A Multi-Modal AI/ML-based Framework for Protein Conformation Selection and Prediction in Drug Discovery Applications

The development of pharmaceutical drugs is a time-intensive and costly process, with more than 90% of drug candidates failing during preclinical or clinical testing. A major challenge lies in accurately predicting protein-ligand interactions, especially given that traditional computational methods often rely on a single protein conformation, failing to capture biologically relevant structural variability. To address this, we present an AI/ML-based multi-modal framework based on Graph Convolutional Network (GCN) that integrates both global and local protein descriptors to classify binding and non-binding conformations more effectively. Global descriptors capture overarching physico-chemical and structural properties of proteins, while local descriptors—such as pharmacophores—provide site-specific information crucial for modeling ligand interactions. Our GCN based approach demonstrates that integrating local and global structural perspectives significantly improves predictive accuracy and robustness. By enabling more reliable protein conformation classification, this work contributes toward scalable, AI-driven drug discovery—an increasingly critical goal in response to global health challenges. February 15, 2026 1/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint

1 Drug development is a critical research area for chemical scientists and pharmaceutical 2 companies; however, it continues to face significant challenges, including low success 3 rates, high costs, and lengthy development timelines [1]. Traditional drug discovery 4 approaches frequently fail due to inaccurate target selection, inadequate safety profiles, 5 limited therapeutic efficacy, and difficulties in identifying appropriate patient 6 populations [2]. A major contributor to unsuccessful target selection is off-target 7 binding, in which drug candidates unintentionally interact with proteins other than 8 their intended molecular targets. These off-target interactions can diminish drug 9 selectivity and potency and often lead to adverse side effects or toxicity, ultimately 10 contributing to late-stage clinical failures [3]. 11 In the rapidly evolving landscape of pharmaceutical research, computational 12

have become a cornerstone of drug discovery and development. 13 Computer-aided drug design (CADD) offers a cost-effective and time-efficient 14 complement to experimental approaches by enabling the prediction of drug behavior, 15 protein–ligand interactions, and pharmacokinetic properties prior to synthesis and 16 experimental validation. Techniques such as molecular modeling, structure–activity 17 relationship analysis, and virtual screening allow researchers to explore vast chemical 18 spaces and make informed decisions early in the drug development pipeline [4] - [7]. 19 Traditional computational drug discovery, however, often relies on docking ligands to 20 a single, static protein conformation, thereby neglecting biologically relevant receptor 21 flexibility and potentially overlooking effective drug candidates. To overcome this 22 limitation, ensemble-based docking strategies have been developed [8], which 23 incorporate multiple protein conformations generated through molecular dynamics (MD) 24 simulations. This approach better captures conformational variability and enables the 25 identification of hidden or transient binding sites that may not be observable in 26 single-structure docking. In ensemble-based workflows, multiple target protein 27 conformations are screened against pre-existing libraries of active compounds and 28 decoys to identify the strongest and most relevant protein–ligand interactions. Although 29 this approach more effectively captures biologically relevant and high-affinity 30 interactions, it is computationally expensive, particularly when multiple target proteins 31 February 15, 2026 2/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint are considered. Moreover, only a small fraction of the sampled conformations—often 32 just a few hundred out of millions—exhibit statistically significant binding activity that 33 is essential for conformational selection, thereby substantially increasing the 34 computational challenge of identifying these rare but critical states and highlights the 35 need for advanced Big Data analytics, particularly artificial intelligence (AI) to 36 effectively process and interpret large-scale datasets. 37 The increasing availability of large-scale biochemical and structural data has driven 38 the widespread adoption of artificial intelligence (AI) and machine learning (ML) in 39 modern drug discovery, with the aim of improving efficiency, accuracy, and 40 cost-effectiveness across the development pipeline. AI/ML techniques have shown 41 significant promise in enhancing target specificity, mitigating toxicity risks, and 42 accelerating the progression of drug candidates through early discovery and pre-clinical 43 stages [9,10]. Ensemble-based docking approaches, which evaluate millions of 44 protein–ligand conformations generated through molecular dynamics simulations, 45 produce extremely large and highly imbalanced datasets. Since only a small fraction of 46 these conformations exhibit meaningful binding activity, efficiently identifying the most 47 relevant candidates remains both computationally intensive and analytically complex. 48 While high-performance computing (HPC) resources are essential to manage the scale of 49 these simulations, intelligent data-driven strategies are equally critical for prioritizing 50 biologically significant conformations. To reduce the substantial computational burden 51 typically associated with ensemble docking, this work introduces performance-oriented 52 AI/ML frameworks that integrate seamlessly with HPC environments and ensemble 53 docking outputs to support protein conformation selection and classification. By 54 directly addressing the challenges of extreme class imbalance and limited experimental 55 validation, these frameworks enable the efficient identification of rare, high-value 56 binding conformations. As a result, they significantly improve the accuracy, robustness, 57 and overall reliability of computational drug discovery pipelines.58 In this work, we introduce a multi-modal framework that integrates global and local 59 descriptors to improve the classification of protein conformations. Global descriptors 60 capture overall structural and physicochemical properties—such as mass, 61 hydrophobicity, and radius of gyration, while local descriptors, including 62 pharmacophores, provide site-specific information critical for identifying ligand-binding 63 February 15, 2026 3/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint interactions. These complementary feature sets offer a more comprehensive view, 64 enhancing classification performance. Like global descriptor datasets, pharmacophore 65 data suffer from class imbalance, with only a limited number of conformations showing 66 strong binding activity. This imbalance can lead to biased models and poor67 generalization. To address this, we apply previously established data-driven68 augmentation techniques [11]- [13] to balance the dataset and improve model reliability. 