Spectral analysis and AI/ML-based data-driven approaches for enhanced protein conformation selection and prediction in drug discovery

Abstract

Artificial intelligence (AI) has become an important tool in drug discovery by enabling the analysis of large-scale molecular dynamics (MD) simulation data and improving the understanding of protein-ligand interactions. However, identifying functionally relevant ligand-binding conformations from highly dynamic MD trajectories remains a major challenge. We propose a spectral analysis-based AI/machine learning (ML) framework to improve the identification of ligand-binding protein conformations. The framework compares the Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT) by transforming protein feature time series into frequency-domain and time-frequency-domain representations, respectively. These spectral features capture global conformational dynamics and localized structural changes. A probabilistic majority-voting decision-fusion strategy integrates predictions from multiple AI/ML models, while the spectral feature space is used to mitigate class imbalance, improve the signalto-noise ratio, and enhance discriminative learning. The framework was evaluated on three G protein-coupled receptors (GPCRs): ADORA2A, OPRD1, and OPRK1. Compared with baseline models trained on raw time-series data, the proposed approach achieved improved predictive performance. FFT-based features effectively captured high-frequency conformational signatures, whereas DWT-based features identified localized high-energy patterns associated with ligand-binding events. The decision-fusion strategy further improved the sensitivity, the area under the receiver operating characteristic curve (AUC), and the overall consistency of the prediction across all targets. Spectral-domain analysis substantially enhances AI-driven identification of ligand-binding protein conformations by providing complementary representations of protein dynamics and improving classification performance. This framework offers a robust approach for analyzing MD simulations in the discovery of structure-based drugs and can facilitate the identification of biologically relevant conformations. Future work will validate the predicted conformations through molecular docking and extend the framework to additional therapeutic protein targets.

1 Introduction

The discovery of new therapeutic compounds depends on the identification of small molecules that selectively and safely bind to their intended protein targets to modulate biological function. Developing such a pharmaceutical drug, defined as a clinically tested and approved therapeutic agent, typically requires more than a decade of research and can cost more than $1-2 billion (Southey and Brunavs, 2023). Despite these investments, more than 90% of candidate molecules fail during preclinical or clinical evaluation (Hughes et al., 2011). A major contributor to this high attrition rate is off-target binding, in which compounds bind unintended proteins, leading to reduced efficacy or adverse effects (Evangelista et al., 2016).

To improve early-stage success rates, modern drug discovery increasingly relies on high-performance computing (HPC) (Liu et al., 2016) to evaluate large chemical libraries. Structure-based approaches such as molecular dynamics simulations (MD) (Karplus and Petsko, 1990) and ensemble-based docking (Amaro et al., 2018) can now screen millions of small-molecule compounds against a protein target in hours. Ensemble docking leverages the protein conformational selection mechanism, in which ligands preferentially bind pre-existing conformational states within a dynamic ensemble (Amaro et al., 2018). These workflows generate massive datasets–often exceeding one million protein conformations and protein-ligand complexes–along with associated interaction energies (Amaro et al., 2018). Only a small fraction of these conformations correspond to true binding-competent states, creating a highly imbalanced classification problem. Machine learning (ML) and artificial intelligence (AI) have emerged as powerful tools to analyze such large-scale biophysical datasets. They enable the identification of structural or physicochemical signatures that distinguish binding from non-binding conformations, thus accelerating the selection of promising protein states for downstream docking or experimental validation. However, the extreme class imbalance (Elrahman and Abraham, 2013), where non-binding conformations - vastly outnumber binding conformations, poses a significant challenge for traditional ML algorithms. Previous work introduced a two-stage sampling-based classifier to mitigate this imbalance and improve the prediction of minority binding conformations (Akondi et al., 2019; 2022; Gupta et al., 2022a). Spectral analysis has recently gained traction in computational biology for uncovering dynamic patterns in protein systems, including protein-protein interactions (Jia et al., 2016), decoy discrimination (Chen et al., 2009), and MD-derived motion analysis (Benson and Daggett, 2012). Building on this direction, our earlier work demonstrated that discrete wavelet transform (DWT) coefficients capture high-energy, time-localized fluctuations associated with ligand binding (Gupta et al., 2022c). In that study, we first applied a feature-ranking methodology (Gupta et al., 2022a; 2022b; 2022c) to identify the most informative protein descriptors across the MD trajectory. These selected descriptors were then transformed into the spectral domain using DWT, allowing us to analyze fluctuations in their time-frequency representations. By examining high-energy spectral coefficients, we developed a targeted re-balancing strategy that simultaneously addressed the severe class-imbalance problem and enhanced the detection of binding-relevant conformations. The motivation for this approach stems from the nature of MD simulations: protein conformations are sampled at uniform time intervals over trajectories that span hundreds of nanoseconds to microseconds. Such time-series data contain subtle but meaningful structural transitions that may be difficult to detect in the raw temporal domain. Spectral representations help reveal these hidden conformational signatures by focusing on periodic, transient, or localized fluctuations that correlate with ligand-binding events. This provides a principled rationale for using spectral analysis to improve protein conformation selection. Building upon our previous work, the present study investigates two prominent spectral analysis techniques, the Discrete Wavelet Transform (DWT) (Mallat, 1989) and the Fast Fourier Transform (FFT) (Harikrishna and Reddy, 2021), to evaluate their effectiveness for dataset rebalancing and improved identification of binding conformations. A comparative analysis of these two spectral approaches is conducted to determine which method provides superior performance in distinguishing binding and non-binding conformations for three targets: ADORA2A, OPRD1, and OPRK1. Furthermore, an ensemble-based decision fusion strategy using probabilistic majority voting is incorporated to enhance the detection and classification of both binding and non-binding conformations, thereby improving the robustness and reliability of the overall prediction framework.

