ProDualNet: dual-target protein sequence design method based on protein language model and structure model

Abstract

Proteins typically interact with multiple partners to regulate biological processes, and peptide drugs targeting multiple receptors have shown strong therapeutic potential, emphasizing the need for multi-target strategies in protein design. However, most current protein sequence design methods focus on interactions with a single receptor, often neglecting the complexity of designing proteins that can bind to two distinct receptors. We introduced Protein Dual-Target Design Network (ProDualNet), a structure-based sequence design method that integrates sequence-structure information from two receptors to design dual-target protein sequences. ProDualNet used a heterogeneous graph network for pretraining and combines noise-augmented single-target data with real dual-target data for fine-tuning. This approach addressed the challenge of limited dual-target protein experimental structures. The efficacy of ProDualNet has been validated across multiple test sets, demonstrating better recovery and success rates compared to other multi-state design methods. In silico evaluation of cases like dual-target allosteric binding and non-overlapping interface binding highlights its potential for designing dual-target binding proteins. Data and code are available at https://github.com/chengliu97/ProDualNet.

Introduction

In living organisms, proteins rarely function in isolation. Instead, they engage in intricate networks of interactions with other proteins, essential for cellular homeostasis and regulating physiological processes [1–3]. A single protein often interacts with multiple distinct partners to carry out a wide range of biological functions. The phenomenon of multi-target binding exemplifies the complexity of biological signaling networks [4, 5]. For instance, Tirzepatide, a co-agonist peptide analog, simultaneously binds to both Gastric inhibitory polypeptide (GIP) and Glucagon-Like Peptide-1 (GLP-1) receptors, harnessing their synergistic effects to improve blood glucose and body weight regulation, offering superior effects over monotherapies targeting either receptor alone [6, 7]. This dual-receptor activation highlights the potential of multi-target drug design as a promising therapeutic strategy [8–14].

Artificial intelligence-driven design methods [15–18], such as ProteinMPNN [19] and ESM-IF1 [20], have introduced new workflows for protein sequence design. This process, known as protein inverse folding (structure-based sequence design), aims to generate protein sequences that fold into specific structures and exhibit desired functions, beginning with a given single backbone [21]. In contrast to the aforementioned methods, which require preexisting structural information for design, there are also structure-free or sequence-based design methods [22–24]. These approaches can be applied to sequence design in the absence of experimental structures or for direct de novo design, without the need for additional backbone design [25, 26]. Several of these methods are discussed in detail in the supplementary materials.

While some methods focus on multi-conformation design [20, 27–29], most structure-based protein sequence design research remains centered on single-target proteins, typically using the structure of a single protein complex [30]. This approach neglects the intricate interactions between the designed protein and multiple target proteins [1, 8], focusing primarily on binding to only one target. In many multi-state design methods, the probabilities of different conformations are simply averaged, without fully exploiting the model to explore the interactions between the designed protein and multiple receptors. As a result, these methods fail to capture the varying binding information when the protein interacts with different receptors. Therefore, multi-target design requires integrating multiple protein complex structures, ensuring that the designed protein can effectively bind to different receptors and perform its intended function. A further challenge lies in the limited availability of experimental datasets for multi-target protein complexes, which significantly hinders effective training.

To address these challenges, we introduce the Protein Dual-Target Design Network (ProDualNet), a structure-based sequence design method for designing dual-target protein sequences. ProDualNet improves the model to design proteins that interact with two targets by jointly modeling the two conformations of dual-target proteins. We first developed a protein sequence design framework based on heterogeneous graph networks, which serves as a pretraining model. This framework excels in capturing geometric features, which are crucial for accurately modeling the relationships between protein sequences, structures, and functions. To overcome the scarcity of dual-target experimental structures, we employed a noise-augmentation strategy (NoiseMix) during fine-tuning, which combines augmented single-target data with experimental dual-target data. This approach improves the model to capture the joint distributions of multiple protein complexes, which is essential for effective dual-target protein design.

We applied ProDualNet to various design tasks, showing its robustness and effectiveness. Additionally, we assessed its performance as a zero-shot predictor on dual-target sequence functional evaluations and mutation datasets, proving its capability to learn complex relationships between protein sequences, structures, and functions [31, 32].