69 To capture both spatial and statistical dependencies, we construct graph-based 70 representations for each descriptor type: a proximity matrix for pharmacophores (local 71 descriptors) and a Pearson correlation matrix for global descriptors. These graphs are 72 processed using a Graph Convolutional Network (GCN) trained with a contrastive loss 73 function, which improves feature discrimination by pulling together similar 74 conformations and separating dissimilar ones in the embedding space. The resulting 75 graph-level embeddings are then classified using traditional ML algorithms, and a 76 decision fusion strategy aggregates their outputs for improved performance. This 77 integrated approach combines structural insight with advanced graph learning to 78 provide a scalable, interpretable, and effective solution for protein conformation 79 classification, with potential to accelerate AI-driven drug discovery. 80

81 0.1 Dataset Overview 82 Four proteins were selected to evaluate the effectiveness of our proposed model: 83 ADORA2A (Adenosine A2A Receptor), ADRB2 (Beta-2 Adrenergic Receptor), OPRD1 84 (Delta Opioid Receptor), and OPRK1 (Kappa Opioid Receptor). These targets were 85 previously studied and reported in our earlier work [11]– [13]. Each protein includes 86 experimentally validated conformations that either (a) bind to ligands (binding 87 conformations) or (b) do not bind to ligands (non-binding conformations), as previously 88 described in [8]. 89 To improve the classification of binding versus non-binding protein conformations,90 we employed both global and local structural descriptors, including pharmacophoric 91 features specific to ligand-binding sites. Global descriptors capture overall structural 92 February 15, 2026 4/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint and physicochemical properties of a protein conformation and were described in our 93 prior work [11]– [13]. In contrast, local descriptors were derived from pharmacophore 94 features using the DB-PH4 module in MOE (Molecular Operating Environment). These 95 features were extracted within a 6.5 ˚A radius of the ligand-binding site using the 96 ”unified scheme,” which includes hydrogen bond donors (Don), acceptors (Acc), cations 97 (Cat), anions (Ani), aromatic centers (Aro), and hydrophobic centers (Hyd). The 98 MD-derived conformational dataset containing these local descriptors has also been 99 documented and published in [14]. 100 Table 1 summarizes the number of binding and non-binding protein conformations, 101 the total number of conformations, the number of local and global descriptors, and the 102 class imbalance ratio for each target protein used in this study.103 T able 1.Dataset description T arget protein # of binding protein conformations # of non-binding protein conformations T otal # of protein conformations # of local descriptors # of global descriptors Class-im balance ratio ADORA2A (adenosine receptor A2A) 850 2,150 3,000 282 50 3:1 ADRB2 ( β2-adrenergic receptor) 154 2,411 2,565 542 51 16:1 OPRK1 ( κ-type opioid receptor) 137 2,862 2,999 697 50 21:1 OPRD1 ( δ-type opioid receptor) 72 2,932 3,004 329 51 41:1 0.2 Pearson Correlation Matrix 104 Our study’s global features—such as protein mass, volume, radius of gyration, 105 hydrophobic surface area, and mobility—are generic to all protein conformations and 106 lack explicit spatial dependency information for GCN input. To address this, we use the 107 Pearson correlation matrix to capture pairwise linear relationships between features, 108 revealing how changes in one relate to another. It measures the linear relationship 109 between two global descriptors. The correlation coefficient rab, ranging from -1 to 1, 110 quantifies the strength and direction of this relationship, showing how changes in one 111 feature predict changes in another. It is calculated as follows [15]: 112 rab = P i(ai − ¯a)(bi − ¯b)qP i(ai −¯a)2(bi − ¯b)2 (1) where ai and bi are individual feature values and ¯a, ¯b their respective means. A value of 113 +1 indicates a perfect positive linear relationship, -1 a perfect negative linear 114 relationship, and 0 no linear correlation. Values closer to ±1 indicate stronger linear 115 February 15, 2026 5/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint relationships, while those near zero indicate weak or no correlation. 116 0.3 Gaussian Naive Bayes Classifier 117 The Gaussian Naive Bayes (GB) classifier is a probabilistic supervised learning method 118 based on Bayes’ theorem, assuming feature independence and Gaussian-distributed 119 continuous variables [16]. For each feature, GB estimates the mean (µ b) and variance 120 (σ2 b) per class b to compute the likelihood using: 121 P (a|b) = 1p 2πσ 2 b exp  −(a−µ b)2 2σ2 b  (2) Classification relies on the maximum a posteriori (MAP) estimate derived from 122 Bayes’ theorem: 123 P (b|a) = P (a|b)P (b) P(a) (3) WhereP(b) is the prior, P (a|b) is the likelihood, and P(b|a) is the posterior 124 probability of class b given data point a. The model predicts the class with the highest 125 posterior probability: 126 class = arg max b (P(a|b)P(b)) (4) This lightweight and interpretable model is especially effective when the features are 127 conditionally independent and normally distributed [17]. 128 0.4 K-Nearest Neighbor 129 K-Nearest Neighbors (KNN) is a straightforward, non-parametric supervised learning 130 algorithm used for classification and regression. It predicts the label of a new data point 131 xby identifying its K closest neighbors using the Euclidean distance metric, 132 D(x, yi) = p (x−y i)2, and assigning x to the class with the majority among these 133 neighbors. Despite its simplicity, KNN faces limitations such as high memory usage and 134 sensitivity to imbalanced data [18], [19], [20]. 