2 Materials and methods

2.1 Dataset overview

Three proteins were used to evaluate the effectiveness of the proposed framework: ADORA2A, OPRD1, and OPRK1. The corresponding Protein Data Bank (PDB) structure IDs used in this study are “3EML”, “4N6H″, and “4DJH”, respectively. For each protein, two categories of conformations were considered: (a) ligand-binding conformations and (b) non-binding conformations. These protein-ligand conformational datasets were previously established and reported by Evangelista et al., 2016. Each data set contains approximately 3,000 protein conformations and 50 physicochemical descriptors. All datasets exhibit significant class imbalance. Specifically, the imbalance ratio for ADORA2A is 3:1, indicating that for every one binding conformation, there are three non-binding conformations. The imbalance ratios for OPRK1 and OPRD1 are 21:1 and 41:1, respectively. The physicochemical descriptors/features extracted using the Molecular Operating Environment (MOE) platform for ADORA2A, OPRD1, and OPRK1 datasets have been previously described and published (Gupta et al., 2022a; 2022b; 2022c).

In this work, the proposed framework was trained on 30% of the data set and evaluated on the remaining 70%. Given the relatively limited size of typical biomedical datasets, this training strategy serves three purposes: first, to assess the model’s ability to achieve strong predictive performance with a reduced training sample size; second, to reduce the risk of overfitting that may arise from using a disproportionately large training set; and third, to maintain consistency with our previously published work. For all three target proteins, the data were randomly partitioned into training and test sets to ensure unbiased evaluation. The distribution of binding and non-binding protein conformations in the training and test sets is displayed in Table 1.

2.2 Feature selection and ranking

The feature selection process involves understanding the contributions of individual biophysical-chemical protein attributes to the prediction rates of potential drug candidates or binding protein conformations. We are specifically interested in delving into multiple facets of protein attributes and evaluating the unique information they contain from an ML and statistical standpoint. In addition, feature selection reduces the number of input variables or protein features, thereby shrinking the learning space of AI/ML models and improving their prediction efficiency while reducing computational and memory overhead. In our work, we have tested four ML-based feature selection methods to identify important protein features, reduce redundancy, and improve the prediction of potential binding protein conformations, thereby reducing the computational complexity of the dataset. The four ML feature selection methods used are Analysis of variance (ANOVA), Mutual Information (MI), Recurrence Quantification Analysis (RQA), and Spearman correlation (SC).

Analysis of variance (ANOVA) (Johnson and Synovec, 2002) was used to determine the linear relationships among the protein’s features, and the significant features associated with the highest F-values were selected. The ANOVA method aids in selecting important physicochemical protein features that play a vital role in protein-ligand interactions and conformational selection. In this method, features are selected based on a user-defined threshold applied to the computed F-value (a feature importance score). Mutual Information (MI) (Mahrooghy et al., 2015; Guyon and Elisseeff, 2003) is a measure of how much knowledge can be gained about a variable ‘P’ using a different random variable ‘Q’. To understand the common information across all protein features, the MI value is first calculated for all protein descriptors to determine how dependent the vectors of physicochemical characteristics are on all attributes. If the MI value is 1, it is assumed that all protein characteristics share the same information, and if it is 0, it is assumed that none of the features share any common (or special) information. In our study, protein features with the highest MI values above a user-defined threshold, as determined by the calculated MI scores, were retained.

A nonlinear data analysis approach for studying dynamical systems is recurrence quantification analysis (RQA) (Eckmann et al., 1987), which first quantifies recurrent patterns in a dynamic system. Entropy, defined as the probability distribution of diagonal line lengths on the RQA plot, was used as the feature-importance metric. In terms of the time-space evolution of protein conformations, it helps investigate the relationship between the RQA-based entropy measure and the likelihood of identifying potential binding protein conformations. Features were selected using a user-defined threshold applied to the entropy-based importance scores.

The Spearman correlation (SC) coefficient (Hauke and Kossowski, 2011) is used to measure the strength and direction of each protein feature’s monotonic relationship with the target variable. Using the absolute value of the correlation coefficient, protein features are selected. This method helps identify physicochemical protein features that are highly correlated with one another. As with other methods, feature selection is performed using a user-defined threshold on absolute correlation values.

Finally, a unified feature ranking score is computed by integrating the feature ranks obtained from all four feature selection methods to identify the most significant protein features. For each method, features that satisfy the user-defined threshold criterion are assigned a binary score of 1 (selected), whereas those that do not are assigned a score of 0 (not selected). The binary scores from all four methods are then aggregated, yielding a cumulative score ranging from 0 to 4 for each feature. Features with a total score of 4, indicating selection by all four methods, are considered the most robust and biologically significant. The rationale for adopting a feature score of 4 is based on two key considerations. First, features consistently selected by all four feature-selection methods are more likely to represent truly informative and biologically relevant characteristics rather than method-specific artifacts. Second, enforcing consensus across all methods minimizes potential biases arising from dataset characteristics or the underlying assumptions of individual feature-selection techniques. Therefore, this consensus-based strategy enhances the reliability, robustness, and generalization of the selected feature set. Table 2 presents the common protein features identified through this unified feature ranking framework. The rationale for using only three descriptors is based on our previous work (Gupta et al., 2022b; 2022c), in which this common subset achieved performance comparable to or better than both target-specific and full-feature sets. Their consistent performance across multiple targets suggests that they capture key characteristics that distinguish binding from non-binding conformations, motivating their use in this study and warranting further investigation of their mechanistic relevance.

