De Novo Design of High-Performance Sec-type Signal Peptide via a Hybrid Deep Learning Architecture

Abstract

The rational design of signal peptides represents a fundamental bottleneck in biotechnology, where sequence optimization directly governs protein secretion efficiency and industrial scalability. Current approaches rely predominantly on natural variants or empirical mutations, constraining the accessible sequence space and limiting performance gains. Here, we develop the SPgo computational framework, which overcomes these limitations by combining rule-based domain assembly with a Transformer-enabled deep generative model to support the design of Sec-type signal peptides. Our hybrid architecture constructs optimal N- and C-terminal regions through biophysical constraints while deploying a BERT-LSTM pipeline to explore vast sequence landscapes within the critical hydrophobic core. Rigorous validation of a variety of protein targets, from fluorescent proteins to industrial enzymes and bioactive peptides, showed that SPgo-designed sequences consistently outperformed natural sequences. Most notably, SPgo was able to achieve secretory production of snake venom peptides at an unprecedented yield of 154 mg/L, a 150-fold increase in target protein yield per unit culture volume compared to traditional intracellular expression, transforming previously intractable targets into viable biotechnology platforms. This work establishes a new paradigm for computational protein design, offering immediate applications in biomanufacturing while revealing the untapped potential of artificial sequence space to surpass natural evolutionary solutions. The SPgo framework data can be found on github (https://github.com/lzlinn801/SPgo).

Industrial biotechnology increasingly demands efficient extracellular protein production to meet growing needs in therapeutics, biocatalysis, and sustainable manufacturing. − Signal peptides, short N-terminal sequences directing nascent proteins through cellular membranesa, serve as the critical gatekeepers of this process, yet their optimization remains a persistent challenge despite decades of research effort. , The apparent simplicity of these 15–30 residue sequences belies extraordinary complexity in their sequence-function relationships, where subtle amino acid changes can dramatically alter secretion outcomes and ultimately determine the commercial viability of bioprocesses.

The tripartite architecture of bacterial signal peptides comprises a positively charged N-region, a hydrophobic H-region forming transmembrane helices, and a polar C-region containing the signal peptidase cleavage site. While this structural framework provides clear design constraints, natural signal peptides display remarkable sequence diversity across these domains, suggesting vast unexplored potential for performance optimization. Traditional design strategies, however, remain fundamentally limited: rational mutagenesis explores only narrow sequence neighborhoods around known variants, while high-throughput screening approaches face combinatorial explosions that render comprehensive exploration computationally intractable. −

Recent advances in protein language models have begun to revolutionize our understanding of sequence-function relationships across diverse protein families, yet their application to signal peptide design has remained largely unexplored territory. − Most computational work in this area has focused on classification tasks rather than the more challenging generative design problem of creating new high-performance variants of. , This represents a missed opportunity to harness the full power of modern deep learning architectures for addressing one of biotechnology’s most persistent bottlenecks.

Recognizing these limitations, we developed SPgo as a hybrid computational framework that strategically partitions the signal peptide design problem according to each region’s distinct biophysical constraints and sequence complexity patterns (Figure ). Our approach exploits the observation that both N- and C-regions operate under clear physicochemical rules (positive charge distribution requirements and specific cleavage motif constraints, respectively), making them amenable to rule-based generation strategies. This targeted approach ensures adherence to fundamental biophysical principles while avoiding unnecessary computational overhead in regions where the design space is already well-constrained. The H-region presents a fundamentally different challenge that drives our hybrid design philosophy. As the primary determinant of membrane insertion efficiency, this hydrophobic core exhibits complex sequence-function relationships that resist simple rule-based approaches while demanding sophisticated modeling of local and long-range dependencies.

1.

Schematic of the SPgo hybrid framework for signal peptide design. SPgo divides the signal peptide into N-, H-, and C-regions based on their distinct constraints. Rule-based methods are applied to the N-region (positive charge) and C-region (cleavage motif), where design rules are well-defined. The H-region, due to its complex impact on membrane insertion, is modeled using data-driven approaches. This hybrid strategy ensures biophysical validity while enabling flexible optimization where needed.

To enable H-region generation, we first curated 12971 validated transmembrane α-helix segments from UniProt (KW_1133), filtering for sequences 10–25 residues in length to ensure structural relevance and consistency during training. These sequences were encoded using a pretrained BERT language model, with only the final two layers fine-tuned to preserve generalization while minimizing computational cost. The resulting embeddings, h = BERT­(x), were passed through a 10-layer bidirectional LSTM network with dropout regularization, producing contextual representations H = BiLSTM₁₀ (h). These representations were projected to token logits via a fully connected layer and trained using cross-entropy loss with the AdamW optimizer and adaptive learning rate scheduling (ReduceLROnPlateau), achieving convergence after 30 epochs. For candidate H-region generation, we applied a hybrid Top-k/Top-p sampling strategy to balance sequence diversity and physicochemical plausibility: the top 15 tokens by logit score (Top-k) and the smallest token set comprising ≥90% total probability (Top-p) were used for multinomial sampling at each position. The final signal peptides were constructed by assembling three regions: N-regions, initiated with methionine and containing 2–3 positively charged residues (fixed length of 5); H-regions generated via the model; and C-regions, which included three polar residues and the conserved “Ala-X-Ala” motif, filtered based on hydrophilicity to ensure proper membrane release. Transmembrane α-helix segment screening and further modeling details are provided in the Supporting Information.

To comprehensively evaluate the performance of each protein sequence generation model, we analyzed the generated signal peptides from multiple perspectives, including amino acid composition entropy, sequence diversity (measured by average Levenshtein distance), n-gram Jensen–Shannon (JS) divergence metrics and SignalP 6.0 score (Table ). The results demonstrate that the original model performs best across all evaluation dimensions. Specifically, its JS divergences are 0.1579, 0.3261, and 0.5037 for 1-g, 2-g, and 3-g levels, respectivelysubstantially lower than those of the other models. Its entropy value of 3.897 closely approximates that of natural signal peptides, indicating a strong ability to capture the global amino acid distribution. Moreover, the average Levenshtein distance is comparable to that of natural sequences, suggesting the model achieves high sequence diversity while preserving biological relevance. Importantly, SPgo obtained the highest SignalP 6.0 prediction rate (50.6%), suggesting that SPgo-designed sequences are not only statistically similar to bona fide signal peptides, but also more consistent with their biological functions.

1. Model Performance Comparison Table.

Model	1-g JS	2-g JS	3-g JS	Entropy	Average Levenshtein distance	SignalP Score
Real-sequences	not applicable			3.832	not applicable
SPgo	0.1579	0.3261	0.5037	3.897	21.89	50.6%
BERTOnly	0.3544	0.6526	0.869	1.558	18.52	25.3%
UniLSTM	0.3921	0.4025	0.703	0.665	25.78	10.2%
Random	not applicable					4.36%

^aNumber of sequence entries identified as Sec-type signal peptides with a predicted probability greater than 0.7 by SignalP 6.0 (per 1000 entries).

In contrast, the BERTOnly model shows limited generative capacity, with much higher JS divergences at all n-gram levels, a significantly lower entropy of 1.558, a reduced Levenshtein distance of 18.52 and the SignalP 6.0 score was reduced to 25.3%, indicating redundancy or limited token diversity in the output. The UniLSTM model performs even more poorly. Although it achieves a high Levenshtein distance, its entropy is extremely low, and its 3-g JS divergence diverges markedly from the natural distribution, and the SignalP 6.0 score is only 10.2%. highlighting its inability to learn meaningful structural features. In summary, the original model achieves an optimal balance between fidelity and diversity, demonstrating superior sequence generation capabilities and significantly outperforming the ablation models across key evaluation metrics. After performing ablation experiments for comparison, we intercepted the H-region sequence of the natural Sec signal peptide and counted its amino acid composition, which was subsequently randomized to form a new H-region. The randomly generated signal peptide was subjected to prediction, and SignalP 6.0 scored only 4.36%. This situation indicates that the SPgo model maintains the functionality of the H-region as a whole by counting the relevant amino acid compositions of the H-region while effectively constraining the amino acids at each position. To illustrate the versatility of the SPgo model, we further evaluated its potential generalizability across diverse microbial hosts. To this end, we generated batches of 1000 signal peptide sequences for four different target proteins using SPgo, and performed in silico predictions with SignalP 6.0 across various taxa, including Eukarya and other prokaryotes. The prediction success rates were comparable among these groups (Figure ), providing preliminary evidence for the versatility of SPgo-designed signal peptides across different microbial hosts.

2.

Cross-species compatibility prediction of SPgo-designed signal peptides. SignalP 6.0 was used to evaluate the predicted performance of SPgo-generated Sec signal peptides for four model secretory proteins across different biological systems. Generation success rates (per 1000 attempts) are shown for eukaryotic (black bars) and other prokaryotic systems (gray bars) including Bacillus and Corynebacterium species. The analysis demonstrates comparable prediction performance between eukaryotic and prokaryotic domains for all tested proteins (McherryT, PETase, Catalase, and α-cobratoxin), supporting the potential universal applicability of the SPgo platform across diverse microbial hosts. Data represent computational predictions of signal peptide functionality across species boundaries.

To specifically evaluate the functional impact of the H-region sequences generated by the model, we fixed the sequences of the N- and C-regions during signal peptide generation to minimize their potential influence on secretion efficiency. The sequences of the N- and C-regions were chosen based on initial random generation and remained constant across constructs. Using the red fluorescent reporter protein mCherry as the secretion target in E. coli, we generated a set of candidate Sec-type signal peptides by fusing these fixed N- and C-regions with variable H-regions to the N-terminus of mCherry. SignalP 6.0-based preliminary in silico screening revealed 14 sequences predicted with >99.9% probability to contain Sec-type signal peptides, suggesting high likelihood of efficient Sec-pathway targeting. These candidates, alongside natural signal peptides LamB and PhoA as controls, were cloned into expression vectors and introduced into E. coli for secretory expression. Fluorescence measurements of the culture supernatant revealed that a majority of model-designed sequences mediated efficient secretion, with SP3, SP4, SP18, and SP33 exhibiting up to 30-fold higher extracellular fluorescence than native signal peptides (Figure S1). This enhancement suggests that the model captured sequence determinants beyond canonical motifs, potentially optimizing hydrophobicity distribution or cleavage efficiency. Importantly, confocal fluorescence microscopy of the culture supernatant confirmed that the luminescent signal was unrelated to intact bacterial cells (Figure S2). SDS-PAGE analysis of cytoplasmic, periplasmic, and culture supernatant fractions, compared with wild-type controls, shows that the target protein is clearly present in the culture supernatant. Total protein content in the periplasm and cytoplasm is comparable to whole-cell protein levels, with no significant decrease observed. This excludes the possibility of target protein release due to cell lysis (Figure S3).

To evaluate the suitability of SPgo-designed signal peptides for extracellular expression of functional enzymes, we selected two representative enzymes: polyethylene terephthalate (PET) hydrolase and catalase, which have distinct catalytic functions. , We fused their respective coding sequences with different SPgo-designed Sec-type signal peptides and expressed them in E. coli under standard induction conditions. The purified protein concentrations were calculated to be 20 mg/L and 60 mg/L, respectively, demonstrating that the generated signal peptide exhibits high secretion efficiency. For PET hydrolase functional verification, 1 mL of concentrated supernatant was incubated with 0.1 g of PET film at 65 °C for 72 h, and the reaction products were detected by electrospray ionization mass spectrometry (ESI-MS), which confirmed the generation of key hydrolyzates- mono­(2-hydroxyethyl)­terephthalate (MHET) and terephthalic acid (TPA) (Figure S4). It is worth noting that the types of hydrolysis products generated did not change due to different signal peptides, indicating that the SPgo-derived sequence did not affect the conformation of the enzyme or the substrate recognition mode. Scanning electron microscopy (SEM) images further verified that the PET film showed significant surface corrosion and pitting characteristics under the action of the secreted enzyme, which directly reflected the effective degradation of the polymer (Figure and Figure S5).

3.

Functional test of secreted PET hydrolases. (a, b) SEM images of PET films degraded by PET hydrolases secreted by two different signal peptides generated by SPgo. (c) SEM image of PET film degraded by the control group without signal peptide.

The catalase fused with three SPgo signal peptides was also successfully secreted into the extracellular space. The secretion yields were calculated to be 170 mg/L, 190 mg/L, and 115 mg/L, respectively, indicating that its secretion level reached the medium-high expression range (Figure S6). The enzyme activity was determined after purification by metal affinity chromatography, and the results showed that the catalytic activity of the secreted enzyme was comparable to that of the free enzyme expressed intracellularly, with the activity of the free enzyme being 57630 U/mL (Table ).

2. Effect of Signal Peptides on the Extracellular Expression and Activity of Catalase.

SPgo	Extracellular yield (mg/L)	Purity (%)	Specific activity (U/mL)
SP1-catalase	170 ± 12	25.5	53218 ± 822
SP2-catalase	190 ± 8	49.2	57096 ± 535
SP3-catalase	115 ± 4	88.7	50218 ± 1173

^aRefers to the amount of target protein contained in the supernatant, and the purity was calculated by grayscale using ImageJ.

To further explore the versatility of SPgo-designed signal peptides, we applied the system to the secretion of a small, bioactive snake venom-derived peptide (cobratoxin, Uniprot ID: P01391). Snake venom peptides have important uses in neuroscience and drug development. However, due to problems such as aggregation, cytotoxicity, and low folding efficiency, it is extremely challenging to express such peptides at high levels in cells. Previous reports have documented maximum intracellular expression yields around 1.0 mg/L, limiting their biotechnological and pharmaceutical development. Using SPgo signal peptides optimized for secretion, we fused the cobratoxin peptide coding sequence and expressed the construct in E. coli. After purification by His-tag nickel affinity chromatography, the target peptide was identified by SDS-PAGE (Figure S7). The protein concentration was detected by BCA, and the highest secretion yield was 154 mg/L and the lowest was 68 mg/L, which was 2 orders of magnitude higher than the intracellular expression benchmark.

To verify correct signal peptide processing, we performed Edman degradation-based N-terminal sequencing and mass spectrometry sequencing on purified snake venom secreted peptides. The results showed that the amino acids of the secreted protein perfectly matched the expected mature sequence, confirming precise cleavage by the signal peptidase (Figure S8), this result is also consistent with our classic cutting site design (A-X-A) for C-region. Furthermore, similar to the results obtained with mCherry, SDS-PAGE analysis of cobra venom toxin in the cytoplasm, periplasm, and culture supernatant also demonstrated that the protein detected in the supernatant was not released due to cell lysis (Figure S9). This significant enhancement highlights the effectiveness of SPgo in facilitating high-yield extracellular production of small peptides that are difficult to produce by other methods. The successful SPgo-mediated secretion of snake venom peptides complements previous examples of enzymatic secretion and highlights the broad applicability of this design platform across a wide range of molecular weights and protein functional classes. Such high-level secretion of bioactive short peptides opens new opportunities for scalable production and downstream applications in drug development.

In summary, SPgo integrates rule-based design of N- and C-regions with a BERT-LSTM deep generative model to create novel signal peptides that preserve essential motifs while exploring expansive sequence space. Model like ProteinBERT and RoseTTAFold are powerful Transformer-based architectures primarily designed for generic protein sequence embedding or structure prediction tasks. While these models excel at learning global sequence representations or modeling protein folding, they do not explicitly address the controlled design of specific functional sequence regions. In contrast, SPgo is specifically tailored for signal peptide generation, incorporating a hybrid rule-based and deep learning framework that enables fine-grained control over distinct signal peptide regions with unique physicochemical properties. This region-specific control allows targeted modification and optimization of particular segments within the signal peptide, a capability not directly achievable by general Transformer-based models treating sequences holistically. Thus, SPgo provides enhanced design flexibility and interpretability by integrating heuristic domain knowledge with deep learning, enabling users to generate customized signal peptide variants with precise regional features. We propose that the high secretion efficiency of SPgo-produced signal peptides stems from two factors: (i) their enhanced ability to direct target proteins from the cytoplasm to the periplasm; and (ii) their subsequent delivery from the periplasm to the extracellular space via the type II secretion system (T2SS). The T2SS utilizes pseudopilus structures to nonspecifically pump proteins into channels formed by GspD. − As the proportion of target proteins in the periplasm increases, the proportion of proteins secreted by the T2SS into the extracellular medium also increases. While certain complex residue patterns remain to be fully captured and validation is so far limited to bacterial systems, SPgo provides a versatile platform for rationally tailoring secretion signals. Future efforts integrating automated workflows and expanding to eukaryotic hosts will broaden its impact. This approach paves the way for improved production of challenging proteins and peptides, advancing protein engineering and biomanufacturing applications.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/jacsau.5c00757.

• Reporter fluorescence under different expression conditions, confocal imaging and visual validation of mCherry secretion, subcellular fractionation confirms SPgo-mediated mCherry secretion, ESI-MS detection of PET hydrolysis products, secretion efficiency of PETase with/without signal peptide, localization and purification of SPgo-catalase, SDS-PAGE analysis of purified SPgo-secreted cobratoxin, sequencing of SPgo-cobratoxin, subcellular fractionation confirms SPgo-mediated cobratoxin secretion, and Western blot analysis of four target enzymes (PDF)

The authors declare no competing financial interest.