135 February 15, 2026 6/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 0.5 Random Forest 136 Random Forest (RF) is a robust ensemble learning method used for classification and 137 regression. It constructs multiple decision trees during training, and for classification 138 tasks, the final class label is determined by majority voting across these trees, 139 improving accuracy [21]. 140 The training process begins by creating bootstrap samples—random subsets of the 141 training data selected with replacement—ensuring each tree is trained on a unique 142 dataset. At each node, a random subset of features is considered, and the best feature 143 from this subset is selected for splitting. This randomness introduces diversity among 144 trees, reducing overfitting and enhancing generalization [22]. 145 Trees are grown fully without pruning to capture complex data patterns. The best 146 splits are chosen by minimizing the Gini impurity G(M), defined as: 147 G(M) = 1 − cX i=1 (p2 i ) (5) where G(M) is the Gini impurity of Node M, c is the total number of classes and pi 148 is the probability of selecting class i. 149 Each tree independently predicts the class of an input sample. For binary 150 classification, the final prediction is the class receiving the majority vote among all trees. 151 For example, if more trees predict Class 1 (binding) than Class 0 (non-binding), the 152 instance is assigned to Class 1. This ensemble approach reduces overfitting, improves 153 accuracy, and increases robustness against noise. 154 0.6 Support Vector Machine 155 Support Vector Machines (SVMs) are widely used for binary classification by finding 156 the optimal hyperplane that separates two classes in an n-dimensional feature space [23]. 157 The hyperplane maximizes the margin, defined as the distance between the hyperplane 158 and the closest points (support vectors) from each class, ensuring reliable class 159 separation. Given a labeled training set {(x1, y1), (x2, y2), . . . ,(xn, yn)}, where xn is a 160 feature vector and yn its class label, the optimal hyperplane satisfies [24]: 161 wxT + m = 0 (6) February 15, 2026 7/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint with w as the weight vector, x the feature vector, and m the bias. The classification 162 constraints are: 163 wxT n +m >0,if yn = 1 (7) wxT n + m < 0, if yn = 0 (8) Training adjusts w and m to maximize the margin 1 /∥w∥2, enhancing generalization. 164 For non-linearly separable data, SVM employs kernel functions to implicitly map 165 inputs to higher-dimensional spaces. The linear kernel computes the inner product: 166 K(xi, xj) = xi · xj (9) and the decision function is: 167 q(x) = w · x + m = nX j=1 βjyj(xj · x) + m (10) where βj are Lagrange multipliers [25]. The linear kernel works well when data is 168 linearly separable. 169 The Radial Basis Function (RBF) kernel extends SVM’s flexibility by using a 170 Gaussian function: 171 K(xi, xj) = exp(−α∥xi − xj∥2) (11) q(x) = nX j=1 βjyj exp(−α∥xj − x∥2) + m (12) where α controls the influence radius of each training point [26]. Low α (gamma) 172 yields smoother decision boundaries with risk of underfitting, while high α creates 173 tighter boundaries that may overfit noise. 174 February 15, 2026 8/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 0.7 Graph Convolutional Neural Network175 Graphs are widely used in domains such as social analysis, bioinformatics, and computer 176 vision to capture structural relationships between data, providing richer insights than 177 analyzing individual data points. A Graph Convolutional Network (GCN) is a deep 178 learning model designed to process graph-structured data, which is inherently 179 non-Euclidean—examples include protein structures, molecules, and social networks. 180 GCNs learn node embeddings by iteratively aggregating information from neighboring 181 nodes, thereby capturing both node features and the overall graph structure [27]. 182 Formally, a graph G consists of nodes N and edges E representing connections 183 between node pairs [28]. In our study, nodes correspond to local or global descriptors, 184 while edges represent spatial distances between local descriptors or linear dependencies 185 expressed as correlation values between global descriptors. The two main components of 186 a GCN are the adjacency matrix A, which encodes graph topology, and the node feature 187 matrix X, which records node attributes. For our datasets, X contains encoded binary 188 vectors for local descriptors and numerical feature values for global descriptors. The 189 GCN aggregates information from neighboring nodes to make predictions. The 190 operation of a GCN layer is described as follows [29]: 191 • The first step is to add self-loops and normalize the adjacency matrix. Graphs 192 may or may not contain self-loops. Self-loops are included to ensure that each 193 node takes into account its own feature. Equation (13) shows the creation of 194 matrix ¯A by adding self-loops to the adjacency matrix A, where E represents the 195 identity matrix used to incorporate self-connections. 196 ¯A = A + E (13) 197 ¯Dii = X j Aij (14) Here, ¯D denotes the degree matrix, which can be computed as shown in Equation 198 (14). The normalized adjacency matrixU is then calculated using Equation (15). 199 U = ¯D− 1 2 ¯A ¯D− 1 2 (15) February 15, 2026 9/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint •The next step is node feature aggregation, where each node combines its own200 feature vector with those of its neighbors to obtain local context. This aggregation 201 process leverages the normalized adjacency matrix U and is expressed in Equation 202 (16): 203 ¯H (k) =U ¯H (k−1) (16) Here, ¯H (k−1) represents node features from the previous layer. This operation 204 successfully propagates and blends information across neighboring nodes, 205 guaranteeing that each node representation is enhanced by its local graph 206 structure. 207 • The following step is feature transformation, which involves the GCN applying a 208 learnable linear transformation on the aggregated node features. This technique is 209 similar to using a Dense (or Linear) layer in a classical neural network. It allows210 the model to modify, combine, and enhance the aggregated data to better fit the 211 task at hand. Equation (17) illustrates the computation of the transformed 212 feature matrix: 213 H (k) = ¯H (k)W (k) (17) Here, W (k) is the learnable weight matrix that determines how the aggregated 214 features are transformed. 