Improving protein and protein interactions using pseudo-dimers derived from monomeric proteins Authors: Hao Du, Xinzhe Zheng, Yuchen Ren, He Huang, Xinqi Gong, Wanli Ouyang, Yang Zhang, Yan Lu Published: 2026 ABSTRACT Accurately predicting protein-protein interactions (PPIs) in dimeric complexes remains a fundamental challenge in computational biology. Although existing PPIs prediction models, such as AlphaFold-Multimer (AF-Multimer) and AlphaFold3 (AF3), have achieved impressive performance, they still suffer from unsatisfactory accuracy due to the limited availability of protein dimer structures, whose collection is both expensive and labor-intensive. Here, we introduce a simple yet effective pre-training method, termed split and merge proxy (SMP), that leverages abundant monomeric proteins to simulate various PPIs tasks for the first time. Specifically, SMP constructs pseudo-dimers by splitting monomer data into two subunits, referred to as pseudo-receptors and pseudo-ligands, and trains models to merge them back by predicting their pseudo interactions (e.g., contact or docking). This proxy task enables large-scale pre-training without additional cost. Models pre-trained with SMP and subsequently fine-tuned on real protein dimer datasets demonstrate consistently improved accuracy and generalization across multiple benchmarks, surpassing strong baselines. Notably, SMP delivers more accurate structure predictions than both AF-Multimer and AF3 on several CASP15 dimer targets. Our findings highlight SMP as a scalable strategy for harnessing monomeric data to advance protein complex modeling, providing insights into the linkage between monomers and multimers. Introduction Identifying protein-protein interactions (PPIs) in dimeric complexes remains a significant and long-standing challenge in computational biology1–3. The notable success of AlphaFold2 (AF2)4 has opened a promising avenue in this area. Building on this advance, a number of deep learning-based methods5–8 have been proposed to predict and analyze PPIs. These approaches are essentially data-driven, depending on well-curated datasets to achieve accurate and generalizable predictions, thereby underscoring the importance of large-scale, high-quality protein complex data. Given the central role of multimeric data, especially dimeric proteins, in training and evaluating PPI prediction models, substantial efforts have been devoted to curating high-quality datasets. High-performing multimer structure prediction models, such as AlphaFold-Multimer (AF-Multimer)5 and AlphaFold3 (AF3)6, are trained on carefully curated entries from the Protein Data Bank (PDB)9, the largest open-access repository of experimentally determined structures, which by 2024 contains more than 100,000 protein complexes. These models have shown strong performance across diverse applications10,11. To support rigorous evaluation, several benchmarks have been established as well3,12,13. A prominent example is the Critical Assessment of Structure Prediction (CASP)14,15, a globally recognized benchmark that evaluates protein structure prediction models against newly solved but unpublished experimental structures, thereby ensuring unbiased assessment. Collectively, these datasets and benchmarks not only enable systematic assessments but also drive methodological advances in computational biology. Despite substantial efforts to assemble high-quality protein complex datasets, current PPIs prediction models still fall short of achieving consistently reliable performance16,17. One major bottleneck is the scarcity of diverse, high-quality dimeric structures. Notably, AF2, the most successful protein structure prediction model to date, is trained on approximately 132,000 experimentally resolved single-chain structures available in the PDB prior to 2018, spanning both monomeric and multimeric entries18. In contrast, only 19,166 high-quality, non-redundant protein dimers were available in the same period19, after stringent filtering steps such as removing chains that were too short or too long and eliminating sequences with > 30% similarity. This represents nearly an order-of-magnitude reduction relative to single-chain proteins. This stark imbalance highlights the difficulty in acquiring high-quality experimental dimers, thereby limiting the performance of deep learning models. To overcome this challenge, cost-effective strategies for generating additional training samples without relying on labor-intensive experimental procedures are urgently needed. In this work, we propose Split and Merge Proxy (SMP), a simple yet effective pre-training strategy that leverages abundant monomeric proteins to enhance PPIs prediction at no additional cost. SMP constructs pseudo-dimer samples by splitting a monomer into two subunits, termed pseudo-receptors and pseudo-ligands, and training an interaction model to merge them back, effectively simulating the process of protein interactions. Although these pseudo-dimers inevitably introduce biological noise, they nevertheless provide rich information that complements existing protein dimer datasets. The pretrained model can be directly fine-tuned on real dimeric structures without any modifications to the model architecture. Crucially, the alignment between the pre-training and downstream objectives eliminates task discrepancies during fine-tuning, enabling more efficient adaptation and yielding higher prediction accuracy. Although some existing approaches fragment full-length proteins into shorter segments for training4–6, with AF2’s cropping strategy being a representative example, the motivation, operation, and learning objective differ fundamentally from SMP. Specifically, (1) Motivation: the cropping strategy of AF2 is primarily introduced to mitigate computational constraints, notably the memory limitations of TPUv3 cores, as the memory usage scales cubically with sequence length4. In contrast, SMP is driven by the scarcity of high-quality dimeric training data. Our aim is to explicitly simulate inter-chain settings and generate additional supervisory data without additional experimental structures. (2) Operation: motivated by distinct goals, the two methods adopt different operational designs. AF2 cropping only shortens sequence length by randomly selecting a contiguous window (typically 256 or 384 residues) from a single-chain protein. In contrast, SMP splits a monomer into two artificial chains, a pseudo-receptor and a pseudo-ligand, to construct pseudo-dimers, making its operation intrinsically different from AF2. (3) Learning objective: the operational distinctions naturally lead to divergent training objectives. AF2 performs intra-chain structure prediction with the cropping input. SMP, however, supervises the model on inter-chain labels, such as contact maps and docking poses, which go beyond the scope of AF2. To evaluate the effectiveness of SMP, we apply it across multiple protein interaction tasks, including protein-protein contact map prediction20, protein-protein docking21, and protein-protein interaction22. Across all tasks, SMP consistently outperforms strong baselines and achieves state-of-the-art (SOTA) results, demonstrating its generality and robustness for both residue-level and sequence-level interaction prediction. We further assess SMP under data-scarce settings with only limited real dimer samples and observe substantial performance gains, highlighting the benefit of the proposed pre-training strategy. Finally, we test SMP on real-world applications using the CASP15 benchmark23 and find that SMP can produce more accurate interaction predictions for some targets compared with AF-Multimer and AF3, despite their significantly larger model capacities and computational budgets. Results The overview of SMP SMP is a simple yet effective proxy strategy designed for pre-training protein interaction models. The overall framework consists of two stages: split and merge. In the split stage, each monomeric protein is partitioned into two subsequences with corresponding structural information, forming a pseudo-dimer referred to as the pseudo-receptor and pseudo-ligand (Fig. 1a). Although such pseudo-dimers inevitably introduce biological noise, they retain meaningful structural patterns for protein interactions. To quantify this effect, we employ Foldseek-Multimer24 to assess structural similarity between pseudo-dimers (generated via random splits) and real protein complexes. As shown in Fig. 1b, 2974 pseudo-dimers achieve a TM-score25,26 greater than 0.5 when matched to real complexes, accounting for