2.3 Data rebalancing

The dataset used in our experiments suffers from class imbalance, in which the distribution of non-binding (higher sample class) and binding (lower sample class) protein conformations is severely imbalanced. This issue often causes AI/ML models to learn in a biased manner, favoring the class with the predominant samples (non-binding protein conformations), leading to a higher misclassifications rate in the class with fewer samples (binding protein conformations). To tackle this issue, our previous work (Gupta et al., 2022a; 2022b; 2022c) has proposed several novel solutions. In this work, we address this class imbalance problem through the introduction of an innovative solution to reduce the non-binding protein conformations and increase the binding protein conformation samples by generating synthetic samples using the information gathered from the time-frequency domain using discrete Wavelet transform (DWT) and Fast Fourier transform (FFT), and Generative adversarial networks (GANs). Through this novel work, we intend to explore the impact of time-frequency and frequency-domain analyses of protein conformations to understand their binding tendencies and enhance the prediction of potential binding protein conformations.

2.3.1 Discrete wavelet transform

Wavelet transforms have been used for decades in image compression, signal and image processing applications due to their ability to provide simultaneous time- and frequency-domain resolution. A multi-resolution discrete wavelet transform (DWT) representation is built on differences in information between successive resolutions (Weeks and Bayoumi, 2003). In multi-resolution DWT, the signal is divided into one or more resolution levels, where the high-pass filter produces the signal’s detail coefficient and the low-pass filter its approximation coefficient. Any level’s absolute coefficient determines which set of filters should be used to produce the absolute and detailed signals at the following level (Mallat, 1989). In our study, we are using only 3 of the 50/51 protein features selected by the AI/ML feature ranking method. Therefore, DWT is applied to all three protein features selected in the dataset. We apply the DWT to the signal at levels 1 and 8, corresponding to the lowest level (time resolution) and the highest wavelet level (frequency resolution), respectively, for our desired time-frequency resolutions (Menon and Yarahmadian, 2018; Mahrooghy et al., 2015; Menon et al., 2018). To maximize the prediction of binding vs. non-binding protein conformations, DWT is applied at two separate layers to preserve high-energy coefficients and extract rich time-frequency information in the wavelet domain. The wavelet coefficients are sorted by level first. Finally, as defined by Equation 1, we determine the total energy TE that is present in the wavelet coefficients at the desired level.where represents the absolute wavelet coefficient present at the desired level, Equation 1 is used to calculate the threshold parameter , which determines the percentage of significant wavelet coefficients to be preserved at the lowest level based on the determined total energy from Equation 2. Finally, we flag the indices of the retained wavelet coefficients and add the energy of the sorted wavelet coefficients (highest to lowest significant coefficients) existing at the lowest level up until the threshold is met. This ensures that only biologically relevant conformations consistent with learned spectral binding patterns are kept. As a result, the dataset is not arbitrarily inflated; instead, it is enriched with representative minority-class samples while reducing noise and limiting spurious or biased synthetic generations.

The visualization in Figures 1a–c illustrate the discrete wavelet transform spectral signatures for the features pro_hyd_moment, pro_asa_hyd, and pro_asa_vdw, respectively, for protein OPRK1. In the figure, the orange line represents the wavelet coefficients at the lowest level, while the identified peaks, denoted by green dots, highlight the chosen significant wavelet coefficients. The selected significant coefficients were subsequently used to rebalance the dataset after being mapped back to the time domain, aiming to address class imbalance. The discrete wavelet transform spectral signatures of proteins ADORA2A and OPRD1 are shown in Supplementary Figures S1, S3, respectively.

FIGURE 1

2.3.2 Fast fourier transform

The Fast Fourier transform (FFT) is a commonly used technique for time-frequency-domain signal and image analysis in biomedical, wireless communication, and signal and image processing applications (Harikrishna and Reddy, 2021). The FFT is a method for calculating the discrete Fourier transform (DFT) using a spectrum to determine the frequency component of a sampled signal. In the literature, many researchers have employed this method in the frequency-domain analysis of biological signals, such as electrocardiogram (ECG) (Kalaivani et al., 2022), photoplethysmography (PPG), and blood pressure (BP) (Aisuwarya et al., 2019). In our work, we have employed the FFT to identify high-energy Fourier coefficients in the frequency domain and to explore whether this information can be used to improve the predictive power of our proposed AI/ML framework.

In this experiment, FFT is applied to 3 out of 50/51 protein descriptors selected from the prior feature selection in Table 2. On the obtained spectral information, the amplitude for each spectral point is computed using Equation 3:where is the amplitude of the signal, we take the absolute value of , which denotes the real, and , which stands for the imaginary part of the computed Fourier frequency coefficient value. The amplitude values are sorted in decreasing order, and we determine the total amplitude in the frequency domain using Equation 4 based on the calculated amplitudes.

The threshold parameter is calculated using Equation 5, which specifies the percentage of significant amplitude values to retain relative to the determined total amplitude. Lastly, we flag the retained value indexes and add the amplitude of the sorted values until the threshold is reached.

The representation in Figures 2a–c showcases the spectral signature obtained through the fast Fourier transform for the features pro_hyd_moment, pro_asa_hyd, and pro_asa_vdw, respectively, for protein OPRK1. Here, the coefficients in the frequency domain are depicted by the orange line, and the selected high-energy coefficients are marked by blue dots indicating identified peaks. Following this, the identified significant coefficients were used to rebalance the data set after being mapped back to the time domain, effectively addressing the class imbalance. The fast Fourier transform spectral signatures of the proteins ADORA2A and OPRD1 are shown in Supplementary Figures S2, S4, respectively.

FIGURE 2

2.4 Generative adversarial network

Generative adversarial networks (GANs) are an unsupervised learning technique that generates new samples that closely resemble the original data set by exploiting regularities or patterns in the input data. Two artificial models used in the GAN method, the discriminator and the generator, compete to learn the data (Wonkeun and Dongil, 2022). While the generator produces synthetic data that resembles real data, the discriminator aims to distinguish genuine from synthetic data. To address class imbalance and improve the prediction of potential binding protein conformations, a Generative Adversarial Network (GAN) is employed to oversample the minority class (binding conformations) in the dataset. The GAN-based approach generates synthetic yet realistic minority-class samples, enabling a more balanced and representative learning process for the proposed AI/ML framework. Unlike traditional oversampling techniques, which may duplicate existing samples and increase the risk of overfitting, GANs can learn the underlying data distribution and generate diverse synthetic conformations that better capture the variability of binding states. In this study, GANs are employed to increase the representation of minority class samples (Binding conformations) identified in the spectral domain. In this study, GANs were employed to augment the minority-class distribution by synthesizing additional binding conformations identified within the spectral domain (Gupta et al., 2022c). GAN models were trained on frequency-domain and time-frequency-domain representations derived from FFT and DWT, respectively. The synthesized spectral-domain samples were subsequently mapped back to the temporal domain to ensure that the generated conformations retained biologically meaningful structural characteristics. This targeted augmentation increased the effective representation of binding conformations and enabled the downstream classifiers to learn more discriminative decision boundaries between the two classes. Incorporating these synthetic samples into the training process improved model generalization, enhanced robustness to severe class imbalance, and strengthened the framework’s ability to reliably detect both binding and non-binding protein conformations. Figure 3 provides a complete overview of the GAN-based workflow.

FIGURE 3

2.5 Classification methods

2.5.1 Gaussian Naive Bayes classifier

Gaussian Naive Bayes Classifier (GB) is a popular supervised data classification technique based on Bayes Rule (Berrar, 2019). For probabilistic classification, it uses maximum a posteriori (MAP) estimation. A test data instance’s posterior class probability is determined using the class-conditional density estimation and class prior probability parameters, and the class with the highest posterior probability is chosen (Ham et al., 2017). The supervised classification and prediction performed in this work use the GB approach, which is probabilistic and geared toward identifying probable binding protein conformations.

2.5.2 Extreme Gradient Boosting

Several weak classifiers are combined into a single strong classifier using Extreme Gradient Boosting (XGBoost), a tree-ensemble boosting technique (Chen and Guestrin, 2016). The strong learner is iteratively trained to achieve the best classification or prediction performance from a base learner. In this study, supervised classification and prediction were performed using the XGBoost classifier, which is effective in identifying probable non-binding protein conformations.

2.5.3 Convolutional neural network

The most popular artificial neural network utilized in a range of computer vision applications, including texture identification (Cimpoi et al., 2016), remote sensing scene classification (Hu et al., 2015), and structure-based protein analysis (Torng and Altman, 2017), is the convolutional neural network (CNN), a supervised deep learning technique (Hossain and Sajib, 2019). In a CNN, several convolutional, pooling, and dropout layers are followed by one or more fully connected (FC) layers (Sultana et al., 2018). To reduce overfitting, our proposed approach uses a convolution layer followed by a dropout layer, then a max pooling layer, a fully connected layer, and an output layer. The rectified linear unit (ReLU) serves as the activation function for the convolutional and fully connected layers. Binary cross-entropy serves as the CNN’s loss function. Figure 4 describes the CNN architecture used in our work.

FIGURE 4

2.6 Evaluation metrics

The accuracy of a spectral-based AI/ML framework is calculated by dividing the sum of correctly predicted binding and non-binding protein conformations by the total number of conformations in the data set, as defined by Equation 6.

The sensitivity of the spectral-based AI/ML framework refers to its ability to accurately forecast binding protein conformations. It is calculated by dividing the total number of binding protein conformations in the data set by the number of successfully anticipated binding protein conformations, as defined by Equation 7.

The AUC score measures the effectiveness of the spectral-based AI/ML framework to distinguish between binding and non-binding protein conformations. The higher the AUC score, the better the correct prediction of binding protein conformations (TP) and non-binding protein conformations (TN).

The Matthews correlation coefficient (MCC) (Chicco and Jurman, 2020) is a contingency-matrix-based metric that computes the Pearson correlation between the predicted and actual binary labels, making it robust to class imbalance. Unlike accuracy or F1-score, MCC yields a high value only when the classifier correctly identifies most positive and most negative samples. It ranges from −1 (complete disagreement) to +1 (perfect classification), with 0 representing performance equivalent to random guessing. In terms of the confusion-matrix entries, MCC is defined by Equation 8:

2.7 Decision fusion

In our proposed work, we use three classifiers to maximize the detection of potential binding and non-binding protein conformations and enhance the AI/ML model’s ability to distinguish between them. Initially, all classifiers are assigned equal probabilities, i.e., = = = 1/3, so that each classifier’s prediction is given equal importance. Here, is the probability of selecting GB predictions, is the probability of selecting XGBoost predictions, and is the probability of selecting CNN predictions.

Due to the inherent class imbalance in the dataset, the individual classifiers exhibit varying biases in their predictions. Some models tend to favor detecting non-binding protein conformations, while others are more effective at identifying binding conformations. To address this imbalance and improve the AI/ML framework’s ability to accurately distinguish between binding and non-binding conformations, a rank-based decision-fusion strategy is introduced.

The proposed strategy evaluates each classifier according to two performance measures: specificity and sensitivity. Specificity reflects the model’s ability to correctly identify non-binding conformations, whereas sensitivity measures its ability to detect binding conformations correctly. Based on these metrics, each classifier is assigned two separate ranks ranging from ‘0’ to ‘2’. The first rank is determined using specificity values, and the second rank is assigned using sensitivity values. A rank of 0 corresponds to the lowest-performing classifier for a given metric, while a rank of 2 corresponds to the best-performing classifier. This ranking mechanism prioritizes classifiers that exhibit stronger predictive performance for either binding or non-binding conformations. Rather than relying on a fixed set of classifiers whose behavior may be biased toward the characteristics of the original training distribution, the framework dynamically selects the highest-performing models as new data becomes available. This adaptive selection strategy reduces dependence on any single model, mitigates performance drift, and improves the reliability of distinguishing between the two classes across evolving datasets.

After the ranking process, the prediction probabilities generated by the individual classifiers are combined using a weighted probabilistic fusion scheme, as defined in Equations 9, 10:where and denote the total probability of each predicted binding and non-binding protein conformations. Based on the computed specificity value of the model, , , and reflect the assigned ranks of the GB, XGBoost, and CNN classifiers, respectively. Based on the model’s computed sensitivity values, , , and denote the assigned ranks of the GB, XGBoost, and CNN classifiers, respectively. The final classification produced by the decision-fusion layer is determined by selecting the class with the highest aggregated probability. A prediction is accepted only when the corresponding probability is greater than or equal to 0.6667 (66.67%), ensuring stronger agreement among the ensemble classifiers and improving the reliability of both binding and non-binding conformation detection.

2.8 Enrichment ratio

To validate the proposed AI/ML framework, the model’s true positives (TP) and false negatives (FN) were used to compute the enrichment ratio. The enrichment ratio provides a biologically meaningful measure of how effectively the model prioritizes true binding conformations relative to random or baseline selection. In a biological context, a higher enrichment ratio indicates that the framework is successfully concentrating experimentally or functionally relevant protein-ligand binding states at the top of the ranked predictions. This is particularly important in drug discovery, where experimental validation is costly and time-consuming, as it enables more efficient identification of promising candidate conformations for downstream studies such as molecular docking and binding affinity analysis. Consequently, the enrichment ratio serves as a practical indicator of the model’s ability to enhance biological signal detection in highly imbalanced, noisy conformational datasets.

As discussed and reported in our earlier work (Gupta et al., 2022a; 2022b; 2022c), the baseline enrichment ratio is used to evaluate overall prediction performance in the absence of an AI/ML framework. We selected subset data based on previously computed and published expected protein-ligand interaction energies reported by Evangelista et al. (2016) to ensure the accuracy of the baseline enrichment ratio. Equation 10, which expresses the baseline enrichment, was obtained by dividing the number of binding conformations by the total number of conformations, as described by Evangelista et al. (2016). This determines the baseline enrichment observed throughout the testing phase without the AI/ML framework.

Equation 12 is used to compute the AI/ML framework enrichment ratios by using various subsets of the TP and FN values. Equation 13 specifies the final enrichment ratios returned by each of the four distinct filters (A, B, C, D), based on the values obtained from Equations 11, 12.

FIGURE 5

Table 3 summarizes the maximum Base Enrichment ratios computed for each target protein.

2.9 Proposed spectral based AI/ML protein conformation selection/prediction framework

3 Results

3.1 Computational evaluation on ADORA2A dataset

Table 4 summarizes the classification performance for ADORA2A using four WaveGAN-based models: WaveGAN-CNN, WaveGAN-GB, WaveGAN-XGBoost, and the WaveGAN-Fusion ensemble obtained through the voting mechanism described in Sections 2, 3. All models were trained on 30% of the dataset (630 non-binding and 268 binding conformations) and evaluated on the remaining 70% (1,517 non-binding and 583 binding conformations) using the feature set in Table 2. Table 5 presents the corresponding results for the FFT-based models: FFT-GAN-CNN, FFT-GAN-GB, FFT-GAN-XGBoost, and FFT-GAN-Fusion. Across both tables, it is evident that the decision fusion strategy consistently improves discrimination between binding and non-binding conformations, as reflected in higher sensitivity, overall accuracy, AUC, and MCC. When comparing the two spectral approaches, WaveGAN-Fusion achieves slightly better sensitivity, accuracy, AUC, and MCC than FFT-GAN-Fusion.

Figure 6 illustrates the total accuracy of the AI/ML frameworks evaluated across varying training sizes. WaveGAN-Fusion performs particularly well in smaller training subsets, whereas FFT-GAN-Fusion shows stronger performance as training size increases. Figure 7 further compares the AUC scores of the baseline models with their corresponding fusion models. The results clearly show that WaveGAN-Fusion provides the strongest class-separation capability, outperforming both its individual base learners and the FFT-based fusion models in terms of AUC scores.

FIGURE 6

FIGURE 7

Finally, Table 6 reports the enrichment ratio computed from the true-positive and false-negative predictions of the decision-fusion model. The proposed spectral-based AI/ML framework achieves a maximum enrichment of 13, which is substantially higher than the baseline enrichment of 9 obtained when no AI/ML-based conformation selection and prediction framework is applied. This demonstrates that the fusion-based spectral models not only improve classification performance but also meaningfully enhance the prioritization of biologically relevant binding conformations.

3.2 Computational evaluation on OPRD1 dataset

Table 7 summarizes the classification performance for OPRD1 for the WaveGAN-based models: WaveGAN- CNN, WaveGAN-GB, WaveGAN-XGBoost, and WaveGAN-Fusion ensemble. Table 8 presents the corresponding results for the FFT-based models: FFT-GAN-CNN, FFT-GAN-GB, FFT-GAN-XGBoost, and FFT-GAN-Fusion. All models were trained on 30% of the dataset (883 non-binding and 18 binding conformations) and evaluated on the remaining 70% (2,049 non-binding and 54 binding conformations). This dataset exhibits a severe 41:1 class imbalance, and under these conditions, the FFT-based fusion model demonstrates markedly superior sensitivity, overall accuracy, AUC, and MCC values compared to WaveGAN-Fusion.