215 • Lastly, a non-linearity, such as ReLU, is applied element-by-element to the 216 transformed features, as indicated in Equation (18). This activation function 217 introduces nonlinearity to the model, allowing it to capture complicated patterns 218 and relationships in the same way that hidden layers in traditional neural 219 networks do. 220 H (k) = ReLU( ¯H (k)W (k)) (18) The GCN model in this study consists of two graph convolutional layers followed by a 221 mean pooling layer, which aggregates node features by averaging them to produce a 222 fixed-size graph-level embedding that captures the overall structure and feature 223 distribution. 224 To enhance the embeddings’ discriminative power, the GCN is trained using a 225 February 15, 2026 10/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint contrastive loss function that pulls embeddings of similar graphs closer while pushing 226 dissimilar ones apart. The loss is defined as follows: 227 Loss = 1 2 yD2 + 1 2(1−y ) (max(0, m− D))2 (19) where D = ∥ei − ej∥2 is the Euclidean distance between embeddings ei and ej, and 228 y indicates pair similarity (y = 1 for similar pairs—both binding or non-binding—and 229 y = 0 for dissimilar pairs). The margin m sets the minimum distance for negative pairs. 230 The first term pulls positive pairs together, while the second pushes negative pairs apart 231 if closer than m. 232 0.8 Decision Fusion 233 In this study, we provide a decision fusion strategy that improves the detection of 234 binding and non-binding protein conformations while also enhancing the AI/ML 235 model’s capacity to distinguish between them. We propose a method that leverages the 236 predictive power of four machine learning techniques which operate on embeddings 237 produced by a unique dual Graph Convolution Network (GCN) model. This dual GCN 238 model consists of two separately trained GCN on a dataset that contains local spatial 239 and global linear relationship respectively, allowing it to capture a more complete 240 description of the data. 241 Let us define the prediction results obtained from the four machine learning 242 models—Gaussian Na¨ ıve Bayes (GB), K-Nearest Neighbor (KNN), Random Forest 243 (RF), and Support Vector Machine (SVM)—applied to GCN embeddings of local and 244 global descriptors as follows: LGB, LKN N, LRF and LSV M represent the predictions 245 obtained using embeddings of GCN model trained on local descriptors and GGB, 246 GKN N, GRF and GSV M represent the predictions obtained using embeddings of GCN 247 model trained on global descriptors. After obtaining the individual predictions from 248 each dataset, the cumulative prediction score is computed by summing all the 249 February 15, 2026 11/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint predictions, as defined by the following equation (19). 250 T otalP rediction = LGB + LKN N + LRF + LSV M + GGB + GKN N + GRF + GSV M (20) where T otalP rediction is used to compute the final classification results. Conformations 251 with a cumulative score of at least 3 are allocated to Class ’1’, indicating binding 252 protein conformation, while those with a cumulative score of no more than 5 are 253 assigned to Class ’0’, indicating non-binding protein conformation. 254 0.9 Evaluation metrics 255 The confusion matrix and its derived metrics—such as accuracy, sensitivity, and 256 specificity—are widely used to evaluate the performance of machine learning classifiers. 257 For binary classification of binding vs. non-binding protein conformations, the confusion 258 matrix includes four outcomes [11]: 259 •True Positive (TP): Correctly predicted binding conformations (Class 1).260 •False Positive (FP): Non-binding conformations incorrectly predicted as binding. 261 • False Negative (FN): Binding conformations incorrectly predicted as non-binding. 262 •True Negative (TN): Correctly predicted non-binding conformations (Class 0). 263 The accuracy measures the overall proportion of correct predictions: 264 Accuracy = T P + T N T P+T N+F P+F N (21) Sensitivity quantifies the model’s ability to correctly identify binding conformations: 265 Sensitivity = T P T P+F N (22) Specificity quantifies the model’s ability to correctly identify non-binding266 conformations: 267 Specificity = T N T N+F P (23) February 15, 2026 12/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 0.10 Enrichment Ratio 268 To validate the proposed AI/ML framework, true positive (TP) and false negative (FN) 269 predictions were used to compute the enrichment ratio. As outlined in our previous 270 work [11], a baseline enrichment ratio—representing the expected performance under 271 random selection—was calculated for comparison. This metric provides a benchmark to 272 assess how effectively the AI/ML framework improves prediction performance relative 273 to random selection of protein conformations. 274 Table 2 summarizes the maximum Base Enrichment ratios computed for each target 275 protein. 276 Protein Maximum Enrichment Ratio ADORA2A ∼9 ADRB2 ∼5.4 OPRD1 ∼2 OPRK1 ∼5.5 T able 2.Maximum Base Enrichment ratios calculated for each target proteins. 0.11 Proposed Work 277 This study introduces a dual-GCN framework that integrates both local and global 278 features for protein conformation prediction. One GCN is trained on 279 pharmacophore-based local descriptors to capture spatial interactions at binding sites, 280 while the other learns from global descriptors to model broader structural patterns. 281 Each GCN is optimized using a contrastive loss to enhance the separation between 282 binding and non-binding conformations in the embedding space. The resulting 283 embeddings are then fed into four traditional machine learning classifiers, whose 284 outputs are combined through decision fusion to yield the final prediction. An 285 enrichment ratio framework is applied to validate binding conformations in test proteins. 286 This approach leverages the representational strength of GCNs and the predictive power 287 of classical models to improve protein conformation selection. An overview of the 288

