].

RBD mutations also reverted depending on the genetic background. The Q493R mutation in BA.1 and BA.2 was found to enhance receptor binding [28]. However, the mutation reverted in BA.2.75 and its XBB sublineages as well as in BA.2.86 ant its JN.1 sublineages. The reversal reduced receptor-binding affinity but was allowed in the context of the overall strong binding of the BA.2.75 and BA.2.86 lineages for the purpose of immune evasion. Beside Q493, six other amino acids in the RBD are known to experience “Yo-Yo” mutations. Notably, G446 and N501 had mutated between two forms three times, and L452 four times by August of 2023. Meanwhile, deletion of HV69-70 in the N-terminal domain of the Spike had appeared and disappeared three times within the same time frame [30].

1.3. Natural Restrictions on the Improvement of Receptor-Binding Affinity

Although in vitro evolution produced an RBD with a 1000-fold increase in receptor-binding affinity, the highest affinity we have observed in natural variants was no more than 20-fold stronger than wild type. The strongest receptor-binding virus in various reports was BA.2.75, XBB.1.5, or BA.2.86, depending on measurement methods and samples analyzed [14,19,20,21,24,31]. Figure 1F shows dissociation constants (Kd) of the major variants according to and [20], with lower Kd indicating higher binding affinity. The discrepancy between in vitro evolution and the actual in vivo outcomes attests to the differences between isolated molecular interactions on the surface of yeast cells and natural selection of replicating viruses with competing functional priorities in the host cell and the need to compromise under immune pressure. For example, the D614G mutation, which became dominant globally within the first few months of the pandemic [32], resulted in an RBD that bound less tightly to ACE2 than wild type, but the mutation increased the density of intact Spike trimers on the viral surface by preventing premature dissociation of S1 from S2 following furin cleavage in the producing cell [33,34]. As discussed earlier, after high-affinity RBDs were produced, they were quickly deoptimized by immune escape mutations that gave the lineages bigger advantages in human populations. The current trend shows an overall tendency of decrease in receptor-binding affinity (Figure 1F) [20,21,25,35].

Of the nine RBD mutations in the optimal in vitro evolution product, RBD-62 [2], five of them are found in the current LP.8.1* lineage. Interestingly, although V445K in RBD-62 has never been reported in nature, mutation of V445 to another basic amino acids, histidine, has been fixed in the BA.2.86/JN.1 lineage. Recently, H445 was replaced by the third basic amino acid, arginine, in LP.8 and enhanced its receptor-binding affinity [35]. Another mutation in RBD-62, I468F, was found in early variants but was not fixed because it rendered the virus more susceptible to antibody neutralization [36]. The other two mutations in RBD-62 (I358F and T470M) have never been reported in natural variants, suggesting that these residues must be conserved for effective viral replication or for infection of the human host. Two mutations found in another in vitro evolution product with improved ACE2-binding (RBD-71), R408D and K417V, have not been reported in nature, but R408S and K417N/T have been fixed in multiple lineages. The fact that the mutations are tolerated suggests that chances of R408D and K417V emerging in the future exist, but they are unlikely because aspartic acid (D) is categorically different from serine (S), so is valine (V) from asparagine (N) and threonine (T). Nevertheless, it is worthwhile to analyze the impact of D408 and V417 in the context of the current JN.1 sublineages.

2. Deleterious Mutations Lead to Degeneration and Attenuation

Even with enormous population sizes, exponential growth, and strong natural selection, viruses still experience significant genetic drift [37]. Viruses are known to be subject to Muller’s ratchet, the process of irreversible accumulation of deleterious mutations due to random genetic drift, especially in asexual organisms, leading to reduced fitness [37,38]. Beside the high error rates of viral RNA-dependent RNA polymerases (RdRp) [39], mammalian cells use mutagenesis as a defense mechanism to dampen viral fitness [3,4,5,38]. All the Variants of Concern (VOC) derived directly from the ancestral virus instead of evolving from a previous VOC, suggesting lowered adaptability of the variants compared to the ancestral virus. The ancestral virus and most of the early variants are long gone from the human population, leaving the extant subvariants at the mercy of Muller’s ratchet, even with occasional recombination between circulating lineages.

2.1. Declining Mutational Fitness Effects

Bloom and Neher estimated the fitness effects of SARS-CoV-2 mutations by comparing independent occurrences of each mutation to an expected number based on mutation rates at the third nucleotide of four-fold degenerate codons. They found that the overall effects of nonsynonymous mutations in each gene were negative. Moreover, plotting their calculated average fitness effects of nonsynonymous mutations in the RdRp gene and in the S gene over time, we show that the overall fitness effects of missense mutations declined during the reign of Omicron from the beginning of 2022 to the end of 2024 (Figure 1E), which is consistent with the observation of Maiti et al of a dn/ds decline within the Omicron VOC [8]. In other words, further mutations are more and more likely detrimental to viral fitness than being adaptive.

The large genome of SARS-CoV-2 contains regions where natural selection is relaxed. Bloom and Neher’s study also found that fitness effects of nonsynonymous mutations in the accessory proteins were comparable to those of synonymous mutations, i.e., near neutral, consistent with the observed accumulation of missense and nonsense mutations in these genes [8,40].

2.2. Declining Codon Adaptation Index in the Human Host

Beside functional improvements and immune escape, another direction of viral evolution after entering a new host species is optimization of codon usage in the cells of the host, which involves synonymous changes. Selections of nonsynonymous mutations typically take priority over codon usage optimization because the fitness effects of the latter are more subtle. Therefore, it is conceivable to expect sacrifices in codon usage in genes where selective pressure for adaptive evolution is high. Codon adaptation index (CAI) of SARS-CoV-2 in the human population experienced a decline in the early variants, restored in Omicron [41,42,43,44,45], and decreased again [44]. After analyzing the CAI and the codon pair adaptation index (CPAI), Padhiar et al reported no drastic net changes in SARS-CoV-2 genes in codon usage from December of 2019 to July of 2024. However, when we plotted the temporal curves with the supplementary data of Padhiar et al, we found that the S gene experienced two phases of decline in both CAI and CPAI with a sharp increase in the middle, and there was an overall negative correlation between CPAI and time (p = 0.0028, Figure 1G) [46]. This suggests that evolution of the S gene not only has to strike compromises between receptor-binding and immune evasion but also must bear the cost of deoptimized codon usage. Timing of the jump in CAI and CPAI in the S gene coincided with the emergence of Omicron. The large number of adaptive mutations and simultaneous optimization of codon usage in the original Omicron variant suggest a unique evolutionary mechanism which is still a puzzle today [38]. Using the same data set, we also observed significant codon degeneration in the ORF1ab (which includes the RdRp gene) and the N gene, where multiple adaptive mutations have been documented [40,47,48,49]. The CPAI of the whole genome also showed a decreasing trend, although not statistically significant (p = 0.15).

2.3. Phenotypical Attenuation

Accumulation of deleterious synonymous and nonsynonymous mutations resulted in attenuation of SARS-CoV-2 as the pandemic unfolded. There was a decline in replication efficiency in cell cultures from the early B.1.1 variant to Omicron [26,50]. The Omicron subvariants, including some dominant XBB family members, BA.2.86, and JN.1, were less pathogenic in hamsters and in mice than the ancestral B.1 variant [24]. Later Omicron subvariants were less virulent than earlier Omicron subvariants, although BA.2.86 and JN.1 demonstrated more efficient replication in human nasal epithelial cells [22,24].

Attenuation of SARS-CoV-2 in humans manifested as reduced mortality rates over time. There seemed to be a rapid decline in mortality rate during the first few months of the pandemic [51]. COVID-19 related mortality in the United States has been declining continuously even though viral prevalence remains high as evidenced by high test positivity rates (Figure 1H) [52]. The mortality decrease does not correlate with vaccination efforts, as intensive vaccination in 2021 failed to prevent the death peaks later in the year, and the federal government completely stopped free vaccination in September of 2023 [53]. The mortality rate correlates partially with natural immunity. After the Omicron wave in the winter of 2021-2022, most Americans had turned seropositive [54], and mortality dropped dramatically ever since. However, the subsequent gradual decline from 2022 till now probably has more to do with viral attenuation than with herd immunity because of declining vaccine coverage and rapid immune escape of the virus.

A global meta-analysis revealed that the global case fatality rates of the ancestral virus, the Alpha, Beta, Gamma, Delta, and Omicron VOCs were 3.64%, 2.62%, 4.19%, 3.60%, 2.01%, and 0.70%, respectively [55]. Note that the more lethal Beta and Gamma variants emerged before Alpha, and Delta was not as virulent as initially reported in Asia [56]. The case fatality rates of the ancestral virus, the Alpha, Delta, and Omicron VOCs in North America were 4.77%, 2.67%, 2.50%, and 0.73%, respectively, which is roughly consistent with Figure 1H considering the lag time between infection and death and the different prevalence rates of the variants [55]. Viral attenuation and immunity buildup probably both contributed to the decrease in case fatality rate from the original virus to Omicron.