Results

Model architecture

ProDualNet is a structure based dual-target protein sequence design framework that integrates the structure and sequence information of two protein complexes. The model is first pretrained on single-target data (Fig. 1a) and fine-tuned on dual-target data using the NoiseMix strategy. This addresses the challenge posed by the limited availability of dual-target complexes in the Protein Data Bank (PDB) and enhances the model’s ability to effectively capture the information of both receptors(Fig. 1b).

Figure 1.

Model framework, fusion strategy framework, and evaluation framework. (a) Schematic representation of ProDualNet pretraining using a standard training strategy. (b) ProDualNet fine-tuning for dual-target tasks, employing a shared encoder–decoder structure and the protein language model ESM2-based recycling strategy. (c) Overview of different dual-target modeling strategies. (d) Three test sets based on different evaluation criteria.

During fine-tuning, ProDualNet adopts a shared encoder–decoder architecture. And the model employs an autoregressive decoder to predict amino acid sequences, while Evolutionary Scale Modeling2 (ESM2) [33] is incorporated to provide additional evolutionary insights (Fig. 1b). This architecture enhances the understanding of the relationship between protein design and functionality. To capture information from different target receptors during decoding, we evaluate four distinct strategies, as shown in Fig. 1c. For detailed information on the model architecture and inference process, please see Methods.

Evaluating ProDualNet on dual-target protein testing sets

To evaluate the model’s robustness under different conditions, we established three test sets based on distinct criteria: Test Set 1 focused on typical dual-target design tasks; Test Set 2 involved proteins that undergo conformational changes upon binding to different receptors; and Test Set 3 examined proteins with non-overlapping binding interfaces to two target proteins (Fig. 1d; see “Methods”). Model performance was assessed using three metrics: global residue recovery, interface residue recovery, and perplexity. We used the default ProDualNet_base architecture (referred to as ProDualNet in the paper) and compared it with ProteinMPNN_mean, a leading approach in multi-state protein design [19, 29].

We first evaluated ProDualNet on Test Set 1, consisting of 159 protein complex pairs. ProDualNet outperformed all methods, achieving a sequence recovery of 0.581 and interface recovery of 0.627, significantly better than ProteinMPNN_mean (sequence recovery: 0.517, interface recovery: 0.541; Fig. 2a; Table 1). Statistical analysis confirmed the significance of these improvements. In terms of perplexity, ProDualNet scored 3.9, outperforming ProteinMPNN_mean (4.8) and ProteinMPNN_single (5.1). BLOSUM62 scores also showed ProDualNet’s superior performance in sequence design (Supplementary Table S3 available online at http://bib.oxfordjournals.org/, Supplementary Table S4 available online at http://bib.oxfordjournals.org/).

Figure 2.

Evaluation results in different design tasks. (a) Global recovery on Test Set 1. (b) Evaluation results on Test Set 2, featuring conformational changes. (c) Evaluation results on Test Set 3, with varying interface combinations. (d) AlphaFold3 results for the designed sequence on Test Set 1. (e) Comparison of AlphaFold2 scores (PAE, pLDDT, iPTM, and TM-score) between ProteinMPNN_b, ProteinMPNN_mean, and ProDualNet across 38 design cases in Test Set 3.

Table 1.

Performance analysis of four variants of ProDualNet and comparison method.

Model	Test 1		Test 2		Test 3
	Rec	IR	Rec	IR	Rec	IR
PorteinMPNN	0.481	0.471	0.465	0.446	0.455	0.412
ProteinMPNN_mean	0.517	0.541	0.521	0.521	0.489	0.481
ProdualNet(base)	0.581	0.627	0.584	0.613	0.556	0.571
ProdualNet_contact	0.591	0.636	0.596	0.627	0.574	0.573
ProdualNet_gate	0.591	0.637	0.593	0.629	0.575	0.596
ProdualNet_cross	0.589	0.621	0.596	0.615	0.577	0.567

Note:

• Rec: recovery.

• IR: interface recovery.

Additionally, we employed AlphaFold3 [34] to predict the structures of the designed sequences and target receptors. We focus primarily on three rank metrics: pLDDT, PAE, and ipTM. These metrics provide indicators for assessing the quality of sequence designs [25, 35]. ProDualNet outperforms the comparison methods in terms of the overall average values of these three metrics across 159 designed protein complex pairs. Specifically, ProDualNet achieves an ipTM of 0.728, a PAE of 8.65, and a pLDDT of 86.52, highlighting the reliability of the designs generated by ProDualNet (Fig. 2d).

We assessed ProDualNet’s effectiveness in dual-target protein design using Test Sets 2 and 3 under two design conditions. For Test Set 2 (52 protein complex pairs), ProDualNet outperformed ProteinMPNN_mean with a sequence recovery of 0.584 and interface recovery of 0.613, compared to ProteinMPNN_mean’s 0.521 and 0.521, respectively (Fig. 2b, Table 1). AlphaFold2 predictions indicated that the designed sequences undergo a conformational shift from alpha-helix to beta-sheet upon receptor binding, supporting ProDualNet’s capability in designing sequences for dual-target conformations (Supplementary Fig. S7 available online at http://bib.oxfordjournals.org/).

In Test Set 3 (38 protein complex pairs), where proteins had non-overlapping binding interfaces with two target proteins, ProDualNet again outperformed ProteinMPNN_mean in all metrics, particularly in interface residue recovery (Fig. 2c). ProDualNet achieved a sequence recovery of 0.556 and interface recovery of 0.571, while ProteinMPNN_mean had 0.489 and 0.481, respectively. ProteinMPNN_single showed even lower performance (0.455 sequence, 0.412 interface).

To further assess ProDualNet’s design capabilities, we employed an autoregressive strategy, designing 10 sequences for each of the 38 cases in Test Set 3, followed by rapid structure predictions with AlphaFold2 [36, 37]. And we compared two ProteinMPNN-based methods: ProteinMPNN_mean and ProteinMPNN_b. ProteinMPNN_b generate sequences independently for each target, neglecting the interrelationship between target proteins and conformational changes across states. ProDualNet outperformed ProteinMPNN_mean in 20 cases, achieving a pLDDT of 88.63 and PAE of 10.79, compared to ProteinMPNN_mean’s 87.86 and 11.49, respectively (Fig. 2e). ProDualNet also outperformed or matched ProteinMPNN_b, despite a slightly lower TM-score [38]. These results demonstrate that ProDualNet is more effective than single-complex-based methods and highlight the advantages of the multi-conformation information in protein sequence design [27, 28].

Case studies in dual-target protein design

To further validate the effectiveness of our approach, we selected two design cases from Test Set 3 for in silico evaluation: Ubiquitin thioesterase Otulin (Fig. 3a) and receptor protein-tyrosine kinase erbB-2 (Fig. 3b), which bind to different receptors at distinct interfaces. For these cases, we designed 100 sequences using both ProteinMPNN_mean and ProDualNet. The protein complex structures of the designed sequences were predicted using AlphaFold3 [34].

Figure 3.

Comparison of protein design methods across different design cases. (a) AlphaFold3 prediction results for designs: ProDualNet versus ProteinMPNN_mean in the ubiquitin thioesterase Otulin case (275 amino acids, PDB id: 3ZNZ, 6SAK). (b) erbB-2 receptor design case (624 amino acids, PDB id: 1S78, 3MZW). (d) Comparison of complex RMSD and success rate in the ubiquitin thioesterase Otulin (length, 275 amino acids) design case: ProDualNet versus ProteinMPNN_mean. The x-axis and y-axis represent protein complex RMSD values between the designed sequences and two different receptors, compared to the native structure. (e) Comparison of complex RMSD and success rate in the erbB-2 receptor (length, 624 amino acids) design case. (f) A comparative design analysis of PepPrCLIP and ProDualNet.

The success rate of each design case was determined by whether the designed proteins could correctly bind within the interaction space of the target receptors. We calculated the Root Mean Square Deviation (RMSD) of the protein complexes, defining a successful design as having an RMSD ≤5 Å for both protein native complexes [35].