is shown in Fig 1. 289 • The summary of Framework 1, which handles the global descriptor data, is as 290 follows: 291 – It begins with the original dataset, which is then refined using the feature 292 February 15, 2026 13/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint scoring method described earlier. 293 –The ML-based feature selection and scoring framework from our prior 294 work [12], [13], is used to identify the most significant protein characteristics. 295 Specifically, four methods are used: Analysis of Variance (ANOVA) [30], 296 Mutual Information (MI) [31], Recurrence Quantification Analysis 297 (RQA) [32], and Spearman Correlation [33]. Each method independently 298 ranks the features based on its criteria. These rankings are then integrated 299 using a majority voting strategy, where features that receive the highest 300 possible vote count of 4—indicating consensus across all methods—are 301 selected. The resulting subset of features is used to construct a new, 302 optimized dataset for further analysis. This framework, along with the 303 selected features for each target protein—ADORA2A, ADRB2, OPRD1, and 304 OPRK1—has been previously described and published in [12].305 –To address the high class imbalance observed in the proteins ADRB2, 306 OPRK1, and OPRD1, a multi-step data re-balancing approach was 307 implemented as detailed in [12]. First, the Extreme Gradient Boosting 308 (XGBoost) [34] classifier was applied to the feature-reduced dataset to 309 distinguish between non-binding (TN) and binding (TP) protein 310 conformations. Next, K-Means [35]- [37] clustering was performed on the 311 non-binding conformations to retain representative samples while eliminating 312 redundant data, thereby reducing dataset bias and size. Finally, Generative313 Adversarial Networks (GANs) were employed to augment the minority class 314 by generating additional binding conformation samples, achieving balanced 315 class representation without increasing the overall dataset size. For the 316 ADORA2A protein, this re-balancing step was omitted due to its low class 317 imbalance, which did not necessitate adjustment for the GCN model. 318 – In parallel with the re-balancing procedure, a Pearson correlation matrix was 319 constructed to capture pairwise linear relationships among the global 320 features. This analysis provides insight into how changes in one feature relate 321 to changes in another, enhancing understanding of feature interactions. 322 Following this, a graph dataset was prepared for input into the GCN model. 323 February 15, 2026 14/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint Here, global features correspond to nodes, the Pearson correlation 324 coefficients serve as edge weights, and the numerical values of the features 325 are assigned as node attributes. 326 – The GCN model was then trained on this graph dataset using 10-fold 327 cross-validation to ensure reliable and generalizable performance. Early 328 stopping was incorporated to prevent overfitting by monitoring the validation 329 loss and halting training when no further improvement was observed. The 330 model with the best validation results was saved and used to generate the 331 final graph-level embeddings. 332 – These graph-level embeddings, derived from the optimized GCN model, were 333 subsequently fed into four traditional machine learning classifiers — 334 Gaussian Na¨ ıve Bayes (GB), K-Nearest Neighbor (KNN), Random Forest 335 (RF), and Support Vector Machine (SVM) — to assess and compare their 336 classification performances. 337 • The summary of Framework 2, which processes local descriptor (pharmacophore) 338 data, is as follows: 339 – The framework begins by applying the multi-step re-balancing technique 340 described in [12]. Unlike the global descriptor dataset, the original local 341 descriptor data is excluded from the feature scoring process due to its high 342 sparsity. Sparse data poses challenges for GCN training because limited and 343 uneven feature availability hampers the aggregation of meaningful patterns 344 across nodes. This issue is exacerbated by conformations containing only one 345 or two pharmacophores in close proximity. Removing such features based on 346 relevance risks losing critical local information necessary for capturing 347 conformation-specific interactions. Therefore, to preserve essential local 348 information, the original local descriptor dataset bypasses the feature scoring 349 framework. 350 –To address class imbalance, the previously described multi-step data 351 re-balancing method is applied. Initially, the feature-reduced dataset is 352 classified with XGBoost to distinguish binding (TP) from non-binding (TN) 353 conformations. K-Means clustering is then used on the TN samples to retain 354 February 15, 2026 15/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint representative instances and reduce redundancy. Finally, GANs augment the 355 minority class by generating additional binding conformations, balancing the 356 classes without increasing overall dataset size. 357 – Next, a graph dataset is prepared for input to the GCN model, where local 358 features serve as nodes, and the distances between them define edge weights. 359 Each node is further characterized by attributes including an encoded binary 360 vector, radius, and frequency corresponding to the respective local feature. 361 – The graph dataset is processed using the GCN model with 10-fold 362 cross-validation and early stopping to ensure stable, reliable performance and 363 prevent overfitting. The model exhibiting the best validation performance is 364 selected to generate the final graph-level embeddings, which are then input365 into four machine learning classifiers for evaluation and classification 366 performance recording. 367 • GEFusion (Graph Embedding Fusion), a decision-level fusion strategy, is applied 368 to combine the classification results from Framework 1 and Framework 2. This 369 integration enables the unique identification of the total number of probable 370 binding and non-binding protein conformations. 371 • Finally, enrichment ratios are calculated using True Positives (TP)—correctly 372 predicted binding conformations—and False Negatives (FN)—incorrectly 373 predicted non-binding conformations—based on the outcomes of the GEFusion 374 decision strategy. 375 Fig 1. Bold the figure title. Figure caption text here, please use this space for the figure panel descriptions instead of using subfigure commands. A: Lorem ipsum dolor sit amet. B: Consectetur adipiscing elit.