approximately 13% of the pseudo-dimer dataset, indirectly indicating that monomer data can complement existing multimer data. Representative examples are illustrated in Fig. 1c, including PDB pairs 2BQI-3ZVQ and 1BGO-5I4H, with TM-scores of 0.95 and 0.87, respectively. In the merge stage, pairwise pseudo-dimers are embedded as the input for sequence-based or structure-based interaction models. The models are trained using labels derived from the split procedure, enabling task-specific pre-training for interaction prediction. Subsequently, the pre-trained models are fine-tuned on real protein dimer datasets without architectural modification, serving as strong initializations for downstream tasks (Fig. 1d). To demonstrate the generality of SMP, we apply it across multiple PPIs tasks, including residue-level tasks (protein-protein contact map prediction and docking) and sequence-level tasks (protein-protein interaction prediction), consistently achieving superior performance (Fig. 1e).Fig. 1Splitting and merging monomer data for the PPIs tasks.a A monomer protein sequence is divided into two subunits, designated as a pseudo-receptor and a pseudo-ligand, which are reflected in their 3D structures. b Foldseek-Multimer is used to align the monomer structure with the multimer structure. A total of 2974 pairs are found with a TM-score larger than 0.5. c Two representative matched monomer-dimer pairs are shown: 2BQI-3ZVQ and 1BGO-5I4H, with TM-scores of 0.95 and 0.87, respectively. d The schematic of the SMP pipeline comprises a pre-training stage using monomer-derived pseudo pairs, followed by a fine-tuning stage on real dimer pairs. The loss is adapted to specific PPIs tasks. e Three PPI-related tasks are conducted in this study: residue-level predictions, including contact map prediction and protein-protein docking, and sequence-level predictions, including protein-protein interaction prediction. Created in BioRender. Zheng, X. (2026) https://BioRender.com/a3rsq3n. Splitting and merging monomer data for the PPIs tasks. a A monomer protein sequence is divided into two subunits, designated as a pseudo-receptor and a pseudo-ligand, which are reflected in their 3D structures. b Foldseek-Multimer is used to align the monomer structure with the multimer structure. A total of 2974 pairs are found with a TM-score larger than 0.5. c Two representative matched monomer-dimer pairs are shown: 2BQI-3ZVQ and 1BGO-5I4H, with TM-scores of 0.95 and 0.87, respectively. d The schematic of the SMP pipeline comprises a pre-training stage using monomer-derived pseudo pairs, followed by a fine-tuning stage on real dimer pairs. The loss is adapted to specific PPIs tasks. e Three PPI-related tasks are conducted in this study: residue-level predictions, including contact map prediction and protein-protein docking, and sequence-level predictions, including protein-protein interaction prediction. Created in BioRender. Zheng, X. (2026) https://BioRender.com/a3rsq3n. Comprehensive analysis of splitting strategies We begin by investigating how monomeric proteins are split, as randomly splitting can increase the number of training samples but may disrupt secondary structures or functional regions, while a domain-guided splitting strategy preserves biologically meaningful fragments, though not all proteins contain multiple domains. To systematically examine these two approaches, we conduct a comprehensive evaluation spanning both residue-level tasks, such as contact map prediction and protein docking, and sequence-level tasks, including interaction classification. For the domain-aware strategy, we adopt a domain segmentation tool built upon the CATH v4.4 database27, yielding 15,602 multi-domain proteins released before 2018-04-30. For the random-split strategy, we use the same protein set but follow a center-based rule, in which each split point is selected near the midpoint of the monomeric protein chain to maintain reasonable fragment lengths. Further analysis of the random split positions is provided in Supplementary Section S1, with detailed results summarized in Supplementary Table S1. We can observe a decline in performance when the split point deviates from the chain center, primarily because the resulting excessively short fragments are considered trivial and are consequently excluded during data pre-processing. The comparisons are illustrated in Fig. 2. Our results show that: (1) the domain-aware split strategy consistently improves the model’s overall performance across multiple downstream tasks. For example, in the contact map prediction task, the domain-aware split outperforms the baseline model DeepInter8, achieving improvements of 14.7% in P@L for homodimers and 5.1% for heterodimers. In the interaction classification task, it also surpasses the baseline PPITrans28 model, yielding gains of 8.2% in recall and 4.3% in F1-score. These results indicate that incorporating monomeric proteins indeed benefits PPIs prediction by providing complementary and informative structural and sequence cues; (2) While the domain-aware split provides better biological interpretability, the random split strategy consistently yields higher performance and robustness in contact map prediction and protein docking tasks, and achieves comparable results in the interaction classification task. Specifically, for the docking task, the random split improves the success rate by an average of 3.0% over the domain-aware split (33% vs. 30%). For contact map prediction, the random split also performs better in both homodimer and heterodimer scenarios (homodimer P@L: 0.706 vs. 0.690; heterodimer P@L: 0.373 vs. 0.353). For the interaction classification task, the two strategies yield similar performance (AUPR: 0.779 vs. 0.772).Fig. 2Investigation of pseudo-dimer construction strategies using random and domain-based splits.We assess how different pseudo-dimer construction strategies, random split and domain-based split, affect model performance across three representative tasks. Each experiment is repeated three times with different seeds to mitigate the effect of randomness. a Contact map prediction on Homodimer300 and Heterodimer99 is evaluated using Top 1, 10, 25, 50, L/10, L/5, and L precision metrics. b Protein docking on the DIPS-Het benchmark is measured by CRMSD, IRMSD, DockQ, and success rate. c PPI prediction on the D-SCRIPT dataset is evaluated using recall, precision, and AUPR. Random split consistently outperforms domain-based split in both contact map prediction and docking tasks, whereas the two strategies yield comparable performance in the PPI prediction task. Source data are provided as a Source Data file. Investigation of pseudo-dimer construction strategies using random and domain-based splits. We assess how different pseudo-dimer construction strategies, random split and domain-based split, affect model performance across three representative tasks. Each experiment is repeated three times with different seeds to mitigate the effect of randomness. a Contact map prediction on Homodimer300 and Heterodimer99 is evaluated using Top 1, 10, 25, 50, L/10, L/5, and L precision metrics. b Protein docking on the DIPS-Het benchmark is measured by CRMSD, IRMSD, DockQ, and success rate. c PPI prediction on the D-SCRIPT dataset is evaluated using recall, precision, and AUPR. Random split consistently outperforms domain-based split in both contact map prediction and docking tasks, whereas the two strategies yield comparable performance in the PPI prediction task. Source data are provided as a Source Data file. Although this finding may appear counterintuitive, it reflects a common property of deep learning-based methods: their ability can benefit from a broader and more diverse set of data patterns during training. In contrast, a domain-guided splitting strategy may inadvertently have structurally or sequentially similar samples together with real protein dimers, thereby reducing the diversity of training examples and limiting the model’s generalization capability. By comparison, the random split strategy introduces greater heterogeneity into the training data, which can enhance the model’s robustness and generalization performance across diverse PPIs scenarios. Hence, we use the randomly split strategy with a center-based rule to construct our pseudo-dimer pre-training dataset. In addition, we analyze the pre-training dataset composed of single-domain and multi-domain proteins for the contact map prediction task, as detailed in Supplementary Section S2. The results (Supplementary Table S2) show that multi-domain proteins