Specifically, in the ubiquitin thioesterase otulin design case (PDB id: 3ZNZ, 6SAK), the designed protein sequence consisted of 275 amino acids. ProDualNet achieved a success rate of 37%, compared to 20% for ProteinMPNN_mean (Fig. 3d). Furthermore, ProDualNet outperformed ProteinMPNN_mean in AF3 metrics such as PAE and pLDDT (Fig. 3b, Supplementary Fig. S10 available online at http://bib.oxfordjournals.org/ and Supplementary Fig. S11a available online at http://bib.oxfordjournals.org/), see supplementary materials.

In the more challenging receptor protein-tyrosine kinase erbB-2 design case (PDB id: 1S78, 3MZW), where the designed protein sequence contained 624 amino acids, ProteinMPNN_mean failed to design any acceptable sequences (Fig. 3e). In contrast, ProDualNet successfully produced 6 high-quality designs out of 100, and outperformed ProteinMPNN_mean in AF3 metrics (Supplementary Fig. S11b available online at http://bib.oxfordjournals.org/) proving its robust design capabilities.

To assess the advantage of dual-target modeling, we compared it directly with single-target design. Specifically, we used ProteinMPNN’s single-target method (ProteinMPNN_a) to design sequences and employed AlphaFold3 to predict their complex structures with other receptors. Both ProDualNet and ProteinMPNN_mean outperformed this single-target approach (Fig. 3c). The inferior performance of the single-target method appears to stem from its inability to capture binding interface information specific to different target receptors. In conclusion, ProDualNet’s joint modeling of dual-target data results in higher accuracy and success rates in challenging multi-target design scenarios compared to ProteinMPNN_mean.

Comparison with structure-free method

To compare sequence-based and structure-based design approaches, we conducted a comparison between ProdualNet and PepPrCLIP [24]. Due to the limitation in PepPrCLIP’s training data, which only includes peptides of up to 25 amino acids, we restricted our evaluation to peptides of this length. Accordingly, we selected Hepcidin (PDB: 4QAE, 6WBV) for our analysis.

Using PepPrCLIP, we generated 250 000 peptides (20–25 aa) and scored their affinity for both receptors. The top 100 sequences by combined affinity score (Fig. S14 available online at http://bib.oxfordjournals.org/) were selected. For structure-based design, ProDualNet generated 100 sequences (temperature = 0.1), and AlphaFold3 predicted structures for all designs.

Although PepPrCLIP-selected sequences showed greater diversity, they had lower AlphaFold3 confidence compared to ProDualNet designs (Fig. 3f). This highlights a trade-off: structure-based methods provide higher confidence complexes, while sequence-based design offers broader diversity. Future work should explore hybrid strategies combining both approaches.

Functional effect prediction for dual-target sequences

Accurately predicting functional changes from protein mutations is crucial for directed evolution and AI-driven design. Zero-shot screening with Negative Log-Likelihood (NLL) scores accelerates in silico screening, speeding up experimental iteration [39, 40].

This study evaluated ProDualNet’s potential for dual-target peptide-based drug design using a benchmark dataset of 46 dual-target agonists derived from Glucagon Receptor (GCGR) and GLP-1R receptors [41](Fig. 4b). Each sequence consists of 29 amino acids, with experimental EC50 data for both receptors. None of the test set sequences were part of the training data, ensuring robust evaluation. ESM2 features were excluded from analysis.

Figure 4.

Functional scoring and mutational fitness evaluation of ProDualNet. (a) Comparative analysis of positive dual-target sequences using the top-k% selection strategy to assess the model’s ability to predict functionality. (b) Sequence similarity in the GCGR/GLP-1R dual-target test set. (c) Comparison of ProDualNet dual agonist sequence rank scores with AlphaFold3 rank scores. (d) Overall Spearman correlation coefficient evaluation of 65 sets of protein mutation data in the Proteingym stability database, each set containing multiple mutant sequences for a protein. (e) Higher-order mutation evaluation of 65 sets of protein mutation data in the Proteingym stability database. (f) Overall evaluation of 298 sets of protein mutation data in the DMS stability dataset. (g) Affinity evaluation of 11 sets of mutant sequences in the SKEMPI V2 database. (h) Evaluation of the relationship between model-predicted scores and binding free energy for 1,877 single-point mutations and 675 multi-point mutations across 11 proteins.

We assessed ProDualNet’s zero-shot prediction capability using NLL scores as a ranking metric for binding affinity to both receptors. Lower NLL scores correlated with higher binding success. Experimental results showed a Spearman correlation of 0.78 between NLL scores and combined log10EC50 values, outperforming AlphaFold3 scores (Fig. 4c). Stronger correlation between NLL and log10EC50 indicates a higher probability of dual agonist production. For further validation, we applied the top-K% strategy, ranking sequences by NLL scores and selecting the top K% for virtual experimental validation. ProDualNet consistently outperformed AlphaFold3’s three scoring metrics (PAE, pLDDT, and ranking_score) in top-K% screening (Fig. 4a). PAE and pLDDT were more effective than ranking_score for sequence ranking (Fig. 4c).

Zero-shot prediction of mutation effects

We assessed the ability of our pretrained model (ProDualNet_single) to predict mutation effects, crucial for understanding protein function. Our model, not relying on ESM2 features, captured sequence-structure–function relationships, providing a framework for optimizing protein stability and function.

Evaluating on two benchmark datasets (DMS [42, 43] and Proteingym Stability [32]), our model achieved a Spearman correlation of 0.67 in thermal stability, outperforming models like ProteinMPNN [19] and ESM-IF1 [20] (Fig. 4d). They have been proven to be effective zero-shot likelihood models [44]. It also performed well in higher order mutation prediction (0.45 correlation, Fig. 4e) and surpassed competitors in the DMS stability database (Fig. 4f). In antibody mutation data from SKEMPI V2 [45], our model showed better performance than structure-based models, with a correlation >0.5 in 5 of 11 cases (Fig. 4g), and a correlation of 0.46 for fitness prediction across 1877 single-point and 675 multi-point mutations (Fig. 4h).

These results confirm that our model effectively captures mutation impact, demonstrating its potential for protein engineering and further research.

Analysis of ProDualNet fusion strategies

We evaluated four feature fusion strategies of the ProDualNet variant, as shown in Fig. 1c, to assess their performance in modeling dual-target interfaces. Specific model details can be found in the supplementary material. Evaluation results across three test sets revealed that while ProDual_contact, ProDualNet_gate, and ProDualNet_cross demonstrated higher recovery rates in interface restoration compared to ProDualNet_base (Table 1), they did not exhibit significant improvements in structural confidence for generated sequences (Fig. 5a). This limitation may arise from the relative scarcity of experimental dual-target structures, which, even with our noisemix strategy, restricts the performance of more complex data-driven models like ProDualNet_gate and ProDualNet_cross.

Figure 5.

Evaluation of ProDualNet’s modeling strategies and ablation analysis. (a) Performance analysis of ProDualNet’s modeling strategies, using AlphaFold2 structure confidence scores and TM-score. (b) Comparative analysis between the ProDualNet pretrained model and ProteinMPNN. (c) Comparison of fine-tuning ProteinMPNN_mean and ProDualNet on training and test set 1. (d) Analysis of the mixed augmented data ratio in the NoiseMix strategy. (e) Evaluation of the effect of the ESM2 language model and validation of the recycling strategy.

Despite this, ProDualNet_gate and ProDualNet_cross showed better capabilities in interface residue design, with their designed sequences yielding higher BLOSUM62 scores (Supplementary Table S6 available online at http://bib.oxfordjournals.org/). However, their design diversity was notably lower than that of ProDualNet_base. During the fine-tuning, ProDualNet_base converged more quickly without the need for additional weight parameters, compared to the pretrained model. Moreover, fine-tuning on dual-target data enabled ProDualNet_base, through its autoregressive decoding mechanism, to dynamically integrate interface information. Overall, although ProDualNet_base was slightly less effective in some aspects of interface modeling, it remains a reliable strategy due to its advantages.