and Discussion 376 0.12 Computational Evaluation on ADORA2A Dataset 377 Table 3 presents the classification performance of the ADORA2A protein across varying 378 training sizes, evaluated using the proposed GEFusion methodology. Notably, the model 379 February 15, 2026 16/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint achieves its highest performance—measured by accuracy, sensitivity, and 380 specificity—when trained on 40% of the dataset. The results for this training size, 381 including the performance of individual classifiers applied to both local and global 382 embeddings, are summarized in Table 4 383 T able 3. Classification performance (in %) across different training sizes for protein ADORA2A T raining Size Accuracy(%) Sensitivity(%) Specificity(%) 0.1 90.93 88.09 92.05 0.2 91.88 84.66 94.72 0.3 88.52 72.90 94.53 0.4 94.94 93.24 95.61 0.5 89.00 70.05 96.22 0.6 87.83 68.96 95.14 0.7 92.00 85.48 94.48 0.8 89.33 68.45 97.45 T able 4.Classification performance for 40% training size for protein ADORA2A Approac h TN FP FN TP Accuracy (%) Sensitivit y (%) Sp ecificity (%) KNN (local embeddings) 303 994 69 434 40.94 86.28 23.36 GB (local embeddings) 687 610 130 373 58.89 74.16 52.97 RF (local embeddings) 0 1297 0 503 27.94 100.00 0.00 SVM (local embeddings) 191 1106 22 481 37.33 95.63 14.73 KNN (global embeddings) 1130 167 432 71 66.72 14.12 87.12 GB (global embeddings) 850 447 278 225 59.72 44.73 65.54 RF (global embeddings) 1297 0 503 0 72.06 0.00 100.00 SVM (global embeddings) 1297 0 503 0 72.06 0.00 100.00 GEF usion 1240 57 34 469 94.94 93.24 95.61 From Table 4, it is evident that while the overall framework delivers strong 384 classification performance, some individual models exhibited extreme predictive 385 behavior. For instance, the Random Forest (RF) model trained on local embeddings 386 classified all conformations as binding, whereas both RF and Support Vector Machine 387 (SVM) models trained on global embeddings predicted all conformations as non-binding. 388 These individual biases are effectively mitigated by the framework’s voting-based 389 ensemble mechanism, which combines predictions from multiple models. This ensemble 390 strategy reduces the impact of the outlier behavior of any single model, resulting in 391 more balanced and reliable classification results. 392 As shown in Table 5, our proposed methodology demonstrated strong performance 393 not only in overall model accuracy but also in achieving a high enrichment ratio. T able 5. Enrichment ratios of protein ADORA2A on a training size of 40% Approac h Maxima Filter % of Data Minima Filter % of Data GEF usion 13.67 Filter A 0.5% 13.07 Filter A 1.0% 394 February 15, 2026 17/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 0.13 Computational Evaluation on ADRB2 Dataset395 Table 6 presents the classification performance of the proposed approach across various 396 training sizes for the ADRB2 protein. Among these, the 40% training size yields the 397 best overall performance, achieving higher accuracy, sensitivity, and specificity. Detailed 398

for this optimal training size—including the performance of the proposed399 framework and individual classifiers utilizing local and global embeddings—are provided 400 in Table 7. 401 T able 6.Classification performance (in %) across different training sizes for protein ADRB2 T raining Size Accuracy(%) Sensitivity(%) Specificity(%) 0.1 84.11 64.08 85.42 0.2 88.35 65.63 89.86 0.3 88.47 64.29 90.08 0.4 92.14 80.61 92.92 0.5 93.06 42.68 96.50 0.6 90.55 56.06 92.92 0.7 86.49 66.00 87.92 0.8 93.37 8.82 99.37 T able 7.Classification performance for 40% training size for protein ADRB2 Approac h TN FP FN TP Accuracy (%) Sensitivit y (%) Sp ecificity (%) KNN (local embeddings) 634 807 49 49 44.38 50.00 44.00 GB (local embeddings) 355 1086 22 76 28.01 77.55 24.64 RF (local embeddings) 348 1093 22 76 27.55 77.55 24.15 SVM (local embeddings) 260 1181 17 81 22.16 82.65 18.04 KNN (global embeddings) 978 463 60 38 66.02 38.78 67.87 GB (global embeddings) 1000 441 62 36 67.32 36.73 69.40 RF (global embeddings) 1376 65 87 11 92.06 11.22 95.49 SVM (global embeddings) 1372 69 92 6 89.54 6.12 95.21 GEF usion 1339 102 19 79 92.14 80.61 92.92 Table 8 provides an overview of the derived enrichment ratios based on the decision 402 model’s prediction outcomes (TP and FN), serving as a validation measure for the 403 predictions generated by our proposed framework. 404 T able 8. Enrichment ratios of protein ADRB2 on a training size of 40% Approac h Maxima Filter % of Data Minima Filter % of Data GEF usion 33.71 Filter A 0.5% 28.10 Filter C 1.0% 0.14 Computational Evaluation on OPRD1 Dataset405 Table 9 presents the classification performance of the protein OPRD1 across various406 training sizes using our AI/ML-based multi-modal framework for protein conformation 407 February 15, 2026 18/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint selection and prediction. The best overall performance—measured by accuracy,408 sensitivity, and specificity—is achieved with a training size of 10%. Table 10 further 409 details the classification results of the proposed framework and the individual 410 performance of classifiers utilizing local and global embeddings at this optimal training 411 size. 412 T able 9.Classification performance (in %) across different training sizes for protein OPRD1 T raining Size Accuracy(%) Sensitivity(%) Specificity(%) 0.1 82.64 90.91 82.44 0.2 83.18 85.96 83.11 0.3 93.15 58.18 94.09 0.4 93.00 67.35 93.72 0.5 87.74 88.64 87.71 0.6 93.42 86.21 93.60 0.7 96.34 39.13 97.84 0.8 90.50 75.00 90.82 T able 10.Classification performance for 10% training size for protein OPRD1 Approac h TN FP FN TP Accuracy (%) Sensitivit y (%) Sp ecificity (%) KNN (local embeddings) 1749 887 20 46 66.43 69.70 66.35 GB (local embeddings) 1783 853 26 40 67.47 60.61 67.64 RF (local embeddings) 1598 1038 12 54 61.14 81.82 60.62 SVM (local embeddings) 2087 549 33 33 78.46 50.00 79.17 KNN (global embeddings) 1308 1328 18 48 50.19 72.73 49.62 GB (global embeddings) 1225 1411 18 38 47.30 67.86 46.47 RF (global embeddings) 1558 1078 23 43 59.25 65.15 59.10 SVM (global embeddings) 1025 1611 11 55 39.97 83.33 38.88 GEF usion 2173 463 6 60 82.64 90.91 82.44 T able 11.Enrichment ratios of protein OPRD1 on a training size of 10% Approac h Maxima Filter % of Data Minima Filter % of Data GEF usion 38.33 Filter A 1.0% 32.38 Filter B 0.5% Table 11 presents the resulting enrichment ratios derived from the predictions made 413 by the GEFusion decision strategy, further supporting the validity and effectiveness of 414 the proposed framework. 