provide more effective supervision than single-domain proteins in both the homodimer and heterodimer settings. Moreover, a mixed pre-training dataset combining single-domain and multi-domain proteins yields the best overall performance, suggesting that single-domain proteins still capture biologically meaningful patterns that are transferable to real protein complexes. We further examined how different pseudo-dimer construction strategies affect the learned PPI representations and what interaction-related properties are captured through SMP pre-training, as detailed in Supplementary Section S7. Our qualitative analysis (Supplementary Fig. S3) indicates that SMP pre-training enables the model to learn meaningful contact and interface priors. Moreover, for certain proteins, the random-split strategy, by introducing more heterogeneous protein data, further enhances the learning of contact-specific features and improves the modeling of interfacial regions. SMP enhances protein-protein contact map prediction The protein-protein contact map prediction task aims to determine whether an individual amino acid on one protein is in contact with a residue on the other, playing an important role in the structural or functional analysis of protein complexes29,30. We combine SMP with the existing strong contact model DeepInter8 to evaluate the effectiveness and robustness of SMP on both the homodimer test set Homodimer30031 and the heterodimer test set Heterodimer998. Fig. 3(a) indicates the detailed contact results on the Homodimer300 dataset, showing that SMP outperforms all models across different metrics. In particular, SMP brings 5.6, 6.4, 5.2, 4.9, 4.8, 4.5, and 4.5% performance increases compared to the well-performing model DeepInter on the Top 1, 10, 25, 50, L/10, L/5, and L precision metrics, obtaining the highest precision of 0.81, 0.80, 0.79, 0.77, 0.79, 0.77, and 0.72, respectively, which means that SMP can learn more discriminative residue representation and useful contact prediction knowledge from extra pseudo-dimer data, and help the model identify the contact binding site. In addition, for the more difficult heterologous test set Heterodimer99, SMP consistently demonstrates superior performance over other models, as shown in Fig. 3b. Notably, SMP improves upon DeepInter by 12.6, 18.8, 18.8, 17.8, 21.4, 16.6, and 15.8% on the Top 1, 10, 25, 50, L/10, L/5, and L precision metrics, achieving the best precision of 0.47, 0.44, 0.43, 0.41, 0.44, 0.43, and 0.37, respectively, indicating that these performance gains may arise from the consistency between pseudo dimers and heterologous proteins since the cut chains usually have low sequence identities, sharing certain similar properties and distributions of the real heterologous data, leading to clear performance improvements on heterologous complexes.Fig. 3Performance of SMP in the protein-protein contact map prediction task.a, b Radar charts comparing performance on the Homodimer300 (a) and Heterodimer99 (b) benchmarks across multiple models, including DeepHomo {\&} 2.0, GLINTER, CDPred, DeepInter, and SMP, evaluated using Top 1, 10, 25, 50, L/10, L/5, and L precision metrics. SMP achieves state-of-the-art performance across these metrics. c Left: Comparison of Top L precision between SMP and DeepInter on Homodimer300. Right: Scatter plot showing Top L precision per sample for SMP and DeepInter on Heterodimer99. d Left: Precision of SMP, DeepInter, CDPred, GLINTER, and DeepHomo {\&} 2.0 across varying numbers of predicted contacts in Homodimer300. Right: Precision of SMP, DeepInter, CDPred, and GLINTER across different thresholds of predicted contacts in Heterodimer99. e Contact map visualization for two representative samples with PDB IDs 3LQN (left) and 6DRR (right) for DeepInter (top), SMP (middle), and Ground Truth (bottom). Source data are provided as a Source Data file. Performance of SMP in the protein-protein contact map prediction task. a, b Radar charts comparing performance on the Homodimer300 (a) and Heterodimer99 (b) benchmarks across multiple models, including DeepHomo {\&} 2.0, GLINTER, CDPred, DeepInter, and SMP, evaluated using Top 1, 10, 25, 50, L/10, L/5, and L precision metrics. SMP achieves state-of-the-art performance across these metrics. c Left: Comparison of Top L precision between SMP and DeepInter on Homodimer300. Right: Scatter plot showing Top L precision per sample for SMP and DeepInter on Heterodimer99. d Left: Precision of SMP, DeepInter, CDPred, GLINTER, and DeepHomo {\&} 2.0 across varying numbers of predicted contacts in Homodimer300. Right: Precision of SMP, DeepInter, CDPred, and GLINTER across different thresholds of predicted contacts in Heterodimer99. e Contact map visualization for two representative samples with PDB IDs 3LQN (left) and 6DRR (right) for DeepInter (top), SMP (middle), and Ground Truth (bottom). Source data are provided as a Source Data file. Furthermore, we analyze the Top L precision of each sample in the Homodimer300 and Heterodimer99 for DeepInter and SMP, as illustrated in Fig. 3c. SMP performs better than DeepInter for 154 out of 297 samples on the Homodimer300 and 48 out of 95 cases on the Heterodimer99. The results show SMP is a robust method for homologous dimers or heterologous dimers, demonstrating that our SMP can be applied to various types of proteins. Meanwhile, Fig. 3d shows the performance comparison of different numbers of top predicted contacts in the Homodimer300 and Heterodimer99. SMP outperforms all current models across different predicted contact numbers, indicating that SMP’s predictions are more reliable and confident, resulting in a higher true positive rate. Finally, Fig. 3e presents the contact map visualizations for two samples with PDB IDs 3LQN and 6DRR in the Homodimer300, predicted by DeepInter and SMP. It can be observed that SMP captures contact regions similar to the ground truth, but DeepInter fails. This demonstrates that SMP can generate more accurate contact maps than other models for certain proteins. The additional contact map visualizations in Heterodimer99 are provided in Supplementary Fig. S4. Overall, this pre-training paradigm plays a considerable role in various contact map benchmarks, including homologous and heterologous proteins, and shows superiority and robustness in these benchmarks. SMP improves structure accuracy during docking Protein-protein docking predicts the bound 3D structure of a protein complex based on the known 3D structures of two individual proteins in their unbound states, which is a crucial and realistic problem in many biological scenarios like drug design32 and the study of disease mechanisms33. This study focuses on the rigid body docking task, a common phenomenon in many biological scenarios, where the conformations during the protein binding process are not considered, as discussed in recent works34,35. To demonstrate the benefit of our method, we incorporate SMP into the popular method EquiDock34 and analyze it on a challenging docking benchmark DIPS-Het36,37. Fig. 4a illustrates the extensive docking results, which show that SMP achieves better performance than current deep learning-based models in terms of Complex and Interface Root Mean Square Deviations (CRMSD and IRMSD), and DockQ metrics. Specifically, compared to the best docking model, EquiDock, the mean and standard deviation of CRMSD for SMP decrease from 12.00 to 10.77 and from 10.18 to 9.21, respectively. Meanwhile, the mean and standard deviation of IRMSD decrease from 10.43 to 9.46 and from 8.72 to 7.87, respectively. In contrast, the mean and standard deviation of DockQ increase from 0.18 to 0.20 and from 0.20 to 0.21, respectively, when compared to EquiDock. These evaluations highlight SMP’s SOTA performance across multiple metrics compared to the existing models, demonstrating that the pseudo-dimer data can provide further structural prior knowledge, and our pre-training strategy is effective for protein complex structure prediction. In addition, Fig. 4a presents a comparison of the success rate (DockQ > 0.23) between SMP and other models. SMP outperforms current methods, achieving the highest success rate at 35%, namely the 5, 13, and 13% improvements over the EquiDock, DiffDock-PP, and GeoDock. We further evaluated the robustness of SMP on the DIPS-Het docking benchmark by conducting five independent runs with different random seeds, as detailed in Supplementary Section S4. SMP consistently improved over EquiDock across complex RMSD, interface RMSD, DockQ, and success rate, with corresponding p-values of 0.0281, 0.0200, 0.0225, and 0.0111, respectively (Supplementary Table S6).Fig. 4Performance comparison in the protein-protein docking task.a Box plots showing the distribution of CRMSD (first column), IRMSD (second column), and DockQ (third column) for four docking methods: GeoDock, DiffDock-PP, EquiDock, and SMP, across 100 independent protein complex samples. Box plots indicate the mean (shown as '+'), median (center line), the interquartile range (IQR; 25th-75th percentiles), whiskers (5th-95th percentiles), and outliers (gray points). The fourth column presents the success rate (DockQ > 0.23) of each method. SMP achieves higher docking success rates and more accurate structural predictions than existing models. b Left: CRMSD comparison per sample between EquiDock and SMP on the DIPS-Het dataset. Right: Zoomed-in view of CRMSD values within the 0 ~ 5Å range for EquiDock and SMP, providing a clearer visualization of the dense region in the scatter plot shown in the left panel. c Performance comparison between EquiDock and SMP across different DockQ score bins. d Docking structure visualizations for three representative samples (PDB IDs: 6FMP, 5MZY, and 3CPT), predicted by GeoDock (first column), DiffDock-PP (second column), EquiDock (third column), and SMP (fourth column). Source data are provided as a Source Data file. Performance comparison in the protein-protein docking task. a Box plots showing the distribution of CRMSD (first column), IRMSD (second column), and DockQ (third column) for four docking methods: GeoDock, DiffDock-PP, EquiDock, and SMP, across 100 independent protein complex samples. Box plots indicate the mean (shown as '+'), median (center line), the interquartile range (IQR; 25th-75th percentiles), whiskers (5th-95th percentiles), and outliers (gray points). The fourth column presents the success rate (DockQ > 0.23) of each method. SMP achieves higher docking success rates and more accurate structural predictions than existing models. b Left: CRMSD comparison per sample between EquiDock and SMP on the DIPS-Het dataset. Right: Zoomed-in view of CRMSD values within the 0 ~ 5Å range for EquiDock and SMP, providing a clearer visualization of the dense region in the scatter plot shown in the left panel. c Performance comparison between EquiDock and SMP across different DockQ score bins. d Docking structure visualizations for three representative samples (PDB IDs: 6FMP, 5MZY, and 3CPT), predicted by GeoDock (first column), DiffDock-PP (second column), EquiDock (third column), and SMP (fourth column). Source data are provided as a Source Data file. To further show the advantage of SMP, we compare the CRMSD of each sample for SMP and EquiDock, as illustrated in Fig. 4b. 100 protein complexes are randomly sampled from the DIPS-Het test set to ensure both computational feasibility and representativeness. In 58 out of 100 cases, SMP achieves a lower CRMSD value and a more accurate complex structure than EquiDock. Moreover, Fig. 4b shows the visualization of 30 protein samples, with CRMSD values ranging from 0 to 5Å for both SMP and EquiDock predicted structures. Among these, 20 out of 30 instances exhibit a lower CRMSD for the SMP predictions compared to EquiDock, indicating that our SMP can obtain more high-quality and reliable structures than EquiDock. Fig. 4c presents the performance comparison between EquiDock and SMP in different DockQ score bins. In detail, 65 samples of SMP and 70 predicted proteins of EquiDock in the 0 ~ 0.23 range, 17 examples of SMP and 15 of EquiDock fall within the 0.23 ~ 0.4 range, 11 of SMP and 10 of EquiDock within 0.4 ~ 0.6, and 7 of SMP and 4 of EquiDock within 0.6 ~ 0.8, demonstrating that our pseudo data and pre-training strategy can help the model improve the quality of predicted protein complex structure. In addition, we evaluate SMP against other models on the full DIPS-Het test set of 1047 samples, with details provided in Supplementary Section S5. The results (Supplementary Fig. S2) show that SMP consistently delivers SOTA performance among existing deep learning-based approaches, achieving a success rate of 40.27%, substantially outperforming GeoDock (27.69%), DiffDock-PP (23.47%), and EquiDock (34.46%). We also visualize three representative samples, PDB IDs 6FMP, 5MZY, and 3CPT, predicted by GeoDock, DiffDock-PP, EquiDock, and SMP, as shown in Fig. 4d. It shows that SMP achieves better alignment with the ground-truth structures compared to other methods and attains the highest DockQ score among deep learning-based models, showing the effectiveness of our approach in certain protein complexes. In summary, our pre-training method achieves a new SOTA performance on the challenging docking benchmark DIPS-Het, improving the quality and accuracy of predicted structure, as well as the docking success rate. Concurrently, it also displays that SMP has a strong potential to benefit the protein complex structure prediction field. SMP learns discriminative protein representations We assess the ability of SMP to perform the protein-protein interaction (PPI) task. This task intends to find out whether two proteins bind, which is another essential problem for discovering signal transduction pathways38 and new drug targets39. The SMP is integrated with the high-performing method PPITrans28 and evaluated on multiple PPI benchmarks with different features like HIPPIE40 (balanced dataset) and D-SCRIPT41 (unbalanced dataset) to show its advantages. In detail, Fig. 5a illustrates the interaction prediction results between SMP and other popular methods in the D-SCRIPT test set, indicating that SMP outperforms four previous models and achieves SOTA performance. For example, SMP obtains the best recall of 0.62, the best F1 score of 0.72, and the best AUPR of 0.788, corresponding to an average improvement of 12.7% (p-value = 0.0026) in recall, 5.9% (p-value = 0.0076) in F1 score, and 3.0% (p-value = 0.0083) in AUPR compared with the current strong model, PPITrans. These results indicate that SMP can provide extra interaction prior to knowledge from pseudo-dimers and have good robustness in the unbalanced scenario. Furthermore, analogous results are also exhibited in the HIPPIE test set, as shown in Fig. 5b. Specifically, SMP performs better than the SOTA model PPITrans again, achieving an average improvement of 10.0% (p-value = 0.0138) in recall, 3.0% (p-value = 0.0080) in F1 score, and 1.8% (p-value = 0.0175) in AUPR, reaching the highest recall of 0.77, F1 score of 0.69, and AUPR of 0.728. These consistent improvements further demonstrate that the additional pseudo-dimer data can enhance the discrimination of protein features and improve the capability of the network to determine whether two proteins interact.Fig. 5Performance of SMP in the protein-protein interaction task.a,b Results on the D-SCRIPT (a, n = 93,024 independent protein pairs) and HIPPIE (b, n = 35,496 independent protein pairs) test sets with five runs using different random seeds. Precision-recall curves, recall, and F1-score show that SMP consistently outperforms other deep learning-based models across multiple metrics, with statistically significant improvements. For the precision-recall curves, results are reported as mean ± standard deviation. For recall and F1-score, statistical significance was assessed using a paired t test. c Cross-species task, where the model is trained on H. sapiens PPIs (HIPPIE dataset) and directly tested on D. melanogaster PPIs (GUO dataset). SMP shows greater transferability and generalization capabilities than other models. d Probability density of interaction scores for predicted positive samples by PPITrans and SMP, visualized as a ridge plot. SMP exhibits a higher score than PPITrans, indicating SMP can improve the model’s confidence score. Source data are provided as a Source Data file. Performance of SMP in the protein-protein interaction task. a,b Results on the D-SCRIPT (a, n = 93,024 independent protein pairs) and HIPPIE (b, n = 35,496 independent protein pairs) test sets with five runs using different random seeds. Precision-recall curves, recall, and F1-score show that SMP consistently outperforms other deep learning-based models across multiple metrics, with statistically significant improvements. For the precision-recall curves, results are reported as mean ± standard deviation. For recall and F1-score, statistical significance was assessed using a paired t test. c Cross-species task, where the model is trained on H. sapiens PPIs (HIPPIE dataset) and directly tested on D. melanogaster PPIs (GUO dataset). SMP shows greater transferability and generalization capabilities than other models. d Probability density of interaction scores for predicted positive samples by PPITrans and SMP, visualized as a ridge plot. SMP exhibits a higher score than PPITrans, indicating SMP can improve the model’s confidence score. Source data are provided as a Source Data file. We can also find that DeepPPI42, Topsy-Turvy43, and D-SCRIPT41 have lower performance than PPITrans. The maximum AUPR of these three models is 0.420 in the D-SCRIPT and 0.693 in the HIPPIE, respectively, which is lower than the AUPR of 0.775 and 0.713 in PPITrans. The reason is that these models have fewer model parameters than PPITrans. Specifically, these models only utilize the limited convolutional layers to extract the relation between various proteins, whereas PPITrans designs eight transformer layers for model training, showing that a deep and strong network can capture a more detailed interaction pattern than a shallow network. To evaluate the transferability and generalization of SMP, we employ it in the cross-species setting, where the model is trained on human PPI data40 and directly tested on the fruit fly species data44. Figure 5c shows the performance comparison between SMP and other methods. We notice that SMP surpasses all other models, achieving the highest precision, recall, F1-score, AUPR, and AUROC scores of 0.65, 0.735, 0.690, 0.672, and 0.702, respectively. Moreover, compared to the model PPITrans, SMP improves the precision, recall, F1-score, AUPR, and AUROC by 4.2, 14, 8.8, 12.2, and 6.8%, respectively. These results demonstrate that SMP presents good cross-species capabilities and has strong potential to be applied to other species. In addition, Fig. 5d illustrates the probability density of interaction scores for predicted positive samples by PPITrans and SMP. SMP achieves higher interaction scores than PPITrans, suggesting that our SMP can improve confidence in model predictions. In conclusion, SMP can bring significant performance gains across diverse PPI datasets with varying characteristics, demonstrating its superior capability and robustness in the PPI task. Surprisingly, SMP also shows remarkable results in cross-species situations, suggesting that it can help the model acquire additional interaction prior knowledge with minimal cost or modification. Applying SMP to data scarcity scenarios The pre-trained model has the potential to achieve satisfying performances trained only on small-scale data. Therefore, we aim to explore the effect of SMP and its corresponding baselines with limited training data. Specifically, we assess the model in data-insufficient scenarios on two tasks, including contact map prediction and protein docking, considering four data ratios, which belong to the set {0.1, 0.2, 0.3, 0.4}. Fig. 6a shows the contact performance comparison between SMP and DeepInter in different fractions of training data on the Homodimer300 test set. SMP outperforms the DeepInter across top 10, 50, L/10, and L precision metrics. Specifically, at data ratios of 0.1, 0.2, 0.3, and 0.4, SMP has average performance improvements of 6.59, 7.96, 4.58, and 5.56% compared to DeepInter, respectively, demonstrating that our pseudo-dimer data can provide prior knowledge that is relevant to the real dimer contact prediction and the pre-trained model can be effectively reused and transferred to the real dimer scenarios.Fig. 6Experiments in data scarcity scenarios.a Evaluation of the Homodimer300 dataset with varying fractions of training data in the contact map prediction task. SMP consistently outperforms the strong baseline model DeepInter across multiple metrics, including top 10, 50, L/10, and L precisions. b The line chart shows the performance comparison on the DIPS-Het dataset between SMP and EquiDock under limited training data in the protein docking task. SMP achieves lower CRMSD and IRMSD values alongside higher DockQ scores and success rates compared to the baseline model, EquiDock, demonstrating its robustness and superiority in low-resource settings. Source data are provided as a Source Data file. Experiments in data scarcity scenarios. a Evaluation of the Homodimer300 dataset with varying fractions of training data in the contact map prediction task. SMP consistently outperforms the strong baseline model DeepInter across multiple metrics, including top 10, 50, L/10, and L precisions. b The line chart shows the performance comparison on the DIPS-Het dataset between SMP and EquiDock under limited training data in the protein docking task. SMP achieves lower CRMSD and IRMSD values alongside higher DockQ scores and success rates compared to the baseline model, EquiDock, demonstrating its robustness and superiority in low-resource settings. Source data are provided as a Source Data file. Similar trends are observed in the docking task, as illustrated in Fig. 6. SMP can achieve lower CRMSD and IRMSD values, and higher DockQ score than EquiDock in varying training data ratios. For instance, SMP reduces the average CRMSD of 3.05, 1.59, 0.89, and 1.55 at data ratios of 0.1, 0.2, 0.3, and 0.4, respectively, relative to EquiDock, and lowers the IRMSD by 1.44, 0.94, 0.75, and 1.51 at the same ratios. Moreover, SMP improves the success rate of 4, 1, 3, and 8% compared to EquiDock at data ratios of 0.1, 0.2, 0.3, and 0.4, further indicating that noisy pseudo-dimer data, along with the model trained on it, can provide preliminary structural knowledge, and our SMP offers an alternative approach for utilizing monomer data. These experiments prove that our pre-training paradigm can effectively reduce the dependence on real data and make the model adapt to different volume-level training data situations, having the potential to save the extra cost of collecting multimer data. SMP in real-world applications We evaluate SMP on real-world protein complexes to assess its generalization capability. Specifically, we compare SMP with several high-performing multimer structure prediction models, including Chai-145, AF-Multimer5, and AF36, using protein dimers from CASP1546. We find that SMP outperforms other methods on two protein dimers from target IDs T1176 and T1113, both of which are intertwined protein complexes in which the two chains are topologically entangled. Figure 7a, b present the predicted contact maps and structures of these two targets. For SMP’s predicted structures, we employ ColabDock47, using SMP-predicted contact sites as constraints to optimize the protein structures. For Target ID T1176, Chai-1, AF-Multimer, and AF3 fail to predict the correct contact positions. In contrast, SMP accurately predicts almost the same contact regions as the ground truth, and the structure predicted using SMP constraints aligns better with the native structure than those from other models, achieving the highest DockQ score of 0.785, far exceeding AF3 (0.008), AF-Multimer (0.011), and Chai-1 (0.009). Similarly, for Target ID T1113, SMP achieves contact map predictions comparable to Chai-1, AF-Multimer, and AF3, while producing a highly accurate protein complex structure. Specifically, it obtains a DockQ score of 0.700, outperforming Chai-1 (0.038) and AF3 (0.616), and slightly surpassing AF-Multimer (0.630).Fig. 7Evaluation on CASP15 targets T1176, T1113, H1151, and T1173.a Contact maps. For T1176, SMP shows close agreement with the ground truth, whereas Chai-1, AlphaFold-Multimer, and AlphaFold3 miss key contacts. In contrast, for T1173, SMP fails to capture key contacts that are identified by the other methods. For T1113, AlphaFold3, AlphaFold-Multimer, and SMP accurately predict the contacts, while Chai-1 fails. For H1151, SMP predicts more contacts than the other methods, leading to increased false positives. b Protein complex structures. For T1176 and T1113, SMP generates structures closely resembling the ground truth and achieves the highest DockQ scores. By contrast, for H1151 and T1173, SMP fails to recover the native complex structure. Source data are provided as a Source Data file. Evaluation on CASP15 targets T1176, T1113, H1151, and T1173. a Contact maps. For T1176, SMP shows close agreement with the ground truth, whereas Chai-1, AlphaFold-Multimer, and AlphaFold3 miss key contacts. In contrast, for T1173, SMP fails to capture key contacts that are identified by the other methods. For T1113, AlphaFold3, AlphaFold-Multimer, and SMP accurately predict the contacts, while Chai-1 fails. For H1151, SMP predicts more contacts than the other methods, leading to increased false positives. b Protein complex structures. For T1176 and T1113, SMP generates structures closely resembling the ground truth and achieves the highest DockQ scores. By contrast, for H1151 and T1173, SMP fails to recover the native complex structure. Source data are provided as a Source Data file. However, SMP underperforms on certain conventional protein dimers, as shown in Fig. 7a, b for Target IDs H1151 and T1173. For H1151, SMP predicts more contacts than the other multimer structure prediction models, resulting in a higher number of false positives. Consequently, the SMP-predicted structure yields the lowest DockQ score of 0.143, substantially lower than AF3 (0.627), AF-Multimer (0.775), and Chai-1 (0.778). Similarly, for T1173, SMP fails to capture key contacts identified by the other models, and its predicted structure again attains the lowest DockQ score of 0.030, compared with AF3 (0.178), AF-Multimer (0.877), and Chai-1 (0.268). The remaining conventional protein dimers are shown in Supplementary Figs. S5 and S6. We hypothesize that the unsatisfactory performance of SMP on conventional complexes (e.g., H1151 and T1173) stems from the fact that current high-performing structure prediction models typically have substantially more parameters and are trained on much larger datasets. Consequently, they tend to achieve superior performance on conventional protein complexes. In contrast, SMP demonstrates superior performance on intertwined protein complexes (e.g., T1176 and T1113), surpassing the other models on this type of structure. This difference arises from the splitting operation of SMP: the random splitting has a high probability of generating pseudo-dimers in which the two chains are intertwined. As a result, pre-training on these randomly split pseudo-dimers, which frequently exhibit intertwined topologies, improves SMP’s capability to model intertwined protein complexes. Discussion In this study, we have introduced the Split and Merge Proxy (SMP), a simple yet effective pre-training framework designed to enhance protein and protein interactions tasks by leveraging the abundant available data from monomeric proteins. The scarcity of protein complex structure data poses significant challenges for predicting protein and protein interactions. SMP addresses this limitation by utilizing rich monomer information to simulate protein interactions as the pre-training goal. In detail, SMP involves splitting the monomeric proteins into pairwise pseudo-dimers with sequence and structure information, and training the model to merge them back by separately predicting the pseudo-interactions (e.g., contact or docking) for different PPIs tasks. Then the pre-trained model serves as an initialization for the respective task and is subsequently fine-tuned on the real protein dimer data. Because the pre-training and fine-tuning stages share a consistent goal, SMP achieves remarkable performance gains over other deep learning-based methods that do not exploit monomeric proteins in this manner. Our comprehensive experiments demonstrate that incorporating monomer-derived pseudo-dimers during pre-training substantially benefits performance across various protein interaction tasks. This improvement is evident not only in residue-level tasks, such as protein-protein contact map prediction and protein-protein docking, but also extends to sequence-level tasks, including protein-protein interaction prediction. In addition, SMP can perform better than AF-Multimer and AF3 on certain intertwined protein complexes that pose substantial challenges for existing structure prediction models, indicating the effectiveness of our pre-training strategy and data. SMP can also reduce reliance on real dimer data and adapt well to data-scarce scenarios, benefiting scenarios where real protein complex data is difficult to obtain. The SMP framework not only enhances predictive accuracy but also proves to be a versatile approach applicable to other multimer-related tasks. Although SMP has achieved impressive performance in various PPIs tasks and scenarios, it still has some limitations and room for improvement: First, our SMP currently mainly focuses on the dimeric proteins, we do not extend it to the multimeric proteins (i.e., three or more proteins), which would limit its applications in real-world scenarios; Second, we now only use monomers from the PDB, which primarily consists of experimentally determined structures, however, the AlphaFoldDB database48, comprising predicted monomeric protein structures generated by AF2, has been released, this database is ten times larger than the PDB, and leveraging this large-scale dataset for pre-training could further enhance the model’s performance; Third, the current SMP paradigm relies on training separate interaction models with different interaction labels generated from pseudo-dimer data, which may not be the most efficient strategy. Developing a unified pre-training framework capable of supporting multiple downstream interaction tasks within a single model may offer a more scalable and efficient solution. Finally, a more systematic investigation of the learned PPIs representations and the specific properties captured through SMP pre-training, while scientifically meaningful and bringing a broad range of interest, remains beyond the primary scope of the present study and warrants deeper exploration. Methods Task definition The PPIs prediction task aims to calculate the interaction relationship I between the two given proteins x1 and x2 based on the designed model f. This computation pipeline can be summarized as follows: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I=f({x}_{1},{x}_{2}),\,{x}_{i}=\left\{{s}_{i},{c}_{i}\right\},\,i\in \left\{1,2\right\},$$\end{document}I=f(x1,x2),xi=si,ci,i∈1,2, Where si and ci represent the amino acid sequences and their corresponding 3D coordinates in the unbound state of the protein xi, often used as input in protein interaction models. Different interaction tasks have various labels I. For example, the interaction label is the contact map in the protein-protein contact map prediction task, while it turns into the protein complex structure in the protein-protein docking task, and transforms into whether two proteins interact in the protein-protein interaction task. They can be expressed as follows: Protein-protein contact map prediction task For this task, the interaction relationship I is the contact map, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I\in {(0,1)}^{{L}_{1}\times {L}_{2}}$$\end{document}I∈(0,1)L1×L2. The L1 and L2 represent the chain lengths of each protein, respectively. The element Ii,j is 0 or 1, indicating whether the residue i-th in one sequence interacts with the residue j-th in another sequence or not, often determined by a distance threshold of 6Å or 8Å between the residue i-th and the residue j-th. Protein-protein docking task The interaction relationship I in the docking task consists of two components: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{1}\in {{\mathbb{R}}}^{{L}_{1}\times 3}$$\end{document}I1∈RL1×3 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}_{2}\in {{\mathbb{R}}}^{{L}_{2}\times 3}$$\end{document}I2∈RL2×3, which are the 3D coordinates of each protein in the bound state. Here, L1 and L2 denote the chain lengths of the respective proteins. Protein-protein interaction task This task aims to predict whether two proteins can interact with one another. Hence, its interaction label I has a binary value. If the I = 0, the two proteins do not interact; otherwise, they interact. Datasets In this section, we provide a detailed description of the datasets used in SMP, including our self-constructed pre-training dataset, as well as the datasets used for training, validation, and testing in PPIs tasks. Pre-training dataset We use the monomeric proteins from the Protein Data Bank (PDB)9, released before 2018-04-30, to generate the pseudo-dimer data. There are 60,206 PDB files in total, each containing the sequence and structural information of a protein. Monomers that cannot be parsed by Biopython49 are filtered out, as they contain unknown or missing atoms and disordered chain numbers. For consistency, only the primary biological assembly (assembly-1) is retained from each protein structure file, while alternative conformations are excluded6. In addition, we set the split range between 1/3 and 2/3 of the entire chain length so that the cut position is close to the middle of the given sequence to get pseudo-dimers. Too short split proteins, whose length of any chain is less than 20, are dropped. Finally, there are 22,589 pseudo-dimers, each comprising a sequence and its corresponding structural information. Because monomeric and multimeric structures are deposited as distinct PDB entries, the constructed pseudo-dimer dataset has no exact PDB-level overlap with the real multimer test sets. We further examined potential leakage using structure- and sequence-level similarity analyses across all downstream PPI tasks, as detailed in Supplementary Section S3. Only a small fraction of test samples showed high similarity to the pseudo-dimer set, defined as TM-score > 0.5 for structure-based benchmarks or sequence similarity > 0.4 for sequence-based PPI benchmarks (Supplementary Fig. S1). After excluding these potentially similar samples, the evaluation results remained essentially unchanged (Supplementary Tables S3, S4, and S5), supporting that the observed performance gains are not driven by data leakage. These findings suggest that the improvements brought by our pre-training dataset are not driven by sequence or structural similarity, but rather by the additional information introduced through pseudo-dimers. Given that different interaction tasks have various labels, the pseudo-dimer data should be processed to adapt to distinct PPIs tasks. For the contact map prediction task, we calculate the Euclidean distance between the heavy atoms of the residues on different pseudo-proteins and utilize a threshold of 8Å to cut off the distance to acquire the contact map, yielding 22,589 labels. The protein docking task aims to regress the bound protein complex structure from the two unbound proteins. Therefore, we randomly rotate and translate the coordinates of each pseudo protein to simulate the unbound state as input and regard their original pose as the ground truth pose, obtaining 22,589 targets. In addition, in the protein-protein interaction task, proteins from the same pseudo-dimer interact with each other and are considered positive samples, while those from different pseudo-dimers are regarded as negative samples. We construct two datasets with different positive-to-negative sample ratios (1:1 and 1:10) to accommodate both balanced and imbalanced protein-protein interaction datasets, resulting in 45,302 and 249,161 samples, respectively. Datasets used in the PPIs tasks Various PPIs tasks utilize different datasets to evaluate the capability of respective models. Here, we introduce the detailed datasets used in the PPIs tasks. Datasets in the contact map prediction task In this task, we conduct the experiments on both the homodimers and heterodimers, utilizing the same datasets as in the previous work, DeepInter8, for training and testing. Specifically, all the homodimers and heterodimers are downloaded from the PDB. MMseqs250,51 is used to cluster the homologous sequences to remove the redundant proteins, while Biopython49 is employed to filter out PDB entries that could not be parsed. This procedure yields a dataset comprising 3,470 homodimers for training, 294 for validation, and 297 for testing, with the test set referred to as Homodimer300. The heterodimers are filtered based on the following principles: (1) the protein complexes only contain two single chains; (2) the X-ray resolution of the experimental protein structure should be better than 3Å; (3) the total length of the complex cannot be too long, which need to less than 1,023 and the length of the single chain should be more than 50; (4) the sequence identity of the two chains needs to be less than the cutoff of 70%; and (5) the PDB entry can be successfully parsed by Biopython49. Then the heterologous sequences are clustered with a sequence identity threshold of 40% to improve protein diversity and reduce bias. The final heterodimers are divided into a training set containing 1881 heterodimers, a validation set including 96 heterodimers, and a test set consisting of 99 heterodimers, with the test set referred to as Heterodimer99. Dataset in the protein docking task DIPS-Het is a challenging benchmark for protein docking, derived from prior studies36,37 and curated as a subset of the DIPS dataset52, which includes experimental protein complexes sourced from the PDB. This dataset primarily focuses on heterodimeric proteins, which are prevalent in biological processes such as antibody-antigen interactions and enzyme-inhibitor complexes. In detail, the PDB entries in the DIPS-Het are included according to the following conditions: (1) the proteins should be collected by diffraction-based methods or electron microscopy; (2) the X-ray resolution of the experimental structure should be lower than 3.5Å; (3) the complex should contain less than 5 chains; (4) the complex only belongs to protein-protein type; and (5) the PDB entry can be successfully parsed by Biopython49. In addition, the proteins sharing the sequence similarity of 30% with the DB5.5 dataset53 are removed. Finally, the DIPS-Het dataset includes a training set of 9064 proteins, a validation set of 661 proteins, and a test set of 1050 proteins. Following previous works34,36, we randomly sample 100 protein complexes from the entire test set to ensure both computational feasibility and representativeness for evaluation. Datasets in the protein-protein interaction task Two datasets from the previous works40,41, HIPPIE and D-SCRIPT, are used to train and evaluate the protein-protein interaction task. The D-SCRIPT is an unbalanced dataset with 1:10 positive-to-negative samples. The interaction proteins, as the positive cases, are downloaded from the STRING dataset54,55. The length of each protein is between 50 and 800 amino acids. Moreover, highly redundant sequences with a sequence identity more than 40% are removed. In the end, the interaction proteins are split into a training set of 38,344 positives and a test set of 4794 positive pairs. The negative samples are generated from the positive set through the 1:10 ratio, leading to 383,448 negatives in the training set and 47,931 negative cases in the test set. The HIPPIE is a balanced dataset, with positive protein samples taken directly from HIPPIEv2.356 and negative samples randomly generated from the positive set. To prevent potential information leakage, the training, validation, and test sets are completely non-overlapping, and the sequence similarity between these three sets is as minimal as possible. The CD-HIT tool57,58 is applied to remove the redundant proteins using a cutoff of 40%. As a result, there are 35,096 positives and 33,562 negatives in the training set, 20,062 positives and 19,389 negatives in the validation set, and 17,898 positives and 17,598 negatives in the test set. Proteins from the CASP15 To guarantee the fair assessment and avoid potential data leakage, we sample the protein complexes that are released after 2023-05-01 from the CASP15, and the length of any chain cannot exceed 1024 amino acids, resulting in 8 multimers: T1113, T1132, H1151, T1176, T1178, T1179, and H1185, T1173, corresponding to PDB entries 7UYX59, 8ECX60, 8D5V61, 8SMQ46, 8UFN62, 8TN863, 8OUY64, and 8ON465, respectively. For evaluation, we use the A and B chains of each PDB target to form dimers. SMP details Model details The Split and Merge Proxy (SMP) is an effective pre-training task designed for interaction prediction tasks (e.g., contact or docking). The overall pipeline includes a split stage and a merge stage, as illustrated in Fig. 1a, d. In the split stage, each monomer is partitioned into two subunits to construct a pseudo-dimer, comprising a pseudo-receptor and a pseudo-ligand, that contains sequence and structural information. In the merge stage, the model is trained directly on the generated pseudo-dimer data to learn the target interaction prediction task, without any modification to the model architecture. Finally, the pre-trained model is fine-tuned on real dimer data. Split stage We utilize monomers from the Protein Data Bank (PDB)9, which provides both amino acid sequences \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\in {{\mathbb{P}}}^{L}$$\end{document}s∈PL and corresponding 3D coordinates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c\in {{\mathbb{R}}}^{L\times 3}$$\end{document}c∈RL×3. The objective of the split stage is to construct pseudo-dimers that share the same data structure as real dimers, consisting of two sub-sequences (s1 and s2) along with their associated structural information (c1 and c2). Specifically, we first partition the amino acid sequence into two sub-sequences: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${s}_{1}=s[:l],{s}_{2}=s[l:],$$\end{document}s1=s[:l],s2=s[l:], where l denotes the split index. This index can be determined either by a domain segmentation tool or by randomly sampling in a position range around the center of the monomer sequence. For the 3D structure, we apply an analogous split operation: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${c}_{1}=c[:l],{c}_{2}=c[l:].$$\end{document}c1=c[:l],c2=c[l:]. Merge stage This stage essentially mimics the standard learning paradigm of interaction prediction. We employ task-specific, high-performing models, keeping their original network architectures, and pre-train them individually on our generated pseudo-dimer data. The model architectures are described below. Protein-protein contact map prediction task We adopt the DeepInter architecture8 for the contact map prediction task. The model integrates multiple input representations, including evolutionary features derived from single and paired multiple sequence alignments (MSAs) generated by HHblits66 and CCMPred67, sequence embeddings extracted from the ESM-MSA-1b model68, and structural priors encoded as intra-protein distance maps, with stacked ResNet-Inception8 and triangle-aware transformer8 blocks to capture both short- and long-range information. Finally, the model is pre-trained as a binary classification task using the focal loss function8,69 to address class imbalance in contact map data. Protein-protein docking task We adopt the Equidock34 framework, the first deep learning-based model for the protein docking task. The model comprises a series of SE(3)-equivariant graph networks 70 to capture the rotation and translation of the unbound protein complex to determine the final structure. The whole framework is pre-trained using a mean squared error (MSE) loss34 to regress the predicted atomic coordinates to the ground-truth structures. In addition, a pocket loss34 based on optimal transport is incorporated to predict the binding pocket on the interaction interface, as well as an intersection loss34 to prevent atomic collisions. Protein-protein interaction task We adopt the PPITrans28 architecture, which contains extracted sequence features using ProtT571, multiple transformer blocks72, and a classifier with multiple multilayer perceptrons (MLPs) to predict the interaction between two proteins. The model is pre-trained through the standard cross-entropy loss function28. Fine-tuning stage After pre-training on the pseudo-dimer dataset, each task-specific model is initialized with the corresponding pre-trained weights. As the pre-training and fine-tuning stages share the same task formulation, no task gap is introduced. The models are then fine-tuned end-to-end on real dimer data. Implementation details All the models are trained using the PyTorch framework73. The experiments of protein-protein contact map prediction and protein-protein docking are performed on the A100 GPU with 80GB memory, and the experiments of protein-protein interaction are performed on the GeForce RTX 4090 GPU with 24GB memory. The MSAs generated by HHblits in the contact map prediction task are based on the UniRef30_2021_0351 database74. The implementation details for these three downstream tasks are provided in Supplementary Section S6, with the corresponding hyper-parameters listed in Supplementary Tables S7, S8, and S9. For the AlphaFold3 and Chai-1 inference, we utilize their respective online servers: AlphaFold3 Server (https://alphafoldserver.com/) and Chai-1 Server (https://lab.chaidiscovery.com). AlphaFold-Multimer inference is conducted locally on an NVIDIA A100-80GB GPU with the v2.3.0 model weights, using the default databases (BFD4, MGnify75 (v2022_05), PDB7076 (cutoff: 1 November 2021), PDB_mmCIF77 (cutoff: 1 November 2021), PDB_seqres77 (cutoff: 1 November 2021), UniRef3074 (v2022_02), UniProt78 (v2020_05), and UniRef9079 (v2020_01)) for MSA and template searches, and enabling AMBER relaxation for structure refinement. All structure prediction models are inferred five times, yielding five structure predictions for each sample. Subsequently, the top-ranked (rank-1) structures from AlphaFold3, Chai-1, and AlphaFold-Multimer are selected for further analysis. Evaluation metrics Different PPIs tasks use various metrics to evaluate the performance of the model. Here, we introduce the detailed metrics in these PPIs tasks. Metrics used in contact map prediction task All the experiments in contact map prediction follow the standard evaluation protocol in previous works8. To assess the accuracy of the prediction, the top k precisions (P@k) are adopted as the evaluation metrics, where k ∈ {1, 10, 25, 50, L/10, L/5, L} with L is the length of the shortest chain, formulated as follows: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P@k=\frac{{\sum }_{i=1}^{k}{{\bf{1}}}(\widehat{{y}_{i}}=1)}{k},$$\end{document}P@k=∑i=1k1(yi^=1)k, where the k means the number of selected residues with the top-k highest predicted score, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{y}_{i}}$$\end{document}yi^ refers to the predicted label, which means that the score of the model output is higher than 0.5. 1 denotes the indicator function. If \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\widehat{{y}_{i}}=1$$\end{document}yi^=1 is true, the function value is 1, and vice versa. Metrics used in protein docking task In this task, Complex Root Mean Square Deviation (CRMSD), Interface Root Mean Square Deviation (IRMSD), and DockQ metrics are used to evaluate the model’s performance, consistent with previous works34,36. Given the ground truth and predicted protein complex structures: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I}^{*}\in {{\mathbb{R}}}^{3\times ({L}_{1}+{L}_{2})}$$\end{document}I*∈R3×(L1+L2) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I\in {{\mathbb{R}}}^{3\times ({L}_{1}+{L}_{2})}$$\end{document}I∈R3×(L1+L2), and then the Kabsch algorithm80 is utilized to align them. The CRMSD can be calculated as follows: 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CRMSD=\sqrt{\frac{1}{{L}_{1}+{L}_{2}}{\left\Vert {I}^{ * }-I\right\Vert }_{F}^{2}},$$\end{document}CRMSD=1L1+L2I*−IF2, where the L1 and L2 are the chain lengths of different proteins, and the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\left\Vert .\right\Vert }_{F}$$\end{document}.F means the Frobenius norm. A similar computational process is also leveraged in the IRMSD, but only considering the coordinates of the interface residues that are within 8Å of the other protein’s residues. The metric DockQ81 is a combination of different independent quality measures: fnat, Lrms, and Irms, which are proposed and standardized by the Critical Assessment of Predicted Interactions82 (CAPRI). The whole calculation can be summarized as follows: 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$DockQ=\frac{1}{1+\left(\frac{1}{3}{\left(\frac{{f}_{nat}+{L}_{rms}+{I}_{rms}}{0.01}\right)}^{-1}\right)},$$\end{document}DockQ=11+13fnat+Lrms+Irms0.01−1, where the fnat denotes the recall of predicted contact residues between the proteins. The Lrms is the RMSD calculated for the ligand (shorter chain) after superposition of the receptor (longer chain), and the Irms is the RMSD of aligned interface residues (residues within a distance of 8Å of the other protein’s residues). Metrics used in protein-protein interaction task This task is a binary classification problem. We follow the same evaluation criteria as prior studies28,41, including precision, recall, F1-score, area under the precision-recall curve (AUPR), and area under the receiver operating characteristic curve (AUROC) metrics. The detailed equations of precision, recall, and F1-score are shown as follows: 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{Precision}}}=\frac{TP}{TP+FP},$$\end{document}Precision=TPTP+FP,8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{Recall}}}=\frac{TP}{TP+FN},$$\end{document}Recall=TPTP+FN,9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{F1}}}-{{\rm{score}}}=\frac{2\times {{\rm{Precision}}}\times {{\rm{Recall}}}}{{{\rm{Precision}}}+{{\rm{Recall}}}},$$\end{document}F1−score=2×Precision×RecallPrecision+Recall, where the TP, FP, and FN represent the true positive samples, false positive samples, and false negative samples. In addition, AUPR denotes the area under the precision-recall curve, which is suitable for the classification task with imbalanced classes, while the AUROC means the area under the receiver operating characteristic (ROC) curve, which is commonly used to evaluate classification performance. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data