Ablation study of ProDualNet enhancements

We trained several ablation models to evaluate our methods. ProDualNet_single (no ESM2) outperformed or matched ProteinMPNN across three protein types (monomers, homologous, and heterooligomers) in sequence recovery (Fig. 5b), demonstrating the effectiveness of our model.

Evaluation on training and test sets (Fig. 5c) showed consistent performance with noise augmentation. ProDualNet outperformed ProteinMPNN_mean (multi-state) fine-tuned with NoiseMix (0.581 versus 0.534) in sequence recovery. Fine-tuning ProteinMPNN (multi-state) for dual-target design further highlighted ProDualNet’s advantages (Fig. 5c).

We tested the NoiseMix strategy by mixing noise-augmented data with real dual-target data. Models trained with augmented data (recovery = 0.581) significantly outperformed those trained on raw data (0.546) or without protein cluster sampling (0.532) (Fig. 5d), demonstrating the positive impact of noise augmentation. Without NoiseMix, fine-tuning methods led to overfitting (Supplementary Fig. S1 available online at http://bib.oxfordjournals.org/).

Finally, we validated the recycling strategy and ESM2 embedding (Fig. 5e). Setting the recycling parameter to 1 yielded the best performance and runtime efficiency, outperforming models without recycling and those using only ESM2 features. ProDualNet trained without ESM2 still outperformed ProteinMPNN_mean (0.54 versus 0.52). This demonstrates that the model’s improved performance results from the complementary nature of geometric information from the full protein backbone and evolutionary features from ESM2.

Conclusion and discussion

In this study, we introduce ProDualNet, a structure-based sequence design model specifically tailored for dual-target protein design. By extending single-target protein design approaches (e.g. ProteinMPNN [19] and ProBID-Net [30]) to the more complex task of dual-target protein sequence design, ProDualNet addresses the challenge of limited experimental dual-target data through a noise-augmentation strategy, NoiseMix. This approach mitigates the scarcity of dual-target protein structures, enhancing the model’s performance and stability in complex design tasks.

Although primarily validated in silico, the effectiveness and reliability of ProDualNet for dual-target protein design surpass those of other comparative methods. Future work will involve wet lab validation and further expansion of the model to handle more complex proteins and interactions, such as those involving RNA, DNA, metal ions, and small molecules [11, 16].

Additionally, structure-based sequence design is constrained by its reliance on existing structure [21]. In contrast, structure-free [22–24] and joint structure-sequence design [46, 47] approaches provide a potential path to overcome these limitations, enabling true de novo design. Future research will focus on advancing the accurate design of de novo multi-target proteins and agonists.

Materials and methods

Pretraining dataset

To construct the dual-target protein dataset, we first curated our dataset from PDBbank. Building on the methodology of ProteinMPNN [19], we clustered proteins in the database into 25 361 clusters based on a 30% sequence identity cutoff using mmseqs2 [48]. From these, we selected 766 protein clusters known to interact with multiple targets. The dataset was then split into 466 clusters for training and 300 clusters for testing (Supplementary Fig. S15 available online at http://bib.oxfordjournals.org/).

For pretraining, we utilized 20 137 clusters from the database as the training set, with 1662 clusters allocated for testing and 1662 clusters for validation. It is important to note that the clusters in the multi-target test set are entirely derived from the pretraining validation and test datasets, ensuring no data leakage. While MMSeqs2 clustering may, in theory, allow for some protein leakage beyond the similarity threshold [49], we followed the construction guidelines established in ProteinMPNN [19] to mitigate this risk.

Fine-tuning dataset and test dataset

After pretraining, the fine-tuning stage involved extracting 300 dual-target protein clusters from the testing and validation sets for the design task. The remaining 466 multi-target protein clusters from the training set were used to fine-tune the model, focusing on dual-target interactions. To mitigate the lack of training data, we applied the NoiseMix strategy, combining the 466 clusters with random protein samples augmented with Gaussian noise during each fine-tuning epoch. Details of the training procedure are in the supplementary materials (Supplementary Table S2 available online at http://bib.oxfordjournals.org/). We defined the interface residues on the unknown chain as those with Cα atoms located within 8 Å of the Cα atoms on the known receptor chain.

To assess the model’s robustness, we created three test sets based on sequence similarity, conformational changes, and binding interfaces. These sets were derived from the same 300 protein clusters using stratified sampling, with some overlap between them:

Test Set 1: This set consists of 159 dual-target protein complex pairs, with homologous similarity between the two target proteins below 30%.

Test Set 2: This set contains 52 protein complex pairs, where the RMSD of conformational changes between protein–protein interactions exceeds 2 Å, and the similarity between the two receptors is below 50%.

Test Set 3: This set includes 38 protein complex pairs, where the binding interfaces of the designed protein do not overlap with either of the two target proteins and receptor similarity is below 30%.

These three test sets were designed to evaluate the model’s performance under varying conditions of target protein similarity, conformational change, and binding specificity.

GCGR-GLP1R dual-target dataset

To assess the performance of ProDualNet in dual-target peptide drug design, we conducted a benchmark evaluation using a dual-target agonist dataset based on GCGR and GLP-1R. The dataset comprises 46 protein sequences, each 29 amino acids in length, with experimentally determined EC50 values for both GCGR and GLP-1R receptors [50]. Importantly, each sequence differs by at least one amino acid, with an average variation of 14 amino acids, rendering this task more complex than single mutation predictions. Notably, all test set sequences were excluded from the model’s training data, ensuring a robust and unbiased evaluation.

Fitness dataset for mutation prediction

For the assessment of thermostability mutations, we used two databases, ProteinGym DMS [32] and DMS_stability [42, 43]. From ProteinGym DMS, we selected 65 proteins with mutations related to thermostability, consisting of 67 636 single-point mutations and 65 579 higher order mutations. For the DMS_stability dataset, we used their processed data, which includes 298 proteins and a total of 271 231 single-point mutations.

Additionally, to evaluate the performance in predicting the fitness of complex mutations, we selected mutation data from the SKEMPI V2 database [45]. This subset includes antibody mutations with more than 100 entries, comprising 11 proteins, 1877 single-point mutations, and 675 higher order mutations.

ProDualNet architecture

The architecture for the dual-target protein design task involves a shared encoder–decoder framework, where the encoder processes the structural and sequence information of both targets. This encoder, composed of four layers of heterogeneous graph networks, captures the relationships between the two target structures through a unified embedding space. Each layer of the heterogeneous graph layer consists of a MPNN block [19] and a Transformer block [51], as shown in the architecture diagram (Supplementary Fig. S12). Model details for different fusion strategy variants that we proposed can be found in the supplementary material.

To capture these interdependencies, for each layer of the heterogeneous graph network, the feature vectors and corresponding to the two different conformations of the designed protein complex are combined via a weighted sum:

(5)

where is a hyperparameter that controls the contribution of each target structure to the final representation. This weighted combination allows the model to learn shared features that are relevant to both targets.

To further model the joint sequence distribution for both targets, the softmax function is applied to the combined decoder feature representation, enabling efficient model optimization:

(6)

Where is the weight matrix, and is the bias term. The vectors and represent the feature vectors of the two distinct target conformations. The value represents the predicted amino acid type at position , based on the feature vector at position of the protein. In this case, .

The mean operation ensures that the features from both targets are effectively fused before prediction. This approach allows the model to account for the structural and sequence dependencies between the two targets, facilitating the design of sequences that can interact with both targets simultaneously.

NoiseMix strategy

To avoid overfitting due to the limited availability of high-quality dual-target protein data, we use a two-phase training strategy. First, the model is pretrained on a larger, general dataset of single-target proteins, similar to ProteinMPNN, to learn basic structure-sequence mappings. This pretrained model, called ProDualNet_single. Afterward, the model is fine-tuned on a smaller, domain-specific dual-target dataset using a noise augmentation strategy called NoiseMix.

(7)

As shown in the equation (7), represents the real dual-target experimental structure pairs, while represents pseudo-dual-target complexes constructed from single–complex structures through noise augmentation. The amount of noise-augmented data, denoted as N, is treated as a hyperparameter that can be adjusted for optimal training performance. This approach enhances the model’s generalization ability and reduces the risk of overfitting, allowing for more effective joint modeling of the two receptors and the designed protein. The ablation study demonstrates the effectiveness of this noise-enhanced fine-tuning strategy in improving ProDualNet’s performance on complex dual-target design tasks (Supplementary Fig. S1 available online at http://bib.oxfordjournals.org/ and Supplementary Algorithm 2).

Noise augmentation

For the NoiseMix strategy, we employ a noise augmentation approach to enhance single-target complexes. In this approach, Gaussian noise is added to the coordinates of non-dual-target protein complexes to generate synthetic structure pairs (, ), where and are noisy versions of the same complex.

Mathematically, this process is represented as:

(8)

where denotes Gaussian noise with mean 0 and variance , and β is a scaling factor that controls the magnitude of the noise (e.g. β = 0.2, =1). These noisy structure pairs are incorporated into the Noisemix training process, augmenting the available dual-target data and enhancing the model to generalize across a broader variety of protein conformations.

Loss function

The loss function used to train the model measures the discrepancy between the predicted and true amino acid types at each position in the protein sequence. For a sequence of length , where each position is represented by a feature vector , the model predicts the probability distribution over the 20 amino acid types using the softmax function:

(9)

where and are the weight matrix and bias term, respectively. represents the prediction of the amino acid type at position based on the feature at position of the protein.

The total loss is the weighted sum of cross-entropy losses for each position , as well as for its neighboring positions and . The individual cross-entropy loss at position is defined as:

(10)

where is the indicator function that equals 1 if the true amino acid at position is , and 0 otherwise. The overall loss is computed as:

(11)

where is a hyperparameter controlling the weight of neighboring positions in the loss calculation. In the pretraining phase, α is set to 0.25. Similarly, during the dual-target fine-tuning phase, α is also set to 0.

Training details

Pretraining: For training, we used a single NVIDIA A100-SXM4-40GB GPU. The model is pretrained using the Adam optimizer with parameters , , and , following the learning rate schedule of ProteinMPNN. The pretraining phase lasts for 5000 K steps, with α set to 0.25. The training dataset consists of 20 137 protein clusters. We set the noise intensity to β = 0.2.

Fine-tuning: Following pretraining, the model undergoes fine-tuning using the Adam optimizer with a learning rate of 0.00001 for an additional 20 K steps. This phase focuses on optimizing the model for dual-target protein interactions, with α = 0. For models utilizing ESM2 protein language model features, we fill the target sequence features with a zero tensor and apply a recycled training approach. Specific training details are provided in the supplementary materials. During fine-tuning, each epoch includes a mixture of dual-target data from 466 protein clusters with multi-target interactions, as well as 4000 protein clusters randomly sampled from the pretraining dataset, augmented with noise, the detailed training algorithms are provided in the supplementary materials. For the noise strategy, we adopt the way of Carbondesign [15, 16].

Key Points

1. In structure-based sequence design, we successfully addressed the challenge of limited experimental structures for dual-target proteins by using the Noisemix strategy and modeled the dual-target protein complex.

2. In multiple independent validations, we confirmed that Protein Dual-Target Design Network (Produalnet) outperforms current multi-state design approaches and single-conformation design methods.

3. As a zero-shot predictor, Produalnet achieved better screening efficiency on the GLP-1/GCGR dual agonist dataset.

Conflict of interest

None declared.

Funding

This work was supported by Shanghai Municipal Science and Technology Major Project, partially by SJTU Kunpeng & Ascend Center of Excellence, the Center for HPC at Shanghai Jiao Tong University, and the National Key Research and Development Program of China (2025YFA0921001 and 2023YFF1205102), by National Natural Science Foundation of China (12171318 and 32171242), by Shanghai Science and Technology Commission (24JS2810200, 21ZR1436300, 23XD1401900, 23DZ2290600), Medical Engineering Cross Fund of Shanghai Jiao Tong University (YG2023ZD21 and YG2023LC03).