415 0.15 Computational Evaluation on OPRK1 Dataset 416 Table 12 summarizes the classification performance across various training sizes for the 417 protein OPRK1 using the proposed AI/ML-based multimodal framework for protein 418 conformation selection and prediction. The highest overall performance—reflected by 419 improved accuracy, sensitivity, and specificity—was achieved with a training size of 30%. 420 February 15, 2026 19/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint Table 13 presents the classification results at this training size, including the 421 performance of individual classifiers applied to both local and global embeddings. 422 T able 12. Classification performance (in %) across different training sizes for protein OPRK1 T raining Size Accuracy(%) Sensitivity(%) Specificity(%) 0.1 83.33 64.00 84.27 0.2 86.04 62.07 87.25 0.3 73.08 81.19 72.67 0.4 77.88 79.31 77.80 0.5 84.66 70.27 85.40 0.6 92.17 53.33 94.21 0.7 78.67 73.47 78.97 0.8 84.17 72.73 84.60 T able 13.Classification performance for 30% training size for protein OPRK1 Approac h TN FP FN TP Accuracy (%) Sensitivit y (%) Sp ecificity (%) KNN (local embeddings) 1005 993 52 49 50.21 48.51 50.30 GB (local embeddings) 1295 703 61 40 63.60 39.60 64.81 RF (local embeddings) 0 1998 0 101 4.81 100.00 0.00 SVM (local embeddings) 745 1253 29 72 38.92 71.29 37.29 KNN (global embeddings) 1066 932 46 55 53.41 54.46 53.35 GB (global embeddings) 1317 681 52 58 60.58 48.51 65.92 RF (global embeddings) 1273 725 54 47 62.89 46.53 63.71 SVM (global embeddings) 987 1011 38 63 50.02 62.38 49.40 GEF usion 1452 546 19 82 73.08 81.19 72.67 As shown in Table 14, our proposed methodology demonstrated strong performance 423 not only in overall model accuracy but also in achieving a high enrichment ratio. 424 T able 14. Enrichment ratios of protein OPRK1 on a training size of 30% Approac h Maxima Filter % of Data Minima Filter % of Data GEF usion 32.91 Filter D 5.0% 29.74 Filter B 1.0% From Tables 3, 6, 9, and 12, it is observed that the model consistently performs well 425 across all training sizes for protein ADORA2A. However, for proteins ADRB2, OPRK1, 426 and OPRD1, better overall performance is achieved with smaller training sizes. This 427 trend likely stems from higher class imbalance ratios in these datasets. At larger 428 training sizes, the increased proportion of synthetic binding conformation data, coupled 429 with inherent data sparsity, challenges the GCN model’s ability to distinguish binding430 from non-binding classes, thus affecting the framework’s predictive accuracy. 431 The enrichment ratios shown in Tables 5, 8, 11, and 14 demonstrate that binding 432 conformations can be identified with significantly higher accuracy compared to random 433 sampling, as evidenced by the baseline results in Table 2. In particular, the highest 434 enrichment is achieved when selecting the top 0.5% to 1% of the binding conformations, 435 February 15, 2026 20/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint indicating that the most predictive and biologically relevant data points are 436 concentrated among the candidates ranked highest. 437

438 This work presented an AI/ML-based multi-modal framework designed to improve the 439 detection of both binding and non-binding protein conformations by integrating local 440 and global structural information. The proposed approach leverages Graph 441 Convolutional Networks (GCNs) trained with a contrastive loss function to capture 442 meaningful connectivity patterns within protein structures. The GCN learns 443 discriminative graph embeddings that effectively group binding conformations while 444 separating non-binding ones. These embeddings are then processed by traditional 445 machine learning classifiers, and their outputs are combined using a decision fusion 446 strategy to enhance overall classification performance. 447 The study demonstrated that incorporating both global descriptors—reflecting 448 protein-level structural and stability features—and local descriptors—such as 449 pharmacophores that encode site-specific binding information—provides complementary 450 insights that significantly improve conformation classification. Additionally, the use of 451 graph embeddings learned via contrastive learning offers a more compact and 452 noise-resistant data representation than conventional graph classification methods. 453 These embeddings preserve essential structural relationships while facilitating a clearer 454 distinction between classes. 455 Overall, the integration of multi-modal data, GCN-based embedding learning, and 456 ensemble decision-making contributes to a robust and generalizable framework capable 457 of accurately identifying protein binding states. This methodology holds promise for 458 advancing structure-based protein function prediction and accelerating drug discovery 459 workflows. 460 Acknowledgments 461 The authors gratefully acknowledge Dr. Armin Ahmadi for providing the 462 Pharmacophore dataset used in this study. 463 February 15, 2026 21/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint

1. Nass, S. J. et al. Accelerating anticancer drug development — opportunities and trade-offs. Nat. Rev. Clin. Oncol. 15, 777–786 (2018). 2. Oca˜ na, A., Garc´ ıa-Alonso, S., Amir, E. & Pandiella, A. Refining early antitumoral drug development. Trends Pharmacol. Sci. 39, 922–925 (2018). 3. W. Evangelista, R. L. Weir, S. R. Ellingson, J. B. Harris, K. Kapoor, J. C. Smith, and J. Baudry. ”Ensemble-based docking: From hit discovery to metabolism and toxicity predictions.” Bioorganic & Medicinal Chemistry, vol. 24, no. 20, October 2016, pp. 4928–4935. 4. Anwar T., Kumar P., Khan A.U. Molecular Docking for Computer-Aided Drug Design: Fundamentals, Techniques, Resources, and Applications. Elsevier; Amsterdam, The Netherlands: 2021. Modern Tools and Techniques in Computer-Aided Drug Design; pp. 1–30. 5. Bajorath J. Deep Machine Learning for Computer-Aided Drug Design. Front. Drug Discov. 2022;2:829043. doi: 10.3389/fddsv.2022.829043. 6. Ye F., Lin M., Jin J., Broussy S. Editorial: Computer-Aided Drug Design: Drug Discovery, Computational Modelling, and Artificial Intelligence. Front. Chem. 2022;10:968687. doi: 10.3389/fchem.2022.968687 7. Oli B. Revolutionizing Drug Discovery: A Comprehensive Review of Computer-Aided Drug Design Approaches. Int. J. Res. Appl. Sci. Eng. Technol. 2024;12:308–317. doi: 10.22214/ijraset.2024.63563. 8. Amaro, R.E.; Baudry, J.; Chodera, J.; Demir, ¨O.; McCammon, J.A.; Miao, Y.; Smith, J.C. Ensemble Docking in Drug Discovery. Biophys. J. 2018, 114, 2271–2278. 