Integrating AlphaFold2, RoseTTAFold2, and HADDOCK to refine Protein-Protein Interaction (PPI) candidate selection: A case study with ZWINT Authors: Nora Abdelfattah, Bridget Broncales, Ivana Albert, Jaimee Kohan, Lei Yu Published: 2026 ABSTRACT AbstractProtein-protein interactions (PPIs) underpin most cellular processes, and disruptions to these interactions can lead to cellular dysfunction and disease. Understanding PPIs is essential for studying disease mechanisms, yet traditional experimental approaches are time- and labor-intensive. Recent advances in AI-based structural prediction tools, including AlphaFold2 and RoseTTAFold2, now enable efficient in silico exploration of potential PPIs. To develop an integrated and practical multi-tool system for PPI investigation, we present a dual-arm computational pipeline centered on the ZWINT (SIP30) kinetochore protein, which we identified as a key gene in neuropathic signaling.The first arm of the workflow generates PPI models using AlphaFold2 and RoseTTAFold2 and evaluates model consistency using TM-Align. The second arm assesses binding affinity by identifying interface residues with PyMOL and calculating docking scores with HADDOCK. Together, these methods provide both quantitative and qualitative evaluations of candidate PPIs. Using this framework, we examined three established interactors (SNAP25, CAMK2A, UBC) and four exploratory proteins (STX1A, VCP, BLOC1S2, ARC) in combination with ZWINT.This study demonstrates that AI-supported in silico analysis can streamline PPI discovery by prioritizing biologically plausible interactors and guiding downstream experimental validation.Keywords: Protein-protein interactions, AlphaFold2, HADDOCK, TM-Align, RoseTTAFold2, PyMOL Introduction Protein-protein interactions (PPIs) serve as the molecular basis for the majority of biological phenomena, from structural assembly and catalytic activity, to signal transduction and intracellular transport [1]. Abnormal PPIs, those displaying absent, improper, or altered folding and binding affinity, are the foundation for various disease states [2]. Therefore, studying PPIs is crucial to medical research and pharmacological development. Traditionally, identifying these interactions utilizes experimental methods such as yeast two-hybrid assays, co-immunoprecipitation, and cross-linked mass spectrometry. These methods, while helpful in determining PPIs, can be costly and time-consuming, especially in the context of the large number of potential PPIs within the human proteome [3]. The advent of AI-based structural prediction programs, primarily through variants of AlphaFold2 and its successors, has the potential to enable the digital pre-screening of entire candidate lists for PPI compatibility before engaging in costly laboratory efforts. Such programs are trained on existing crystallography structures and are subsequently capable of predicting the tertiary structure with exceeding accuracy based solely on the amino acid sequence. These efforts have paved the way for subsequent modeling of interaction [1]. Although the precision of structure prediction has improved with tools such as AlphaFold2 and RoseTTAFold2, it is not easy to transform these predicted tertiary structures into biologically meaningful PPI models [4]. The structural plausibility does not necessarily guarantee that the interactions are stable or physiologically relevant [5]. In addition, it has been noted that multimer confidence scores such as ipTM and pTM scores might not sufficiently distinguish between alternative models of interaction, especially where proteins contain intrinsically disordered regions [6]. There are docking tools such as HADDOCK can predict interaction energetics. However, it has been noted that docking predictions are heavily dependent on the input structure [7,8]. Recent studies have also demonstrated that machine learning approaches, including unsupervised learning methods applied alongside atomistic simulations and docking calculations, can improve the analysis of bioactivity and molecular interactions [9]. These methods can help identify patterns within complex molecular datasets and extract relationships that may not be readily apparent from docking scores or structural predictions alone. Therefore, it is not advisable to rely on a single tool. Instead, it is advantageous to combine several tools and to validate, analyze, and enhance each other's predictions. The SNAP25 Interacting Protein 30 or “SIP30” [[10], [11], [12]], also known as the ZW10 Interactor “ZWINT”, ensures the proper separation of chromosomes during mitosis by stabilizing kinetochore-microtubules. Beyond this conventional mitotic role, ZWINT has been implicated in neuropathic pain pathways and synaptic regulation [11,[13], [14], [15]], indicating a more multifunctional role than previously thought. Several key interactors of the ZWINT protein have been discovered [14,16,17], but many other possible interactors have yet to be described. Such is the case for several promising, but understudied proteins; hence our case for assessing the abilities of an integrated AI-based PPI pipeline. ZWINT was chosen as our model gene because, despite its established role in mitosis, its role and involvement in neuropathic pain signaling and synaptic plasticity remains poorly defined [18]. This makes it an ideal candidate in testing whether an AI-assisted workflow can reveal interactors that would be more difficult to find using traditional approaches. This case study uses ZWINT to test the workflow of in silico tools, including both AI and traditional software tools, to determine the likely binding candidates of SIP30 prior to experimental verification. This workflow is put forward in the form of a two-pronged computational pipeline involving several tools used consecutively. These software programs, built based on AI, though helpful, are still in the process of being perfected. Another consideration in utilizing these tools is that each program possesses certain strengths and weaknesses. By using multiple tools in tandem with one another instead of separately, one might hope to find an optimal combination of tools that may supplement and complement each other. The first arm is intended to measure model consistency and confidence, while the second predicts the binding probability of the generated complex to select the most promising interactants from a curated set of candidates. This paper puts forward the prospect that AI is not to be utilized as an absolute force for determining these PPI candidates, but rather as the basis to develop a formalized, hypothesis-generating filter, thereby optimizing and accelerating real-world PPI discovery by prioritizing candidates with higher binding probability and confidence. Methods Novel methodological considerations This work provides a novel structural modeling pipeline in which AlphaFold2 and HADDOCK are used not independently, but synergistically. While earlier work independently used AlphaFold2 for prediction of structures and HADDOCK for docking, and some studies have placed outputs from both programs side-by-side [19], our approach integrates the two frameworks into a single pipeline, wherein the output from one tool directly affects and optimizes the performance of the other. Given that our approach integrates previously established tools in a single workflow, we are presenting the methodology under the Results. Software access Access to AlphaFold2 and RoseTTAFold2 was provided by our collaborator, Neurosnap [20]. TM-Align [21] and HADDOCK [22] are accessible online, courtesy of the Zhang Group and Bonvin Lab, respectively. PyMol was accessed through a publicly available open source one as a one-click installer [23]. The interface residues script used in PyMol was accessed from a publicly available GitHub repository [24]. The iterate residues script was produced by our lab, and has also been made available on GitHub [25]. Protein multimer modeling steps with AlphaFold2 and RoseTTAFold2 Binding complexes were selected by having one recurring protein of interest, in our case, “ZWINT”, accompanied by either known binding proteins or exploratory candidate proteins identified via Qiagen's Ingenuity Pathway Analysis (IPA) and confirmed through subsequent literature review to have an interactor in common with the initial protein. The UniProt entries for the Rattus norvegicus version of each protein would be used to find the sequence for each protein. Afterwards, AlphaFold2 (AF) would be opened on the Neurosnap platform. Note that AlphaFold2-Multimer version 2.x was utilized. Once available, the amino acid sequence of the protein of interest (ZWINT) and the exploratory candidate or known binder would be inputted in the designated area and separated by a semicolon. The names of the proteins were recorded in the “notes” field, but all other settings were left on default. The job would then be permitted to run. RoseTTAFold2 (RF) was accessed and run in parallel to AF through Neurosnap. It had the same input, with the protein of interest and potential binder being fed to RoseTTAFold2 and run with default settings. Protein model cross-comparison with TM-Align Following the completion of AF and RF tasks, the generated PDB files were downloaded (five AlphaFold2 and One RoseTTAFold2 structure). TM-Align was accessed virtually. The RoseTTAFold2 PDB file was uploaded as “Sequence 1” and each AlphaFold2 program database (PDB) file uploaded as “Sequence 2” on different runs. TM-Align would be run, and the N-value on the results page was recorded. The highest-ranked AF structure was selected for subsequent analysis, which would be conducted using PyMOL and HADDOCK. Determining binding interface with PyMOL The Open Source variant of PyMOL would be used for the remainder of this study. The selected AlphaFold2-generated PDB file was opened in PyMOL, and an interfaceResidues.py script was run to identify residues within a 5 Å radius of one another. This was followed by an iterate command, which listed all the identified residues. The positions of these residues and the protein or “chain” they were on were recorded for future use. These residues comprised the binding interface. Determining binding score with HADDOCK Binding interface analysis was also performed using the online tool “High Ambiguity Driven protein-protein DOCKing”, or HADDOCK. A new job would be submitted utilizing the selected PDB file as both Molecule 1 and Molecule 2, albeit when prompted “Which chain of the structure must be used?” Chain A and Chain B were selected, respectively. When prompted “What kind of molecule are you docking?” each molecule was indicated to be “Protein or Protein-Ligand”. Neither molecule was coarse-grained or indicated to be a cyclic peptide. On the next page, the amino acid sequences identified via PyMOL as participating in the binding interface were inputted for each chain in a comma-separated format. All other settings were left on default. The research was not marked for COVID-19 research which was the final question prompted by the HADDOCK system. The job would then be submitted and left to run, and after several hours, the HADDOCK scores would be available. These scores would be noted alongside the TM Scores and the HADDOCK reports saved, as they would be removed from the HADDOCK webserver after a period of 30 days. Availability of resources The workflow described above was performed manually, including the input, retrieval, and extraction of files, statistics, or graphics from multiple websites and software platforms. No automated process currently exists to link all the steps together. A detailed standard operating procedure (SOP), as well as a screencast demonstrating how to access and use the various tools comprising the workflow, is available as supplementary material to this journal article. Results Overview of workflow integration The overall workflow is shown in Fig. 1. To explore protein-protein interactions (PPIs) for ZWINT, we utilized a two-path workflow approach, integrating AlphaFold2, RoseTTAFold2, and TM-Align into one arm, followed by AlphaFold2, PyMOL, and HADDOCK into a second arm. With this approach, we were able to examine two essential properties for each pair of protein: structural model accuracy based on model similarity and binding affinity based on predicted interaction strength. Combined workflow diagram of the workflow of AlphaFold2, RoseTTAFold2, TM-Align, PyMOL, and HADDOCK. The left one is the AlphaFold2-RoseTTAFold2 estimate of model accuracy. The right one is the HADDOCK docking score, which estimates the binding affinity of PPI models. Each of the principal software steps has a graphical output, demonstrated here using the ZWINT–SNAP25 complex. The figure was created with Biorender [26]. For the confidence path, amino acid sequences for the binding sequences were obtained from UniProt, then submitted to AlphaFold2 and RoseTTAFold2 through Neurosnap's online platform. AlphaFold2 (AF) generated five models; RoseTTAFold2 (RF) generated one. Each AF model was compared with the RF model using TM-Align, which would superimpose the structures atop one another. The AF model with the highest TM score, thereby having the closest similarity to the RF structure, was selected. For the binding affinity path, the highest-ranked AF model from the confidence path was viewed in PyMOL, which was then asked to highlight and input the residues participating in binding via an “interface residues” script. These residues and models were uploaded into HADDOCK to model binding and generate a binding score. This strategy not only permits structural prediction but also ranks candidate protein-protein interactions (PPIs) for experimental confirmation, hence the notion of an AI-enabled hypothesis pipeline. Experimental methodology: Confidence scoring via model agreement AlphaFold2 generates five structural predictions for each protein dimer, with each model arising from a different random initialization of its neural network. All models are trained on roughly the same publicly available crystallographic data, but the stochastic elements of the network introduce variation in the resulting structures. These differences are usually subtle but become more noticeable in flexible or intrinsically disordered regions of the proteins. AlphaFold2's 3D models are usually ranked by the composite score 80ipTM plus 20pTM [27]. But due to the disordered regions present in several proteins modeled, these values and other confidence scores from multimers are no longer believed to be reliable measures that truly reflect structural uncertainty between different PPI candidates [6]. Therefore, the confidence and rankings from the initial AlphaFold2 analysis were not used, except to identify the origin of each model. However, an ideal model would eventually account for ipTM scores. Instead, each AF model was compared with a reference structure generated by RoseTTAFold2 (RF), an externally trained artificial intelligence prediction tool. While AF would yield five possible tertiary structures, RF would generate only one structure per input. Given the different training models and parameters of each AI, RF was used as an external control by comparing it against each AF model. This comparison would be done using the tool TM-Align, which would superimpose each AF model-generated structure over that of the RF reference and return a TM Score based on how well the structures aligned. The AF model with the highest TM Score was selected for further analysis. Using the technique described with the tools in Table 1A, we bypassed AlphaFold2's ranking and opted for an external validation process based on structure. Table 1Bioinformatics tools used in dual-arm pipeline.Table 1Table 1A. Bioinformatics Tools Used to Assess Model confidenceToolPurposeOutputApplicationAlphaFold2Predicts binding configurations5 structural modelsGenerate and compare candidate models [4]RoseTTAFold2Alternate prediction for comparison1 structural modelExternal control for TM score comparison [3]TM-AlignMeasures similarity (AF vs RF)TM Score (%)Determines the highest-confidence AF model [28]Table 1B. Bioinformatics Tools Used to Assess Binding StrengthToolPurposeOutputApplicationAlphaFold2Provides top model3D structureInput for interface analysis and docking [4]PyMolExtracts binding residuesInterfacing residuesRequired Input for HADDOCK [[29], [30], [31]]HADDOCKSimulates binding & scoringHADDOCK ScoreMeasures interaction strength [32,33] Bioinformatics tools used in dual-arm pipeline. Experimental methodology: Binding affinity estimation After our pipeline picked the best AlphaFold2 model through the TM-Scoring process, we used PyMOL to identify the binding interface of the resulting structure. A Python script called “interface residues” was run on the AF model with the highest TM Score. This script identified amino acid residues within 5 Å of the opposite chain [24]. We highlighted these amino acids on each chain and then listed them. In short, PyMOL helped to highlight the binding interface of a protein complex and printed out the participating amino acid residues. These residues, as described in Table 1B, along with the structural model, would be submitted to the final software tool in this pipeline, the “High Ambiguity Driven protein-protein DOCKing” software, or HADDOCK. This program can simulate the docking of two or more proteins by taking each protein and coming up with many different configurations using the identified interface as the “Active Residues”. Afterwards, these configurations are clustered based on similarity and energy profile. Finally, each cluster would be identified with a HADDOCK score, which accounts for variables such as the van der Waals energy, electrostatics, desolation energy, buried surface area, RMSD from the lowest-energy conformer, cluster size, restraints violation energy, and Z-score or cluster quality. For our purposes, the lowest HADDOCK score, which corresponds to the highest-ranking cluster, was used to evaluate binding probabilities. Negative or lower HADDOCK scores indicate more stable and strong binding interactions. Pilot study: ZWINT and partner proteins To test our workflow, we chose ZWINT because of its known but not fully defined roles in the outer kinetochore and neuropathic pain signaling [11,34,35]. We used the rat (Rattus norvegicus) version of the ZWINT protein sequence. These models and data generated during this experiment are available via the Figshare data repository [36]. Three known interactors of ZWINT were selected from its Biogrid profile [37] to act as positive controls: CAMK2A, SNAP25, and UBC. A direct interaction between SNAP25 and ZWINT was demonstrated by a GST-pull-down and immunoprecipitation assay [10]. Similarly, experimental evidence supports protein-protein interaction between CAMK2A and ZWINT [38]. Interaction between ZWINT and UBC was validated by synaptic protein ubiquitination [39]. Thus, these three partner proteins were chosen as positive controls. Four additional proteins (ARC, STX1A, VCP, and BLOC1S2) were selected as exploratory candidates of potential ZWINT interactors, as suggested through Ingenuity Pathway Analysis (IPA, by QIAGEN, Germantown, MD) [40,41]. IPA was utilized to determine secondary interactors of ZWINT which had an established interaction with the primary interactors of ZWINT (CAMK2A, SNAP25, UBC). These secondary interactors were subsequentially treated as exploratory candidates to determine if they may have a direct interaction with ZWINT. For instance, the molecular mechanism of cholesterol establishment involves the protein-protein interaction between SNAP25 and STX1A [42]. Another interaction is between VCP and UBC, interconnected in the study of retrotranslocation from the endoplasmic reticulum to the cytosol [43]. Also, ARC and CAMK2A are integrated by ARC ubiquitination, regulating endoplasmic reticulum mediated calcium ion release and CAMK2A signaling [44], reinforcing the connectivity of these candidates and ZWINT. BLOC1S2 has been implicated in neuropathic pain overlapping with the process of ZWINT [45]. These exploratory candidates were tested using the same experimental methodology as the known interactors. For negative controls, three proteins from distinct biological backgrounds were selected because they were considered unlikely interactors of ZWINT, which is a nuclear protein associated with neuronal systems. Two of these proteins, KRT83 and COL1A1 are fibrous proteins found in the connective tissue. The third, HSD17B10, is a mitochondrial enzyme. While it is impossible to guarantee the absence of an interaction between two proteins without explicitly testing for it, selecting proteins that had minimal biological or functional overlap with ZWINT proved to be a reasonable alternative strategy. The proteins selected, reasoning behind their selection, and subsequent TM Scores and HADDOCK Scores are described in Table 2. Table 2Proteins Selected to Model Dimerization and associated Confidence and Binding Strength Scores.Table 2ZWINT Dimerizing ProteinCategoryReason SelectedPose Confidence (TM-Score)Binding Strength (HADDOCK Score)SNAP25 [46]Positive ControlMass Spectrometry Data and Reconstituted Complex indicated on Biogrid [12].n = 0.985−196.0 ± 15.1CAMK2A [47]Positive ControlMass Spectrometry data indicated on Biogrid [38].n = 0.928−148.5 ± 6.7UBC [48]Positive ControlMass Spectrometry data indicated on Biogrid [39].n = 0.550−63.1 ± 12.2STX1A [49]Exploratory CandidateBoth proteins interact with SNAP25 [42].n = 0.826−158.8 ± 9.8VCP [50]Exploratory CandidateBoth interact with UBC [43].n = 0.930−82.8 ± 5.6BLOC1S2 [51]Exploratory CandidateBoth proteins are implicated in neuropathic pain pathways [45].n = 0.820−49.3 ± 9.5ARC [52]Exploratory CandidateBoth interact with CAMK2A [44].n = 0.095−206.5 ± 12.8KRT83 [53]Negative ControlKeratin is restricted to the hair, skin, and nails [54].n = 0.133119.1 ± 36.8COL1A1 [55]Negative ControlCollagen is expressed predominantly extracellularly [56].n = 0.091128.7 ± 11.3HSD17B10 [57]Negative ControlGene expresses a mitochondrial enzyme [58].n = 0.223−49.5 ± 23.0 Proteins Selected to Model Dimerization and associated Confidence and Binding Strength Scores. The TM Scores and HADDOCK scores of the positive and negative controls were used as references to compare with the exploratory candidates, including setting a soft baseline for an acceptable score. Note that no hard cutoffs were used in any component of this study, as our objective was to develop a ranking system that would depend on the context of each individual protein. Pilot study: TM scores vs HADDOCK scores The data from this pilot study are described in Table 2. A correlation was observed between the TM Score and the HADDOCK score. This means that AF models that are more like RF also had better docking results, as seen in Fig. 2. Both HADDOCK and TM Scores consider the interfaces. If the amino acids are more closely aligned, fewer steric clashes can occur, which would raise TM Scores and lower HADDOCK scores [[61], [62], [63]]. However, there is still some variation in this correlation. For instance, SNAP25 has a lower HADDOCK score, even though its TM Score is 5% higher than ARC's. Other factors, besides model quality and steric hindrance, might also affect docking results. Therefore, it makes sense to combine the two measures and use the TM Score as a filter rather than a strict criterion. Rubric and Rating of Predicted Protein-Protein Interaction Models using TM Score and HADDOCK Metrics. Left Panel. This is a rubric used to assign meaning to TM Scores (x-axis) and. HADDOCK scores (y-axis) scores. Structures belonging in the top left quadrant (high TM, low HADDOCK) have the presumed best binding potential, due to the low HADDOCK score, and presumed best accuracy, due to the high TM Score. While these would be considered ideal candidates for testing, their interaction will not be confirmed until they are proven to interact through real-world experimentation. The Bottom-left quadrant (low TM, high HADDOCK), contains docking candidates with high model agreement and binding strength that may be worthy of exploring depending on how their data compares to known binding interactors. The Top-Right quadrant (low TM, low HADDOCK) indicates low structural agreement despite the high binding affinity. Here, factors such as the interactor's promiscuity should be considered. The Bottom-Right quadrant (low TM, high HADDOCK) indicates low structural agreement and bad binding, making it the worst possible candidate to pursue and the only one that can be entirely ruled out. Overall, this rubric is meant to enable triage and prioritization of PPI candidates for experimental follow-up. Right Panel. HADDOCK vs. TM-Align Scatterplot of all dimer pairs. A Pearson Correlation Coefficient of r = −0.86 was found, indicating a strong negative correlation between the TM Score and HADDOCK scores [59]. Cluster 1 (upper-left) has high modeling and high Pose Confidence pairs consistent with the top-left quadrant. Cluster 2 has moderate to weak binding with low model consensus, consistent with the bottom-left quadrant. Cluster 3A and 3B have weak binders with low confidence, consistent with the bottom right quadrant. Figures created using Biorender [60]. Discussion Explanation of approach The novelty of this study lies in the specific workflow and sequential integration of established computational tools into this pipeline, rather than the development of new software or the general concept of combining multiple prediction platforms or in silico prediction and assessment platforms. Previous studies have discussed or integrated these tools together in a variety of manners. These include head-to-head comparisons of RoseTTAFold and AlphaFold [[64], [65], [66]], using HADDOCK to score AlphaFold prediction [[67], [68], [69]], and using TM-Align to evaluate structural similarity of predicted complexes generated by RF or AF [[70], [71], [72]]. The contribution of this work therefore lies in the structured and sequential application of these tools in a unified pipeline. Organizing these existing methods into a single, systematic workflow enables this project to provide a practical framework in prioritizing PPI candidate selection. While HADDOCK scores of low values are thought to correlate with higher binding affinity, there are no hard cutoffs in HADDOCK scores for what would predict a binder in an empirical experiment [[73], [74], [75]]. In this case study, the applied computational pipeline gave relatively negative binding scores for all the confirmed interactors (SNAP25, CAMK2A, UBC) and similar scores for most of the exploratory candidates. While the candidate with the best binding score, ARC (HADDOCK = −215), had not been previously reported in the literature in connection with ZWINT, this pipeline does indicate that physical binding would be possible. In biological environments, ZWINT and ARC may exist in the same cellular compartment because they share a role in kinetochore assembly or possibly in neuronal function, such as synaptic plasticity. This aligns with what we know about activity-dependent cytoskeleton regulation. STX1A also had a high docking score that is in line with SNARE-complex function, as it is well-documented in the literature to form this complex alongside ZWINT and SNAP25. Regardless of biological interaction, however, both ARC and STX1A require experimental follow up to determine their biological relevance and ability to bind ZWINT. Another important point, which is highlighted by AI tools in experimental probing in this study, is that hard numerical cutoffs in docking scores often cannot capture important biological context. For example, enzymes UBC and VCP both generated rather modest docking scores (∼–70), which may correspond to the intrinsically reversible and transient nature of enzyme interaction rather than poor binding. Taken together, these observations further highlight the importance of ensuring that AI-predicted interactions are translated as hypothesis-generating leads rather than proof. One of the negative controls selected, HSD17B10, is a mitochondrial enzyme. It had a higher HADDOCK score comparable to those of UBC and VCP, and a substantially lower TM Score. Our case study example further highlights the point that strong AI-predicted interactions are often, but not always, in agreement with existing experimental data. The absence of a predicted interaction does not eliminate the possibility of a protein–protein interaction occurring in vivo as subcellular localization, cofactor availability, and protein concentration all play an important role in binding and will not be fully recapitulated by HADDOCK's rigid-body docking. As a result, this type of pipeline is better applied as a decision-aid to filter and rank candidates for empirical follow-up. Study limitations Limitations of software Even though the case study reported here suggested a low-cost, scalable solution to PPI assessment using currently available AI-based algorithms, there are limitations that should be noted. Several of the tools utilized here make static, idealized predictions based on structure that do not fully reflect the dynamic, context-dependent nature of PPIs in cells. Many biologically interesting protein-protein interactions (PPIs) are temporary, depend on specific conditions, undergo post-translational modifications, or happen in situations not shown here. First, using TM score as a measure of structural correctness assumes that similarity to an independently generated RoseTTAFold2 model directly relates to biological assembly. While this can help select the most likely AlphaFold2-derived structure, it cannot replace experimental confirmation and may unintentionally disadvantage conformational flexibility [5]. Second, HADDOCK provides a useful assessment of predicted binding free energy but is not definitive proof of biological interaction. Docking algorithms also rely on the selection of input residues, and automated binding site definitions may not always identify biologically relevant contact surfaces. It remains to be seen to what extent the in vitro docking findings correlate with in vivo binding, since subcellular targeting, expression levels, and biochemical conditions all play a role in interaction potential [8]. Third, use of AlphaFold2 and RoseTTAFold2 may possibly generate false negatives. Despite their extensive training data, until experimentally confirmed, the structures alone are only educated guesses. As such, failure to observe a predicted interaction in silico is not necessarily proof of failure for proteins to interact in vivo. However, optimization of ranking matrices will hopefully increase sensitivity as well as specificity to reduce the chance of false negatives occurring. Limitations of resources A major limitation of this study is the lack of in vitro validation for the exploratory candidates. Because this work is meant to serve as a proof of concept for the hypothesis generation engine, establishing and testing the process relied on existing experimental data to serve as positive controls. It is our hope that both our lab and others will be able to empirically test and cross-reference predictions produced by this pipeline in future experiments. Importantly, this methodology is not intended to replace empirical experimentation, but rather to function as a preliminary screening step that prioritizes candidate genes and interactions for subsequent laboratory validation. Another limitation of the methodology was that all steps require manual input, making it a labor-intensive process. Full automation was not feasible for batch processing because multiple components were utilized through online/web-server versions rather than local APIs or command-line installations. Beyond this, software licensing constraints present challenges for integrating several of these tools into a unified system. Finally, the pilot case study reported here was limited to only a few numbers of ZWINT complexes. Broader extension to diverse families of proteins and experimental conditions will be needed to evaluate the generalizability of this approach. Future updates should also include molecular dynamics, AlphaFold variants with better accuracy, and larger validation sets. Implications for experimental triage This two-pathway workflow offers a practical solution for prioritizing protein-protein interactions that need experimental verification. By combining structure alignment and energetic scoring, it can sift through the large number of potential interactors and categorize them into high-and low-priority candidates. Good candidates like ARC and SNAP25-ZWINT display both structural and energetic compatibility, making them suitable for further investigation through wet-lab experimentation. In contrast, proteins with weak or unclear predictions, such as UBC, should be evaluated in their biological context rather than dismissed outright. Failing to do so could lead to missing an important discovery. This process should also be repeated with both positive and negative controls to better understand the protein's known and baseline behavior. This integration of AlphaFold2, RoseTTAFold2, TM-Align, PyMOL, and HADDOCK into a complete workflow framework establishes a reproducible prioritization pipeline. Our results show that such AI-aided screening can be a hypothesis generator yielding tractable leads without premature discard of biologically relevant yet low-confidence interactions. Conclusion and future directions This case study demonstrates that, while AI tools are continuing to develop, a combination of certain tools can intuitively come up with powerful and promising predictions of PPI. Integration of AlphaFold2, RoseTTAFold2, TM-Align, PyMOL, and HADDOCK into an all-encompassing workflow framework creates a replicable prioritization pipeline. The method is especially useful in less-characterized systems where initial empirical evidence is lacking, but should be considered a hypothesis-generating tool, not an eventual proof of itself. Predicted structural feasibility alone does not ensure biological relevance. Actual interactions depend on changing cellular environments, expression patterns, post-translational modifications, and spatial and temporal control, which current modeling does not fully account for. Modeling of additional proteins identified through IPA, for instance, reflects the type of prospective leads such a pipeline can yield, but each must be specifically confirmed. These PPI predictions will still need to have their merit evaluated through real-world testing and empirical evidence to determine their biological relevance. As these refinements come into play, we hope the methodology presented here will not only accelerate PPI research but also help other scientists to prioritize their most promising exploratory proteins. However, more work is required to ensure that the high ranked predictions match real world PPIs. The broader significance of this work is that it demonstrates a practical, yet scalable way to triage PPI candidates using in silico approaches prior to engaging with labor-intensive wet bench experiments. The use of AI as a frontend filter offers researchers the potential to focus their resources on their most structurally and energetically promising interaction, which has the potential to accelerate discovery, reduce cost, and guide early exploration in systems and experiments where empirical data would be otherwise limited. Data link Our models and data are available through the Data Repository "Figshare": •All data produced from our novel Pipeline Can be found and downloaded here: https://figshare.com/articles/dataset/ZWINT_Interactome_Exploration_Supporting_Data_for_a_Novel_AI-Integrated_PPI_Workflow/30020866 All data produced from our novel Pipeline Can be found and downloaded here: https://figshare.com/articles/dataset/ZWINT_Interactome_Exploration_Supporting_Data_for_a_Novel_AI-Integrated_PPI_Workflow/30020866 Funding Student researchers for this project received a free Academic Plan from Neurosnap collaborators to access pertinent research software required for the study. Prior elements of the work were supported in part by funding from The National Institutes of Health of the United States (NS120617). CRediT authorship contribution statement Nora Abdelfattah: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Bridget Broncales: Data curation, Formal analysis, Investigation, Validation, Visualization, Writing – original draft, Writing – review & editing. Ivana Albert: Data curation, Formal analysis, Software, Writing – review & editing. Jaimee Kohan: Data curation, Formal analysis, Software, Writing – review & editing. Lei Yu: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Writing – original draft, Writing – review & editing. Declaration of competing interest The authors declare no conflict of interest. Acknowledgements We would like to thank Saule Mikaliunaite for feedback on preliminary drafts of this publication; Doria Elhalafawy, Lauren Houvener, Meghana Rajesh, and Anisha Singhal for technical assistance; and Keaun Amani and the staff at Neurosnap for lending their expertise and resources. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.bbrep.2026.102612. Supplementary data The following are the Supplementary data to this article: Multimedia component 1 Multimedia component 2 Data availability Our models and data are available through the Data Repository "Figshare" and are attached to the manuscript. FigsharePipeline Research Data (Original data) FigsharePipeline Research Data (Original data) References Associated Data Supplementary Materials Multimedia component 1 Multimedia component 2 Data Availability Statement Our models and data are available through the Data Repository "Figshare" and are attached to the manuscript. FigsharePipeline Research Data (Original data) FigsharePipeline Research Data (Original data)