Figure 8 illustrates the overall accuracy of the evaluated AI/ML frameworks across varying training sizes. Consistent with the tabulated results, FFT-GAN-Fusion outperforms WaveGAN-Fusion for both smaller and larger training subsets, indicating that FFT-derived spectral features provide a more stable basis for classification under extreme imbalance. Figure 9 also compares the AUC scores of the baseline models with their corresponding fusion models, showing that the fusion strategy substantially improves class discrimination. Notably, FFT-GAN-Fusion achieves the strongest separation between binding and non-binding conformations, outperforming both its individual base learners and the WaveGAN-based fusion model.

FIGURE 8

FIGURE 9

Finally, Table 9 reports the enrichment ratio computed from the true-positive and false-negative predictions of the decision-fusion model. The proposed spectral-based AI/ML framework achieves a maximum enrichment of 40, “20 times” higher than the baseline enrichment of 2.

3.3 Computational evaluation on OPRK1 dataset

Table 10 summarizes the classification performance for OPRK1 using the four WaveGAN-based models: WaveGAN-CNN, WaveGAN-GB, WaveGAN-XGBoost, and WaveGAN-Fusion, whereas Table 11 presents the corresponding results for the FFT-based models: FFT-GAN-CNN, FFT-GAN-GB, FFT-GAN-XGBoost, and FFT-GAN-Fusion. All models were trained on 30% of the dataset (864 non-binding and 37 binding conformations) and evaluated on the remaining 70% (1998 non-binding and 101 binding conformations). This dataset exhibits a substantial 21:1 class imbalance, and under these conditions, WaveGAN-Fusion achieves notably higher sensitivity, AUC, and MCC values, whereas FFT-GAN-Fusion attains the highest overall accuracy.

Figure 10 shows the overall accuracy of the evaluated AI/ML frameworks across varying training sizes. Consistent with the tabulated results, FFT-GAN-Fusion maintains superior accuracy for both smaller and larger training subsets. At the same time, WaveGAN-Fusion continues to provide stronger discrimination between binding and non-binding conformations. Figure 11 further highlights that the fusion models consistently outperform their individual base learners, with WaveGAN-Fusion showing the strongest class-separation capability for OPRK1.

FIGURE 10

FIGURE 11

A notable observation across the individual models is the presence of negative MCC values, indicating that several base learners perform worse than random classification. One potential explanation for the weaker performance of certain WaveGAN- and FFT-GAN-based individual models is the non-uniform time-domain evolution of binding versus non-binding conformations in the training dataset. Because the selection of samples for spectral-based oversampling depends on high-energy coefficients, fewer binding conformations may satisfy the high-energy threshold, resulting in limited oversampling and reduced representation of binding-specific spectral signatures. Despite this challenge, the proposed spectral-analysis-based sampling and classification framework shows strong potential for distinguishing binding from non-binding conformations, and the fusion mechanism effectively compensates for these limitations, as reflected in the improved AUC and MCC values.

Finally, Table 12 reports the computed enrichment ratio where our proposed framework achieves a maximum enrichment of 26, substantially higher than the baseline enrichment of 5.5, demonstrating that, despite the imbalance, our proposed framework can not only improve predictive performance but also significantly enhance the prioritization of biologically meaningful binding conformations.

4 Discussion

In this work, we introduced a spectral-based AI/ML framework that integrates DWT, FFT, and a decision-fusion strategy to address severe class imbalance in protein: ligand binding datasets. Across three GPCR targets-ADORA2A, OPRD1, and OPRK1 - the proposed WaveGAN-Fusion and FFT-GAN-Fusion models consistently outperformed single-classifier baselines in sensitivity, specificity, overall accuracy, AUC, and MCC. WaveGAN-Fusion performed best for ADORA2A and showed superior binding conformation detection for OPRK1. In contrast, FFT-GAN-Fusion achieved the highest overall accuracy and AUC for OPRD1 and remained more stable across varying training sizes. For OPRK1, several individual models produced negative MCC values due to extreme imbalance and non-uniform time-domain evolution of binding versus non-binding conformations. Because spectral oversampling depends on high-energy coefficients, fewer binding conformations met the oversampling threshold. The fusion strategy mitigated these limitations by combining complementary decision boundaries, yielding consistently positive MCC values and an improved AUC. The enrichment ratios further demonstrate the practical value of the framework. For ADORA2A, OPRD1, and OPRK1, the fusion models achieved enrichment values of 13, 40, and 26 - Ovsubstantially higher than baseline enrichment of 9, 2, and 5.5 without AI/ML-based conformation selection and prediction. These results show that the proposed approach not only improves predictive accuracy but also effectively prioritizes biologically meaningful binding conformations.

A significant advantage of the proposed framework is its ability to reduce the computational burden of ensemble docking, one of the most expensive steps in structure-based drug discovery. Traditional ensemble docking requires evaluating thousands of MD-derived conformations, many of which are non-binding and contribute minimally to downstream screening accuracy. Prior studies have attempted to mitigate this bottleneck through clustering-based selection, RMSD filtering, or energy-based pruning—approaches exemplified by (Amaro et al., 2018), who highlighted the high computational cost of ensemble docking and the need for improved conformation-selection strategies. Additional work by (Sinko et al., 2013) showed that RMSD based clustering often produces redundant conformations, while (Sherman et al., 2006) demonstrated that induced-fit docking improves accuracy but remains computationally intensive when applied to large ensembles. These long-standing challenges are now being revisited in light of emerging AI-based conformational-landscape predictors. As Abriata (2026) emphasize, “no large-scale method can achieve this other than slow and computer-intensive MD simulations,” underscoring why ensemble docking remains computationally demanding. Recent Critical Assessment of Structure Prediction (CASP) assessments (Kryshtafovych et al., 2023; 2026) similarly show that “the results on these tracks were not outstanding,” highlighting the persistent difficulty in capturing alternative conformations. Abriata (2026) also identifies Microsoft’s BioEmu (Lewis et al., 2025) as a potential breakthrough, noting that it can emulate what “would take 100,000 GPU hours of atomistic MD in minutes” by generating thousands of Boltzmann-weighted conformations. However, BioEmu also inherits limitations common to AI-based sampling—most notably memorization bias. Together, these observations reinforce the need for frameworks that can intelligently filter or prioritize conformations before expensive docking or refinement steps. By using spectral-domain signatures to identify binding-relevant conformations early, our framework directly addresses the computational inefficiencies highlighted in both MD-based and AI-based ensemble generation. This positions our approach as a practical, scalable solution aligned with the broader push toward efficient conformational-landscape prediction and the reduction of unnecessary computational overhead in drug-discovery pipelines.

The proposed framework can not only alleviate the computational bottleneck associated with ensemble docking but can also be readily integrated into the early stages of the drug-discovery pipeline. A key application is the automated labeling of unlabeled MD conformations, followed by validation of predicted binding states through ensemble docking. Non-binding conformations can be efficiently filtered out, reducing computational overhead, while predicted binding conformations can be selectively docked to confirm their relevance. This establishes a feedback loop in which docking outcomes iteratively refine the spectral-domain classifiers, progressively improving predictive accuracy. Overall, this work demonstrates that spectral analysis, combined with GAN-based sampling and decision fusion, provides a scalable, biologically grounded approach to protein-conformation selection. It advances methodological capabilities for handling extreme class imbalance, enhances the detection of binding-relevant conformations, and opens new avenues for understanding the physicochemical evolution of proteins while streamlining early-stage drug-discovery workflows.

References

• 1

  AbriataL. A. (2026). The latest AI breakthroughs in structural biology: protein binder design and conformational state prediction. Commun. Biol.9, 627. 10.1038/s42003-026-10112-3

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  AisuwaryaR.HendrickH.MeiotizaM. (2019). “Analysis of cardiac frequency on photoplethysmograph synthesis for detecting heart rate using fast fourier transform,” in 2019 International Conference on Electrical Engineering and Computer Science (ICECOS), 391–395.

  • Google Scholar
  • View reference in article

• 3

  AkondiV. S.MenonV.BaudryJ.WhittleJ. (2019). “Novel K-means clustering-based undersampling and feature selection for drug discovery applications,” in 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2771–2778.

  • Google Scholar
  • View reference in article

• 4

  AkondiV. S.MenonV.BaudryJ.WhittleJ. (2022). Novel big data-driven machine learning models for drug discovery application. Molecules27 (3), 594. 10.3390/molecules27030594

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  AmaroR. E.BaudryJ.ChoderaJ.DemirÖ.McCammonJ. A.MiaoY.et al (2018). Ensemble docking in drug discovery. Biophysical J.114 (10), 2271–2278. 10.1016/j.bpj.2018.02.038

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  BensonN. C.DaggettV. (2012). Wavelet analysis of protein motion. Int. J. Wavelets Multiresolution Inf. Process.10 (4), 1250040. 10.1142/s0219691312500403

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  BerrarD. P. (2019). “Bayes’ theorem and naive bayes classifier,” in Encyclopedia of Bioinformatics and Computational Biology.

  • Google Scholar
  • View reference in article

• 8

  ChenT.GuestrinC. (2016). XGBoost: A Scalable Tree Boosting System. arXiv:1603.02754.

  • Google Scholar
  • View reference in article

• 9

  ChenM.LiuB.YanW.ShenB. (2009). “Wavelet transform based protein decoy discrimination,” in 2009 3rd International Conference on Bioinformatics and Biomedical Engineering, 1–4.

  • Google Scholar
  • View reference in article

• 10

  ChiccoD.JurmanG. (2020). The advantages of the matthews correlation coefficient over F1 score and accuracy in binary classification evaluation. BMC Genomics21 (1), 6. 10.1186/s12864-019-6413-7

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  CimpoiM.MajiS.KokkinosI.VedaldiA. (2016). Deep filter banks for texture recognition, description, and segmentation. Int. J. Comput. Vis.118 (1), 65–94. 10.1007/s11263-015-0872-3

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  EckmannJ. P.KamphorstS. O.RuelleD. (1987). Recurrence plots of dynamical systems. Europhys. Lett.4 (9), 973–977. 10.1209/0295-5075/4/9/004

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  ElrahmanS. M. A.AbrahamA. (2013). A review of class imbalance problems. J. Netw. Innovative Comput.1, 332–340.

  • Google Scholar
  • View reference in article

• 14

  EvangelistaW.WeirR. L.EllingsonS. R.HarrisJ. B.KapoorK.SmithJ. C.et al (2016). Ensemble-based docking: from hit discovery to metabolism and toxicity predictions. Bioorg. and Med. Chem.24 (20), 4928–4935. 10.1016/j.bmc.2016.07.064

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  GuptaS.BaudryJ.MenonV. (2022a). Using big data analytics to “back engineer” protein conformational selection mechanisms. Molecules27 (8), 2509. 10.3390/molecules27082509

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  GuptaS.BaudryJ.MenonV. (2022b). Big data analytics for improved prediction of ligand binding and conformational selection. Front. Mol. Biosci.9, 953984. 10.3389/fmolb.2022.953984

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  GuptaS.MenonV.BaudryJ. (2022c). “Wavelet-based spectral analysis for protein conformation selection and prediction using AI in drug discovery applications,” in 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2595–2602.

  • Google Scholar
  • View reference in article

• 18

  GuyonI.ElisseeffA. (2003). An introduction to variable and feature selection. J. Mach. Learn. Res.3, 1157–1182.

  • Google Scholar
  • View reference in article

• 19

  HamJ.ChenY.CrawfordM. M.GhoshJ. (2017). “A non-parametric mixture of gaussian naive bayes classifiers based on local independent features,” in 2017 Artificial Intelligence and Signal Processing Conference (AISP), 209–212.

  • Google Scholar
  • View reference in article

• 20

  HarikrishnaE.ReddyK. A. (2021). “Use of transforms in biomedical signal processing and analysis,” in Real Perspective of Fourier Transforms and Current Developments in Superconductivity.

  • Google Scholar
  • View reference in article

• 21

  HaukeJ.KossowskiT. (2011). Comparison of values of Pearson’s and Spearman’s correlation coefficients on the same sets of data. Quaest. Geogr.30 (2), 87–93. 10.2478/v10117-011-0021-1

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  HossainM. A.SajibM. (2019). Classification of image using convolutional neural network. Glob. J. Comput. Sci. Technol.19, 13–18.

  • Google Scholar
  • View reference in article

• 23

  HuF.XiaG.-S.HuJ.ZhangL. (2015). Transferring deep convolutional neural networks for scene classification of high-resolution remote sensing imagery. Remote Sens.7 (11), 14680–14707. 10.3390/rs71114680

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  HughesJ. P.ReesS.KalindjianS. B.PhilpottK. L. (2011). Principles of early drug discovery. Br. J. Pharmacol.162 (6), 1239–1249. 10.1111/j.1476-5381.2010.01127.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  JiaJ.XiaoX.LiuB. (2016). Prediction of protein–protein interactions with physicochemical descriptors and wavelet transform via random forests. J. Laboratory Automation21 (3), 368–377. 10.1177/2211068215581487

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  JohnsonK. J.SynovecE. R. (2002). Pattern recognition of jet fuels: comprehensive GC × GC with ANOVA-based feature selection and principal component analysis. Chemom. Intelligent Laboratory Syst.60 (2), 225–237. 10.1016/s0169-7439(01)00198-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  KalaivaniK.MaheswariN. U.VenkateshR. (2022). Heart disease diagnosis using optimized features of hybridized ALCSOGA algorithm and LSTM classifier. Netw. Comput. Neural Syst.33 (1–2), 95–123. 10.1080/0954898x.2022.2061062

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  KarplusM.PetskoG. (1990). Molecular dynamics simulations in biology. Nature347, 631–639. 10.1038/347631a0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  KryshtafovychA.SchwedeT.TopfM.FidelisK.MoultJ. (2023). Critical assessment of methods of protein structure prediction (CASP)—Round XV. Proteins Struct. Funct. Bioinform.91, 1539–1549. 10.1002/prot.26617

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  KryshtafovychA.SchwedeT.TopfM.FidelisK.MoultJ. (2026). Progress and bottlenecks for deep learning in computational structure biology: CASP round XVI. Proteins Struct. Funct. Bioinform.94, 5–14. 10.1002/prot.70076

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  LewisS.HempelT.Jiménez-LunaJ.GasteggerM.XieY.FoongA. Y. K.et al (2025). Scalable emulation of protein equilibrium ensembles with generative deep learning. Sci. (New York, N.Y.)389 (6761), eadv9817. 10.1126/science.adv9817

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  LiuT.LuD.ZhangH.ZhengM.YangH.XuY.et al (2016). Applying high-performance computing in drug discovery and molecular simulation. Natl. Sci. Rev.3 (1), 49–63. 10.1093/nsr/nww003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  MahrooghyM.YarahmadianS.MenonV.RezaniaV.TuszynskiJ. A. (2015). The use of compressive sensing and peak detection in the reconstruction of microtubules length time series in the process of dynamic instability. Comput. Biol. Med.65, 25–33. 10.1016/j.compbiomed.2015.07.014

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  MallatS. (1989). A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Analysis Mach. Intell.11 (7), 674–693. 10.1109/34.192463

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  MenonV.YarahmadianS. (2018). “Wavelet-based compressed sensing sampling and estimation of N-states random evolution model parameters in microtubule signal,” in IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2009–2016.

  • Google Scholar
  • View reference in article

• 36

  MenonV.YarahmadianS.RezaniaV. (2018). Novel EM-based ML kalman estimation framework for superresolution of stochastic three-states microtubule signal. BMC Syst. Biol.12 (6), 112–131. 10.1186/s12918-018-0631-5

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  ShermanW.DayT.JacobsonM. P.FriesnerR. A.FaridR. (2006). Novel procedure for modeling ligand/receptor induced fit effects. J. Med. Chem.49 (2), 534–553. 10.1021/jm050540c

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  SinkoW.LindertS.McCammonJ. A. (2013). Accounting for receptor flexibility and enhanced sampling methods in computer-aided drug design. Chem. and Biol. Drug Des.81 (1), 41–49. 10.1111/cbdd.12051

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  SoutheyM. W. Y.BrunavsM. (2023). Introduction to small molecule drug discovery and preclinical development. Front. Drug Discov.3, 1314077. 10.3389/fddsv.2023.1314077

  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  SultanaF.SufianA.DuttaP. (2018). “Advancements in image classification using convolutional neural networks,” in 2018 Fourth International Conference on Research in Computational Intelligence and Communication Networks (ICRCICN), 122–129.

  • Google Scholar

• 41

  TorngW.AltmanR. B. (2017). 3D deep convolutional neural networks for amino acid environment similarity analysis. BMC Bioinforma.18, 302. 10.1186/s12859-017-1702-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 42

  WeeksM.BayoumiM. (2003). Discrete wavelet transform: architectures, design and performance issues. VLSI Signal Process.35, 155–178. 10.1023/a:1023648531542

  • CrossRef
  • Google Scholar

• 43

  WonkeunJ.DongilK. (2022). OBGAN: Minority oversampling near borderline with generative adversarial networks. Expert Sys. Appl.197, 116694.

  • Google Scholar