9. M. Parvathaneni, A. K. Awol, M. Kumari, K. Lan, and M. Lingam. ”Application of artificial intelligence and machine learning in drug discovery and development.” Journal of Drug Delivery and Therapeutics, vol. 13, no. 1, January 2023, pp. 151–158. February 15, 2026 22/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 10. R. Awasthi, S. Mishra, R. Grasfield, J. Maslinski, D. Mahapatra, J. B. Cywinski, A. K. Khanna, K. Maheshwari, C. Dave, A. Khare, F. A. Papay, and P. Mathur. ”Artificial intelligence in healthcare: 2023 year in review.” medRxiv (Cold Spring Harbor Laboratory), February 2024. 11. Gupta S, Baudry J, Menon V. Using Big Data Analytics to “Back Engineer” Protein Conformational Selection Mechanisms. Molecules. 2022; 27(8):2509. https://doi.org/10.3390/molecules27082509 12. Gupta, S., Baudry, J., & Menon, V. (2022). Big data analytics for improved prediction of ligand binding and conformational selection. Retrieved from https://www.frontiersin.org/articles/10.3389/fmolb.2022.953984/full 13. S. Gupta, V. Menon and J. Baudry, ”Wavelet-based Spectral Analysis For Protein Conformation Selection and Prediction Using AI in Drug Discovery Applications,” 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Las Vegas, NV, USA, 2022, pp. 2595-2602, doi: 10.1109/BIBM55620.2022.9995169. 14. A. Ahmadi, S. Gupta, V. Menon, and J. Baudry. Linking machine learning and biophysical structural features in drug discovery. Frontiers in Molecular Biosciences, 11 (2025), 1305272. doi:10.3389/fmolb.2024.1305272. 15. P. Schober, C. Boer, and L. A. Schwarte. Correlation Coefficients: Appropriate Use and Interpretation. Anesthesia & Analgesia, 126(5) (2018), 1763–1768. doi:10.1213/ANE.0000000000002864. 16. A. H. Jahromi and M. Taheri. A non-parametric mixture of gaussian naive bayes classifiers based on local independent features. Artificial Intelligence and Signal Processing Conference (AISP), IEEE (2017), 209–212. 17. J. Ren, S. D. Lee, X. Chen, B. Kao, R. Cheng, and D. Cheung. Naive bayes classification of uncertain data. Ninth IEEE International Conference on Data Mining, IEEE (2009), 944–949. 18. L. Wang, L. Khan, and B. Thuraisingham. An effective evidence theory based k-nearest neighbor (KNN) classification. IEEE/WIC/ACM International February 15, 2026 23/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint Conference on Web Intelligence and Intelligent Agent Technology, 1 (2008), 797–801. 19. S. Chen. K-nearest neighbor algorithm optimization in text categorization. IOP Conference Series: Earth and Environmental Science, 108 (2018), 052074. 20. S. Taneja, C. Gupta, K. Goyal, and D. Gureja. An enhanced k-nearest neighbor algorithm using information gain and clustering. International Conference on Advanced Computing & Communication Technologies, IEEE (2014), 325–329. 21. G. Louppe. Understanding Random Forests: From Theory to Practice. arXiv preprint, 2014. 22. R. Couronn´ e, P. Probst, and A. L. Boulesteix. Random forest versus logistic regression: a large-scale benchmark experiment. BMC Bioinformatics, 19, 270 (2018). https://doi.org/10.1186/s12859-018-2264-5 23. W. S. Noble. What is a support vector machine? Nature Biotechnology, 24(12) (2006), 1565–1577. 24. S. Huang, N. Cai, P. P. Pacheco, S. Narrandes, Y. Wang, and W. Xu. Applications of support vector machine (SVM) learning in cancer genomics. Cancer Genomics & Proteomics, 15(1) (2018), 41–51. 25. A. F. Rochim, K. Widyaningrum, and D. Eridani. Performance Comparison of Support Vector Machine Kernel Functions in Classifying COVID-19 Sentiment. 2021 4th International Seminar on Research of Information Technology and Intelligent Systems (ISRITI), Yogyakarta, Indonesia, 2021, pp. 224–228. doi:10.1109/ISRITI54043.2021.9702845. 26. B. Scholkopf et al. Comparing support vector machines with Gaussian kernels to radial basis function classifiers. IEEE Transactions on Signal Processing, 45(11) (1997), 2758–2765. doi:10.1109/78.650102. 27. S. Zhang, H. Tong, J. Xu, et al. ”Graph convolutional networks: A comprehensive review.” Computational Social Networks, vol. 6, article 11, 2019. https://doi.org/10.1186/s40649-019-0069-y February 15, 2026 24/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint 28. U. A. Bhatti, H. Tang, G. Wu, S. Marjan, and A. Hussain. Deep Learning with Graph Convolutional Networks: An Overview and Latest Applications in Computational Intelligence. International Journal of Intelligent Systems, 2023, Article ID 8342104, 28 pages. https://doi.org/10.1155/2023/8342104 29. F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, and K. Weinberger. Simplifying Graph Convolutional Networks. Proceedings of the 36th International Conference on Machine Learning, in Proceedings of Machine Learning Research 97 (2019), 6861–6871. 30. K. J. Johnson and R. E. Synovec. Pattern recognition of jet fuels: comprehensive GC×GC with ANOVA-based feature selection and principal component analysis. Chemometrics and Intelligent Laboratory Systems, 60(1–2) (2002), 225–237. 31. J. Song, Z. Zhu, P. M. D. Scully, and C. Price. Modified mutual information-based feature selection for intrusion detection systems in decision tree learning. Journal of Computers, 9(7) (2014), 1542–1546. 32. J.-P. Eckmann, S. O. Kamphorst, and D. Ruelle. Recurrence plots of dynamical systems. Europhysics Letters, 4(9) (1987), 973–977. 33. J. Hauke and T. Kossowski. Comparison of values of Pearson’s and Spearman’s correlation coefficients on the same sets of data. Quaestiones Geographicae, 30(2) (2011), 87–93. 34. T. Chen and C. Guestrin. Xgboost: A scalable tree boosting system. arXiv:1603.02754 (2016). 35. O. J. Oyelade, O. O. Oladipupo, and I. C. Obagbuwa. Application of k means clustering algorithm for prediction of students academic performance. arXiv preprint arXiv:1002.2425 (2010). 36. V. S. Akondi, V. Menon, J. Baudry, and J. Whittle. ”Novel big data-driven machine learning models for drug discovery application.” Molecules, vol. 27, no. 3, January 2022, article 594. 37. V. S. Akondi, V. Menon, J. Baudry, and J. Whittle. ”Novel K-means clustering-based undersampling and feature selection for drug discovery February 15, 2026 25/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint applications.” In Proceedings of the 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), San Diego, CA, USA, 18–21 November 2019, pp. 2771–2778. February 15, 2026 26/26 .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint .CC-BY 4.0 International licenseavailable under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint