In recent years, the study of protein interactions within cells has gained considerable attention due to their essential roles in cellular processes. Proteins rarely function in isolation; instead, they engage in diverse interactions, from transient contacts to stable assemblies. These interactions can form intricate and dynamic networks that influence critical biological processes [1,2]. Among these, direct protein-protein interactions and modular relationships within protein complexes have been a primary focus for understanding their impact on the stability of interaction networks. Studies employing topological and network-based approaches [[3], [4], [5], [6]] have highlighted the importance of interactions directly between proteins and within modules for maintaining structural integrity and influencing the functionality of complexes.

A notable example of such a network is the INO80 chromatin remodeling complex, a conserved protein complex with significant roles in nuclear processes such as transcription, DNA replication, and repair [[7], [8], [9]]. In Saccharomyces cerevisiae, INO80 is particularly crucial for maintaining genome stability [10]. Extensive structural characterization has provided insight into its organization, revealing distinct modules, including the core ATPase Ino80, the Rvb1/Rvb2 heterododecamer, and the Arp5-Ies6 module [[11], [12], [13], [14]]. Additionally, topological data analysis has refined our understanding of interaction strengths within these modules, identifying key structural relationships that influence INO80's stability and function [15,16].

Another key approach in recent research for studying protein complex dynamics is network perturbation. Perturbation of protein interaction networks can be achieved through biochemical or genetic approaches, including inhibitors, genetic mutations, and genetic deletion [15,[17], [18], [19], [20]]. In S. cerevisiae, genetic deletion has been proven to be particularly effective in studying protein complex dynamics [[21], [22], [23], [24], [25]]. However, while this technique has significantly contributed to the characterization of network structures, there remains a gap in understanding how structural changes within protein complexes affect their functionality. Given the critical roles of protein interactions in cellular processes, this knowledge is essential for understanding the impact of these networks on organismal health.

To bridge this gap, integrative machine learning techniques offer promising solutions by enabling predictive modeling of network perturbation outcomes. Analyzing protein networks involves examining interactions among numerous proteins, resulting in complex and high-dimensional data. Techniques such as feature selection [26] and dimensionality reduction [27] play a crucial role in managing this complexity by pinpointing the most pertinent features, including specific proteins or interactions, within the network. Furthermore, rare events, such as mutations that lead to disease, adverse drug reactions, or unusual phenotypic responses, are challenging to predict due to their infrequency and the high dimensionality of biological data [28,29]. Machine learning models, particularly those designed for anomaly detection or rare event prediction, can be trained to recognize the subtle signals associated with these events [30].

Ensemble learning techniques can be employed to enhance the performance of these models. These methods can integrate both supervised and unsupervised machine learning approaches, leveraging their strengths to improve prediction accuracy and model interpretability [[31], [32], [33]]. Supervised learning techniques are particularly effective for predicting rare events, such as specific protein interactions, by learning from labeled data [30,34]. In contrast, unsupervised learning methods excel in discovering hidden patterns and structures within unlabeled data, making them valuable for exploratory analysis [34,35]. Here, we will use both techniques to analyze the dynamics of protein complex networks. In their recent work, Dutta et al. [36] propose an integrative machine learning approach employing twelve supervised machine learning methods to simultaneously allow for the prediction of rare events and feature structure selection. This approach not only enhances the accuracy of predicting rare events but also improves model interpretability through the incorporation of feature structures [36]. Thus, it is well-suited for the complexity and high dimensionality of omics data and the dynamic nature of biological networks, particularly when dealing with perturbed networks.

In this study, we leverage perturbation network analysis in conjunction with an integrative machine learning approach to predict the outcomes of perturbing the subunits of the INO80 complex in S. cerevisiae. By combining advanced statistical methodologies with machine learning techniques, we comprehensively analyze how genetic perturbations influence protein networks. First, we apply the statistical framework QPROT for differential expression analysis to quantify changes in protein abundance between wild-type and perturbed networks. We also explore the functional pathways associated with proteins exhibiting significant changes. We then incorporate structural data to provide mechanistic explanations for the observed genetic perturbations between specific subunit interactions within the INO80 complex. Finally, we utilize a machine learning approach that integrates supervised learning techniques with feature selection to predict network perturbations within the complex while considering its biological structure. Our findings significantly enhance the understanding of protein network dynamics and underscore the transformative potential of integrating network perturbation analysis with structural data and machine learning techniques in areas such as cross-species analysis and disease research.