AI-directed gene fusing prolongs the evolutionary half-life of synthetic gene circuits

Overview of the STABLES fusion strategy

Here, we introduce STABLES, a comprehensive approach to enhancing evolutionary stability through gene fusion. It is host and GOI agnostic and robust to many mutations and provides a generic, systematic, simple framework. Our design includes the following components (Fig. 1):

1) The GOI to be expressed in the host organism.

2) An EG selected for optimal gene expression and mutational stability using an ML model. This model is based on meaningful bioinformatic features and empirical data in S. cerevisiae. The GOI and EG are expressed on a shared promoter, on a single ORF, where the GOI’s C terminus is fused to the EG’s N terminus.

3) The linker is selected to minimize disruption to protein folding by comparing disorder profiles of the GOI and EG before and after fusion using biophysical models. A commercial linker yielding a minimal change is chosen (see Methods) (53–60).

4) The fusion gene is optimized for gene expression and the avoidance of mutationally unstable sites (21, 61–63). This includes optimization of the GOI, linker, and depending on use case, the EG as well.

5) A leaky stop codon is placed after the GOI. This is a stop codon with a positive rate of read-through. This leads to the generation of two proteins—either the GOI alone or the fusion protein. The rate of expression for the two proteins can be controlled through the selection of an appropriate rate of read-through (49–52). A codon is selected such that the fusion protein is produced in barely viable quantities for the host’s growth while maintaining much higher expression in favor of the GOI’s protein alone—as this is the derived, relevant product. By ensuring just barely viable quantities of the fusion protein, the host’s mutational stability is further enhanced, as many more mutations would prove deleterious.

6) The EG in its native form is deleted from the host and replaced by this gene fusion. The host is now dependent on the fusion protein to provide the original EG function. Many mutations that would have reduced production or caused the misfolding of the GOI, whether in the GOI or promoter, would now reduce the production of the fusion protein beneath viable quantities. Because of the host’s dependence on this protein, this leads to host lethality, and these mutations do not take hold in the population.

Fusion strategy improves evolutionary stability

To assess the impact of the fusion strategy, we evaluated the stability of green fluorescent protein (GFP) fused to various EGs by examining 10 strains from a previously described library of N-terminally GFP-tagged genes in S. cerevisiae (65, 66). Fluorescence was used as a proxy for expression for 15 days.

This experiment was conducted before the creation of the EG selection mechanism—it was designed to give an indication of the need for such a model and its potential impact. For this purpose, meaningful bioinformatic features were generated for all EGs, and they were clustered in many clustering configurations. Ten strains were selected such that they were consistently classified to different clusters and were near the centroids of these clusters. This generated a set of strains that were highly varied between them while being representative of many similar genes. They were compared with a baseline strain of unfused GFP (see Methods).

Fluorescence intensity was used as a proxy for functional GFP levels, following established SWAp-Tag library protocols (65, 66). Given that GFP fluorescence only occurs after proper β barrel folding and chromophore maturation (67–71), it directly reflects the abundance of correctly folded, functional protein. This approach has been validated in large-scale GFP-tagging screens (65, 66, 72, 73) and supported by studies showing that fluorescence changes arise from protein folding and production rather than intrinsic chromophore brightness (74). By contrast, Western blotting detects all GFP-containing species, including misfolded forms, and is prone to technical variability such as antibody binding and transfer efficiency (75–79).

The experiment (Fig. 2A) validated the following conclusions, emphasizing the need for an EG selector:

1) Most strains demonstrated decline over the course of the experiment, emphasizing the existence of mutational instability and the need to improve it.

2) GOI-EG fusions exhibited slower declines in fluorescence compared to unfused GFP, confirming that the fusion of genes enhances stability.

3) Different EGs yielded varying degrees of stability, demonstrating that the stability of a fused gene depends on the EG selected.

4) One gene displayed a statistically significant (Student’s t test, P ≈ 0.047) advantage over unfused GFP, emphasizing the need for more informed gene selection.

For statistical analysis, see Supplementary 1.

ML predicts optimal EG-GOI combinations

The variability in stability observed across different EGs highlighted the importance of systematic EG selection. To address this, we developed an ML model to predict EG-GOI fusions that maximize expression and stability (Fig. 2B; see Methods). The model was trained on fluorescence data collected from GOI-EG fusion libraries under various conditions in S. cerevisiae (see Supplementary 2) (65, 66). As the fluorescence was measured after the variants had time to mutate, this is assumed to capture a combination of both expression and stability.

Using features such as codon usage bias [tRNA adaptation index (80) and codon adaptation index (81)], GC content (82), mRNA folding energy (83, 84), ChimeraARS scores (85), and other meaningful bioinformatic features, the model successfully ranked potential fusion pairs.

In a reasonable use case, a user would generate and use very few genetic designs. The model would recommend one to three EGs, which the user would validate experimentally, and proceed with the design exhibiting best performance. As the models were trained in cross-validation, each EG received a score equal to the quantile of its expression within the test set (e.g., 1.0 for the EG with the highest expression). Models were evaluated both on their expected performance (median score of EGs recommended by the model among cross-validations) and on their robustness (likelihood of recommending an EG with a low score). For performance measurement with more common and less relevant metrics, see the Supplementary Materials.

On the basis of these evaluations, an ensemble model combining k-nearest neighbors (KNN) (86) and XGBoost (XGB) (87) was selected and trained. The KNN model exhibited a high median score, while the XGB model improved the model’s robustness. Selecting the best performance among top 3 candidates, the median score was 0.995, and the scores were above 0.98 ( $P<0.05$, Fig. 2C). When selecting only the top performer, the median score was 0.939, and the scores were above 0.92 ( $P<0.001$, Fig. 2D). These results underscore the predictive power of the ML model and its ability to systematically identify highly performing gene fusions, enhancing the efficiency of fusion design. Feature importance was calculated for the features for further insight (Fig. 2E).

To evaluate whether EG selection depends on the GOI, we first assessed universal features of high-performing EGs. The union of the top 50 EGs across GFP, RFP (red fluorescent protein), and human proinsulin exhibited shorter lengths [mean: 962 nucleotides (nt) versus 1354 nt; P < 1 × 10⁻⁸], higher codon adaptation [tRNA adaptation index (tAI): 0.568 versus 0.360; P < 1 × 10⁻⁵⁶), and slightly higher GC content (43.4% versus 40.3%; P < 1 × 10⁻¹⁹) compared to the yeast genome, indicating that compact, codon-optimized EGs are universally favorable.

We then compared the top 10 EGs for each GOI to assess specificity. The overlap among the highest-ranked EGs was limited: No gene appeared in the top position for more than one GOI, and only GFP and insulin shared two of five top EGs (Jaccard: 0.25), while other pairs showed no top 5 overlap. At larger sets, moderate overlap emerged (e.g., GFP-insulin top 20 overlap: 14 of 20; Jaccard: 0.54), suggesting that certain EGs are broadly strong performers, but the very top predictions remain GOI-specific.

Feature analysis revealed that RFP-optimized EGs were longer and more codon adapted than GFP and insulin EGs (length: P < 0.001; tAI: P < 0.005), while the GC content did not differ (P > 0.4). All top 10 EGs fall within the top 4% of genome-wide tAI scores (quantile: ≥0.96), highlighting strong codon optimization across targets. These findings underscore the value of ML-based ranking to prioritize optimal, GOI-specific EGs.

Validation with proinsulin production

To demonstrate the real-world applicability of our approach, we applied it to stabilize human proinsulin expression in yeast, a biotechnologically relevant system. EGs were selected for high performance in both the XGB and KNN models and additional engineering needs (see Methods). Thus, our model identified two EGs—CAF20 and ARC15—as suitable fusion partners for proinsulin.

A 30-day in-lab evolution experiment was conducted, where expression level measurements were taken every 5 days, using the enzyme-linked immunosorbent assay (ELISA) protocol. The strains tested were the original proinsulin (as patented by Novo Nordisk), the proinsulin as optimized by the Evolutionary Stability Optimizer (ESO) (21), and the optimized sequence fused to CAF20 and ARC15 as fusion genes (Fig. 3A). As in the 10-gene experiment, the unfused proinsulin replaced CAN1.

The need for a leader secretion sequence in proinsulin production in yeast has been well documented (88–94). However, it was not clear whether it may disrupt the efficacy of our design. We measured the expression at initiation for the original proinsulin and two fusion genes with and without a leader sequence (Fig. 3B). All variants exhibited much higher expression in the presence of a leader sequence, and thus, all further experiments were conducted as such.

The expression patterns approximately followed an exponential decay pattern, where the fused genes displayed a much slower decay rate (Fig. 3C). This enabled us to estimate the cumulative expression of proinsulin over time. These quantities were normalized by the expression estimated for 10 days for the original proinsulin. The gene fusions showed a fivefold increase in total proinsulin yield over the experimental period (Fig. 3D). ARC15 demonstrated better performance for shorter durations (higher initial expression), while CAF20 demonstrated better performance for longer durations (slower decay)—the selection of the optimal gene will depend on the industrial use case.

For the fusion gene with ARC15, we measure 98.5 mg/liter at initiation, where for the original sequence, we measure 72.9 mg/liter—a ~35% increase at the initial expression. For the fusion gene with ARC15, we measure 68.7 mg/liter after 30 days, meaning that ~70% of expression is maintained. For the original sequence, we measure 4.8 mg/liter, meaning that only ~7% of expression is maintained.

Nanopore sequencing was conducted on the final sequences. It reveals that following the experiment, the fused proinsulin suffered few mutations, while the unfused proinsulin sequence was lost completely (Fig. 4A) (95–98). This has been further validated by Western blotting (Fig. 4B).

We identified several mutations in the CAF20 and ARC15 fusion constructs through nanopore sequencing (Fig. 4A and table S4). Notably, all mutations within the proinsulin region were present at similar frequencies both before and after the 10⁵-generation evolution experiment, suggesting that they were introduced during cloning or represent sequencing artifacts.

This includes a shared 1–base pair deletion (~20% frequency) in a noncoding spacer between the insulin preleader and the B subunit found in both the CAF20 and ARC15 constructs; this ostensible deletion of adenine occurs in a homopolymer run of eight adenines, an area highly prone to sequencing error—making this deletion likely to be an artifact.

The ARC15 construct contained no additional mutations within the proinsulin region. The CAF20 construct, however, exhibited five silent single-nucleotide polymorphisms in the proinsulin coding region; these mutations were likely introduced during cloning, because they appeared in very high frequencies (>99%), and were stable over time. We also observed a 2-nt deletion in codon 9 of CAF20, which increased in frequency from ~40 to ~80% over the course of the experiment and is located within a homopolymer run of five thymines, meaning that this deletion is likely to be at least partially a sequencing error. In addition, the relatively high proinsulin protein levels we observed during the experiment support the conjecture that this is completely or predominantly a sequencing error. This mutation, if it is real and not a sequencing error, introduces a premature stop codon. Despite potential sequencing noise, the observed frequency shift suggests that it may be under positive selection.

Leaky stop codon enables GOI production and overexpression

To enable the generation of both the fusion protein and GOI alone, we used leaky stop codons, which allow partial translational read-through (49–52). The ability to enforce low but nonzero read-through rates is a key factor in improving the stability and industrial viability of our design. If the fusion protein’s abundance is reduced to barely viable, then it is highly likely that any mutations affecting the GOI’s expression would lead to host lethality, promoting mutational stability. This is in tandem with the enforcement of much higher expression of the isolated GOI, which is the necessary and desired product.

Thus, informed selection of the stop codon is necessary. To enable this informed decision, we conducted an experiment, creating a fusion protein of blue fluorescent protein (BFP) and mCherry, with different leaky stop codons in between. This was expressed on a plasmid (Fig. 4C). The red fluorescence is expressed only in the fusion protein, while the blue fluorescence is expressed in both the fusion protein and BFP alone. Thus, the ratio between the fluorescence measurements can be used to derive the rate of read-through (Fig. 4D).

We selected the three stop codon designs expressing the lowest mCherry fluorescence (to minimize read-through), which was still detectable (to increase cell viability). We conducted another evolution expression, attaching the proinsulin gene to ARC15, either without a stop codon or with one of the three selected. This is in accordance with the STABLES design. Using a similar protocol, the expression of proinsulin was measured over 50 days, this time capturing the presence of both the fusion protein and isolated proinsulin. In addition to the secretion of isolated proinsulin (which is a requirement in and of itself), we observed a much higher expression rate and cumulative expression accordingly (Fig. 4E).

For the fusion gene without a stop codon, we measure 102 mg/liter at initiation, where for the best candidate L3, we measure 98 mg/liter—a ~4% decrease at the initial expression. For the best candidate L3, we measure 83 mg/liter after 50 days, meaning that ~85% of expression is maintained. For the fusion gene without a stop codon, we measure 15 mg/liter, meaning that only ~15% of expression is maintained.

Organism-agnostic strategy to tackle evolutionary instability

Previous studies proposed methods for increasing the mutational stability of the GOI. Some of these methods either are dependent on host-specific pathways or require design and implementation of complex parts (e.g., biosensors and overlapping genes), which may not be possible. Others require developing multiple strains and managing population dynamics or provide robustness only to specific mutation types (e.g., repetitive parts and mutations in the promoter). This study introduces STABLES, a simple, robust, generic, host and GOI-agnostic strategy to address evolutionary instability in synthetic biology. By fusing the GOI to an EG on the same ORF under a shared promoter, many deleterious mutations would lead to host lethality, leading to higher GOI mutational stability. By incorporating a leaky stop codon, we enable these gene fusions to generate high expression of the GOI alone while increasing their mutational stability. These, together with informed selection of the linker, sequence optimization, and the use of a leader sequence, are key in the improved stability and performance observed in our experiments.

We find that following all our design principles, we observe a 15% decline in the expression of proinsulin over 50 days. This contrasts with the original design, which declined by 93% over 30 days. This is in addition to a ~30% increase in the initial expression.

Our ML tool enhances the utility of this approach by selecting better GOI-EG pairings on the basis of biologically meaningful features. Translational efficiency metrics, such as tAI, emerged as dominant predictors. A possible explanation for this result may be related to the fact that higher translation efficiency tends to be associated with genes that are highly expressed, have higher mRNA levels (because of higher mRNA stability, among other factors), and are more conserved and thus tend to fold more efficiently (99–103). Other features include GC content, RNA folding energy, alternative shifted ORFs, and amino acid composition that are probably also associated with higher expression, mRNA stability, and robust protein folding.

Protein disorder predictions were used to guide linker selection, increasing the likelihood that the gene fusions maintain native protein folding and functionality (see Methods). Combined with the avoidance of mutational hotspots and sequence optimization for expression and stability, these elements ensure the practical applicability of the approach across a broad range of use cases.

Although our experiments focus on S. cerevisiae, the STABLES framework is broadly generalizable to other organisms. First, the general idea of coupling a heterologous GOI to an essential gene is applicable to any living cell. If performed accurately, such a coupling will induce an addiction of the cell to the fused protein, which will prolong the half-life of the GOI. Moreover, its computational components are designed for cross-species applicability: The essential gene selection model uses organism-agnostic, sequence-based features; linker design relies on biophysical and ML models of protein disorder, which are host-independent; and codon optimization adapts directly to the target organism via the tAI maximization. Leaky stop codons, a key feature of STABLES, have been reported in diverse taxa, including bacteria (104, 105), fungi and animals (49), insects and nematodes (106), and mammals (107, 108). Together, these properties make STABLES a flexible and transferable strategy requiring minimal host-specific adaptation.

Future directions

Using our ML model, we have already seen empirical evidence for good performance in predicting high-performing GOI-EG pairs for different GOIs, as shown in our experimental validation. Its reliance on sequence-based features enables effective application to nonmodel organisms, even in the absence of extensive empirical data. However, expanding the dataset used for training the model remains a valuable avenue for further enhancement. Systematic testing for libraries from more host organisms and a broader range of target genes would improve the model’s generalizability and ensure its utility across an even wider variety of synthetic biology applications. Furthermore, in many organisms, epigenetic silencing must be considered when optimizing expression and stability (24, 109–116). While we have designed our model with capabilities to avoid epigenetically silencing motifs, this application should be tested in vivo.

While our method has shown much promise, further experimental validation is needed to refine certain aspects. For example, empirical testing of linker selection will help confirm computational predictions and guide future improvements. An in-depth analysis and testing of the leaky stop codons would also enable further optimization.

Broader implications

This study demonstrates the potential to stabilize synthetic genes in both industrial and environmental contexts. In biomanufacturing, the ability to maintain stable GOI expression over extended periods can reduce costs, improve scalability, and simplify regulatory processes. In environmental applications, robust gene fusions could support long-term deployment under dynamic and uncontrolled conditions, enabling breakthroughs in bioremediation and biosensing.

Overview of gene fusion design

We developed a multistep workflow to design and optimize GOI-EG fusion genes for enhanced evolutionary stability and expression. This process included the selection of EGs using an ML model, the identification of optimal linkers to minimize misfolding, sequence optimization of the synthetic gene to maximize expression and stability, and addition of a low–read-through leaky stop codon.

ML model for EG selection

The ML model was trained on fluorescence datasets from yeast libraries (see Supplementary 1) (65, 66). The dataset included GOI-EG fusion genes for 5185 EGs (~78% of all EGs in yeast), fused with either GFP or mCherry. Fluorescence was measured across multiple time points.

Features and model architecture

The model utilized a diverse set of bioinformatic features derived from the sequences of each GOI and EG. Key feature families included the following:

1) Codon usage bias: (i) tAI: calculated for the full sequence, first 17 amino acids, and sliding windows across the GOI and EG (80); (ii) relative codon adaptation (RCA): evaluated for the GOI, EG, and sliding windows (117); (iii) codon adaptation index (CAI): assessed for the full sequence and sliding windows (81); (iv) effective number of codons: measures diversity in codon usage (118).

2) Sequence composition: (i) GC content: calculated globally and in sliding windows for both the GOI and EG (82); (ii) k-mer frequencies: counts of nucleotide or amino acid substrings of lengths 3 to 5; (iii) amino acid frequency correlation: Spearman correlation between GOI and EG amino acid compositions.

3) Thermodynamic properties: (i) local folding energy: calculated using ViennaRNA (119) in sliding windows (50 base pairs).

4) Chemical properties: (i) molecular weight and hydrophobicity: computed for both GOI and EG sequences (120); (ii) isoelectric point: assessed at physiological pH.

5) Translation context: (i) start codon context: features describing the nucleotide flanking regions of start codons; (ii) shifted ORF length: alternative reading frame lengths.

6) ChimeraARS score: (i) The ChimeraARS score quantifies sequence similarity to a reference set of genes with high codon usage bias. This metric enhances predictions of sequence stability beyond standard codon usage features (85).

The ML pipeline used an ensemble model combining KNN and XGB architectures (Fig. 3A). The KNN model emphasized features of the EG and provided high median performance—but was prone to failures. The XGB model emphasized features of the GOI, EG, and interaction features and provided much higher robustness to failure. Hyperparameter tuning and cross-validation were conducted using Optuna to ensure robust performance (details in the Supplementary Materials).

Feature importance in ML predictions

Feature importance was analyzed using SHAP (Shapley Additive Explanations) (121) for XGB and forward feature selection for KNN. SHAP provided a global ranking of feature contributions, while forward selection iteratively identified the most impactful features for KNN predictions. The XGB results, which were more robust, provided the primary conclusions, while KNN analysis offered additional insights where consistent or complementary.

XGB feature importance

The SHAP analysis (Fig. 3D) revealed the tAI (80) as the dominant feature, greatly outpacing all others. Variants with high tAI scores consistently showed enhanced stability and expression, underscoring its central role in optimizing translation. The influential features were as follows: (i) GC content (82) and RCA (117), both reflecting translational optimization and sequence stability; (ii) amino acid frequency correlation between the GOI and the EG, suggesting reduced metabolic burden for similar protein compositions; (iii) local folding energy (119) downstream in the GOI and near the initiation site, emphasizing the importance of avoiding unstable mRNA secondary structures; and (iv) shifted ORF length (122–124), ribosomal coverage bias score (125), and tAI in the first 17 amino acids (126–129), which further highlighted the importance of efficient translation and successful initiation.

KNN feature importance

KNN forward selection supported the importance of codon usage–related features, particularly in the EG. Key features included RCA, tAI, and CAI (81) metrics within the EG and in specific regions, such as the first 17 amino acids. These findings complemented the more GOI-focused insights from XGB by emphasizing the role of the EG in the overall gene fusion stability.

Linkers for protein fusion

Protein misfolding remains a challenge in gene fusion strategies (130–133). While tools such as AlphaFold (134) offer an accurate prediction of protein structure, they are too computationally intensive to enable a reasonable comparison between many linkers. Rather, we used tools such as IUPred2A (84), a biophysical model, and MoreRONN (83), an ML model, to predict protein disorder profiles and assess the impact of linkers. As protein disorder profiles have a profound effect of protein folding, it is taken as a proxy variable—linkers with a smaller effect on the disorder profile are less likely to influence the protein folding and thus less likely to cause misfolding (57, 58, 135, 136). Linkers were selected to minimize disruptions to the disorder profile upon fusion, likely preserving native folding patterns of both the GOI and EG (53–59). This minimization was applied by calculating the Euclidean distance between the disorder profiles before and after fusion for the 1280 linkers in a linker database (59). The linker inducing the smallest distance was selected. Although experimental validation of the linker selection step remains pending, these predictions provide an essential foundation for mitigating misfolding risks in future applications.

Scoring and selection

Linkers were scored on the basis of the L2 distance between disorder profiles of the unfused and fused states. The optimal linker was selected to minimize disruptions to native folding patterns, preserving the functionality of both the GOI and EG. The following is the calculation

where ${\mathbf{G}}_{\mathrm{\ell }}\left(x\right)$ is the disorder profile of gene $x$ when fused with linker $\mathrm{\ell }$ , $\mathcal{H}$ is the set of commercial linkers reported in (59), and ${\mathrm{\ell }}^{\ast }$ is the linker yielding the minimum total disruption.

To compute disorder profiles, two complementary models were used to assign a per-residue disorder probability. IUPred2A (84) predicts disorder from estimated interresidue interaction energies, providing context-sensitive scores (0 to 1) for both short and long disordered segments. MoreRONN (83) applies a neural network trained on curated datasets to assign disorder probabilities via a sliding-window approach, enabling sensitive detection of flexible regions relevant for fusion design.

Optimizing DNA sequences for stability and expression

Sequence optimization was performed to enhance the stability and expression of the GOI-EG fusion genes using the ESO (21) pipeline. ESO integrates DNAChisel (137), a tool that optimizes genetic sequences while maintaining biological constraints. ESO has been previously validated for its ability to improve the evolutionary stability of synthetic genes. The ESO pipeline focuses on optimizing codon usage, minimizing mRNA instability, and ensuring the long-term expression of engineered genes by accounting for evolutionary pressures.

Optimization objectives

Experimental validation

This PDF file includes: