{"paper_id":"0bd321c6-7b9b-4b9c-a30e-76e3fdfe9ba6","body_text":"A Multi-Modal AI/ML-based Framework for Protein\nConformation Selection and Prediction in Drug\nDiscovery Applications\nShivangi Gupta1*, Vineetha Menon 1*, Jerome Baudry 2*\n1Department of Computer Science, University of Alabama in Huntsville, Huntsville,\nUSA\n2Department of Biological Sciences, University of Alabama in Huntsville, Huntsville,\nUSA\n* sg0097@uah.edu(SG); vineetha.menon@uah.edu(VM); jerome.baudry@uah.edu(JM)\nAbstract\nThe development of pharmaceutical drugs is a time-intensive and costly process, with\nmore than 90% of drug candidates failing during preclinical or clinical testing. A major\nchallenge lies in accurately predicting protein-ligand interactions, especially given that\ntraditional computational methods often rely on a single protein conformation, failing to\ncapture biologically relevant structural variability. To address this, we present an\nAI/ML-based multi-modal framework based on Graph Convolutional Network (GCN)\nthat integrates both global and local protein descriptors to classify binding and\nnon-binding conformations more effectively. Global descriptors capture overarching\nphysico-chemical and structural properties of proteins, while local descriptors—such as\npharmacophores—provide site-specific information crucial for modeling ligand\ninteractions. Our GCN based approach demonstrates that integrating local and global\nstructural perspectives significantly improves predictive accuracy and robustness. By\nenabling more reliable protein conformation classification, this work contributes toward\nscalable, AI-driven drug discovery—an increasingly critical goal in response to global\nhealth challenges.\nFebruary 15, 2026 1/26\n.CC-BY 4.0 International licenseavailable under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nIntroduction 1\nDrug development is a critical research area for chemical scientists and pharmaceutical 2\ncompanies; however, it continues to face significant challenges, including low success 3\nrates, high costs, and lengthy development timelines [1]. Traditional drug discovery 4\napproaches frequently fail due to inaccurate target selection, inadequate safety profiles, 5\nlimited therapeutic efficacy, and difficulties in identifying appropriate patient 6\npopulations [2]. A major contributor to unsuccessful target selection is off-target 7\nbinding, in which drug candidates unintentionally interact with proteins other than 8\ntheir intended molecular targets. These off-target interactions can diminish drug 9\nselectivity and potency and often lead to adverse side effects or toxicity, ultimately 10\ncontributing to late-stage clinical failures [3]. 11\nIn the rapidly evolving landscape of pharmaceutical research, computational 12\nmethods have become a cornerstone of drug discovery and development. 13\nComputer-aided drug design (CADD) offers a cost-effective and time-efficient 14\ncomplement to experimental approaches by enabling the prediction of drug behavior, 15\nprotein–ligand interactions, and pharmacokinetic properties prior to synthesis and 16\nexperimental validation. Techniques such as molecular modeling, structure–activity 17\nrelationship analysis, and virtual screening allow researchers to explore vast chemical 18\nspaces and make informed decisions early in the drug development pipeline [4] - [7]. 19\nTraditional computational drug discovery, however, often relies on docking ligands to 20\na single, static protein conformation, thereby neglecting biologically relevant receptor 21\nflexibility and potentially overlooking effective drug candidates. To overcome this 22\nlimitation, ensemble-based docking strategies have been developed [8], which 23\nincorporate multiple protein conformations generated through molecular dynamics (MD) 24\nsimulations. This approach better captures conformational variability and enables the 25\nidentification of hidden or transient binding sites that may not be observable in 26\nsingle-structure docking. In ensemble-based workflows, multiple target protein 27\nconformations are screened against pre-existing libraries of active compounds and 28\ndecoys to identify the strongest and most relevant protein–ligand interactions. Although 29\nthis approach more effectively captures biologically relevant and high-affinity 30\ninteractions, it is computationally expensive, particularly when multiple target proteins 31\nFebruary 15, 2026 2/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nare considered. Moreover, only a small fraction of the sampled conformations—often 32\njust a few hundred out of millions—exhibit statistically significant binding activity that 33\nis essential for conformational selection, thereby substantially increasing the 34\ncomputational challenge of identifying these rare but critical states and highlights the 35\nneed for advanced Big Data analytics, particularly artificial intelligence (AI) to 36\neffectively process and interpret large-scale datasets. 37\nThe increasing availability of large-scale biochemical and structural data has driven 38\nthe widespread adoption of artificial intelligence (AI) and machine learning (ML) in 39\nmodern drug discovery, with the aim of improving efficiency, accuracy, and 40\ncost-effectiveness across the development pipeline. AI/ML techniques have shown 41\nsignificant promise in enhancing target specificity, mitigating toxicity risks, and 42\naccelerating the progression of drug candidates through early discovery and pre-clinical 43\nstages [9,10]. Ensemble-based docking approaches, which evaluate millions of 44\nprotein–ligand conformations generated through molecular dynamics simulations, 45\nproduce extremely large and highly imbalanced datasets. Since only a small fraction of 46\nthese conformations exhibit meaningful binding activity, efficiently identifying the most 47\nrelevant candidates remains both computationally intensive and analytically complex. 48\nWhile high-performance computing (HPC) resources are essential to manage the scale of 49\nthese simulations, intelligent data-driven strategies are equally critical for prioritizing 50\nbiologically significant conformations. To reduce the substantial computational burden 51\ntypically associated with ensemble docking, this work introduces performance-oriented 52\nAI/ML frameworks that integrate seamlessly with HPC environments and ensemble 53\ndocking outputs to support protein conformation selection and classification. By 54\ndirectly addressing the challenges of extreme class imbalance and limited experimental 55\nvalidation, these frameworks enable the efficient identification of rare, high-value 56\nbinding conformations. As a result, they significantly improve the accuracy, robustness, 57\nand overall reliability of computational drug discovery pipelines.58\nIn this work, we introduce a multi-modal framework that integrates global and local 59\ndescriptors to improve the classification of protein conformations. Global descriptors 60\ncapture overall structural and physicochemical properties—such as mass, 61\nhydrophobicity, and radius of gyration, while local descriptors, including 62\npharmacophores, provide site-specific information critical for identifying ligand-binding 63\nFebruary 15, 2026 3/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\ninteractions. These complementary feature sets offer a more comprehensive view, 64\nenhancing classification performance. Like global descriptor datasets, pharmacophore 65\ndata suffer from class imbalance, with only a limited number of conformations showing 66\nstrong binding activity. This imbalance can lead to biased models and poor67\ngeneralization. To address this, we apply previously established data-driven68\naugmentation techniques [11]- [13] to balance the dataset and improve model reliability. 69\nTo capture both spatial and statistical dependencies, we construct graph-based 70\nrepresentations for each descriptor type: a proximity matrix for pharmacophores (local 71\ndescriptors) and a Pearson correlation matrix for global descriptors. These graphs are 72\nprocessed using a Graph Convolutional Network (GCN) trained with a contrastive loss 73\nfunction, which improves feature discrimination by pulling together similar 74\nconformations and separating dissimilar ones in the embedding space. The resulting 75\ngraph-level embeddings are then classified using traditional ML algorithms, and a 76\ndecision fusion strategy aggregates their outputs for improved performance. This 77\nintegrated approach combines structural insight with advanced graph learning to 78\nprovide a scalable, interpretable, and effective solution for protein conformation 79\nclassification, with potential to accelerate AI-driven drug discovery. 80\nMaterials and methods 81\n0.1 Dataset Overview 82\nFour proteins were selected to evaluate the effectiveness of our proposed model: 83\nADORA2A (Adenosine A2A Receptor), ADRB2 (Beta-2 Adrenergic Receptor), OPRD1 84\n(Delta Opioid Receptor), and OPRK1 (Kappa Opioid Receptor). These targets were 85\npreviously studied and reported in our earlier work [11]– [13]. Each protein includes 86\nexperimentally validated conformations that either (a) bind to ligands (binding 87\nconformations) or (b) do not bind to ligands (non-binding conformations), as previously 88\ndescribed in [8]. 89\nTo improve the classification of binding versus non-binding protein conformations,90\nwe employed both global and local structural descriptors, including pharmacophoric 91\nfeatures specific to ligand-binding sites. Global descriptors capture overall structural 92\nFebruary 15, 2026 4/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nand physicochemical properties of a protein conformation and were described in our 93\nprior work [11]– [13]. In contrast, local descriptors were derived from pharmacophore 94\nfeatures using the DB-PH4 module in MOE (Molecular Operating Environment). These 95\nfeatures were extracted within a 6.5 ˚A radius of the ligand-binding site using the 96\n”unified scheme,” which includes hydrogen bond donors (Don), acceptors (Acc), cations 97\n(Cat), anions (Ani), aromatic centers (Aro), and hydrophobic centers (Hyd). The 98\nMD-derived conformational dataset containing these local descriptors has also been 99\ndocumented and published in [14]. 100\nTable 1 summarizes the number of binding and non-binding protein conformations, 101\nthe total number of conformations, the number of local and global descriptors, and the 102\nclass imbalance ratio for each target protein used in this study.103\nT able 1.Dataset description\nT\narget protein #\nof binding\nprotein conformations\n#\nof non-binding\nprotein conformations\nT\notal # of\nprotein conformations\n#\nof\nlocal descriptors\n#\nof\nglobal descriptors\nClass-im\nbalance\nratio\nADORA2A\n(adenosine\nreceptor A2A) 850 2,150 3,000 282 50 3:1\nADRB2\n(\nβ2-adrenergic receptor) 154 2,411 2,565 542 51 16:1\nOPRK1\n(\nκ-type opioid receptor) 137 2,862 2,999 697 50 21:1\nOPRD1\n(\nδ-type opioid receptor) 72 2,932 3,004 329 51 41:1\n0.2 Pearson Correlation Matrix 104\nOur study’s global features—such as protein mass, volume, radius of gyration, 105\nhydrophobic surface area, and mobility—are generic to all protein conformations and 106\nlack explicit spatial dependency information for GCN input. To address this, we use the 107\nPearson correlation matrix to capture pairwise linear relationships between features, 108\nrevealing how changes in one relate to another. It measures the linear relationship 109\nbetween two global descriptors. The correlation coefficient rab, ranging from -1 to 1, 110\nquantifies the strength and direction of this relationship, showing how changes in one 111\nfeature predict changes in another. It is calculated as follows [15]: 112\nrab =\nP\ni(ai − ¯a)(bi − ¯b)qP\ni(ai −¯a)2(bi − ¯b)2\n(1)\nwhere ai and bi are individual feature values and ¯a, ¯b their respective means. A value of 113\n+1 indicates a perfect positive linear relationship, -1 a perfect negative linear 114\nrelationship, and 0 no linear correlation. Values closer to ±1 indicate stronger linear 115\nFebruary 15, 2026 5/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nrelationships, while those near zero indicate weak or no correlation. 116\n0.3 Gaussian Naive Bayes Classifier 117\nThe Gaussian Naive Bayes (GB) classifier is a probabilistic supervised learning method 118\nbased on Bayes’ theorem, assuming feature independence and Gaussian-distributed 119\ncontinuous variables [16]. For each feature, GB estimates the mean (µ b) and variance 120\n(σ2\nb) per class b to compute the likelihood using: 121\nP (a|b) = 1p\n2πσ 2\nb\nexp\n\u0012\n−(a−µ b)2\n2σ2\nb\n\u0013\n(2)\nClassification relies on the maximum a posteriori (MAP) estimate derived from 122\nBayes’ theorem: 123\nP (b|a) = P (a|b)P (b)\nP(a) (3)\nWhereP(b) is the prior, P (a|b) is the likelihood, and P(b|a) is the posterior 124\nprobability of class b given data point a. The model predicts the class with the highest 125\nposterior probability: 126\nclass = arg max\nb\n(P(a|b)P(b)) (4)\nThis lightweight and interpretable model is especially effective when the features are 127\nconditionally independent and normally distributed [17]. 128\n0.4 K-Nearest Neighbor 129\nK-Nearest Neighbors (KNN) is a straightforward, non-parametric supervised learning 130\nalgorithm used for classification and regression. It predicts the label of a new data point 131\nxby identifying its K closest neighbors using the Euclidean distance metric, 132\nD(x, yi) =\np\n(x−y i)2, and assigning x to the class with the majority among these 133\nneighbors. Despite its simplicity, KNN faces limitations such as high memory usage and 134\nsensitivity to imbalanced data [18], [19], [20]. 135\nFebruary 15, 2026 6/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n0.5 Random Forest 136\nRandom Forest (RF) is a robust ensemble learning method used for classification and 137\nregression. It constructs multiple decision trees during training, and for classification 138\ntasks, the final class label is determined by majority voting across these trees, 139\nimproving accuracy [21]. 140\nThe training process begins by creating bootstrap samples—random subsets of the 141\ntraining data selected with replacement—ensuring each tree is trained on a unique 142\ndataset. At each node, a random subset of features is considered, and the best feature 143\nfrom this subset is selected for splitting. This randomness introduces diversity among 144\ntrees, reducing overfitting and enhancing generalization [22]. 145\nTrees are grown fully without pruning to capture complex data patterns. The best 146\nsplits are chosen by minimizing the Gini impurity G(M), defined as: 147\nG(M) = 1 −\ncX\ni=1\n(p2\ni ) (5)\nwhere G(M) is the Gini impurity of Node M, c is the total number of classes and pi 148\nis the probability of selecting class i. 149\nEach tree independently predicts the class of an input sample. For binary 150\nclassification, the final prediction is the class receiving the majority vote among all trees. 151\nFor example, if more trees predict Class 1 (binding) than Class 0 (non-binding), the 152\ninstance is assigned to Class 1. This ensemble approach reduces overfitting, improves 153\naccuracy, and increases robustness against noise. 154\n0.6 Support Vector Machine 155\nSupport Vector Machines (SVMs) are widely used for binary classification by finding 156\nthe optimal hyperplane that separates two classes in an n-dimensional feature space [23]. 157\nThe hyperplane maximizes the margin, defined as the distance between the hyperplane 158\nand the closest points (support vectors) from each class, ensuring reliable class 159\nseparation. Given a labeled training set {(x1, y1), (x2, y2), . . . ,(xn, yn)}, where xn is a 160\nfeature vector and yn its class label, the optimal hyperplane satisfies [24]: 161\nwxT + m = 0 (6)\nFebruary 15, 2026 7/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nwith w as the weight vector, x the feature vector, and m the bias. The classification 162\nconstraints are: 163\nwxT\nn +m >0,if yn = 1 (7)\nwxT\nn + m < 0, if yn = 0 (8)\nTraining adjusts w and m to maximize the margin 1 /∥w∥2, enhancing generalization. 164\nFor non-linearly separable data, SVM employs kernel functions to implicitly map 165\ninputs to higher-dimensional spaces. The linear kernel computes the inner product: 166\nK(xi, xj) = xi · xj (9)\nand the decision function is: 167\nq(x) = w · x + m =\nnX\nj=1\nβjyj(xj · x) + m (10)\nwhere βj are Lagrange multipliers [25]. The linear kernel works well when data is 168\nlinearly separable. 169\nThe Radial Basis Function (RBF) kernel extends SVM’s flexibility by using a 170\nGaussian function: 171\nK(xi, xj) = exp(−α∥xi − xj∥2) (11)\nq(x) =\nnX\nj=1\nβjyj exp(−α∥xj − x∥2) + m (12)\nwhere α controls the influence radius of each training point [26]. Low α (gamma) 172\nyields smoother decision boundaries with risk of underfitting, while high α creates 173\ntighter boundaries that may overfit noise. 174\nFebruary 15, 2026 8/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n0.7 Graph Convolutional Neural Network175\nGraphs are widely used in domains such as social analysis, bioinformatics, and computer 176\nvision to capture structural relationships between data, providing richer insights than 177\nanalyzing individual data points. A Graph Convolutional Network (GCN) is a deep 178\nlearning model designed to process graph-structured data, which is inherently 179\nnon-Euclidean—examples include protein structures, molecules, and social networks. 180\nGCNs learn node embeddings by iteratively aggregating information from neighboring 181\nnodes, thereby capturing both node features and the overall graph structure [27]. 182\nFormally, a graph G consists of nodes N and edges E representing connections 183\nbetween node pairs [28]. In our study, nodes correspond to local or global descriptors, 184\nwhile edges represent spatial distances between local descriptors or linear dependencies 185\nexpressed as correlation values between global descriptors. The two main components of 186\na GCN are the adjacency matrix A, which encodes graph topology, and the node feature 187\nmatrix X, which records node attributes. For our datasets, X contains encoded binary 188\nvectors for local descriptors and numerical feature values for global descriptors. The 189\nGCN aggregates information from neighboring nodes to make predictions. The 190\noperation of a GCN layer is described as follows [29]: 191\n• The first step is to add self-loops and normalize the adjacency matrix. Graphs 192\nmay or may not contain self-loops. Self-loops are included to ensure that each 193\nnode takes into account its own feature. Equation (13) shows the creation of 194\nmatrix ¯A by adding self-loops to the adjacency matrix A, where E represents the 195\nidentity matrix used to incorporate self-connections. 196\n¯A = A + E (13)\n197\n¯Dii =\nX\nj\nAij (14)\nHere, ¯D denotes the degree matrix, which can be computed as shown in Equation 198\n(14). The normalized adjacency matrixU is then calculated using Equation (15). 199\nU = ¯D− 1\n2 ¯A ¯D− 1\n2 (15)\nFebruary 15, 2026 9/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n•The next step is node feature aggregation, where each node combines its own200\nfeature vector with those of its neighbors to obtain local context. This aggregation 201\nprocess leverages the normalized adjacency matrix U and is expressed in Equation 202\n(16): 203\n¯H (k) =U ¯H (k−1) (16)\nHere, ¯H (k−1) represents node features from the previous layer. This operation 204\nsuccessfully propagates and blends information across neighboring nodes, 205\nguaranteeing that each node representation is enhanced by its local graph 206\nstructure. 207\n• The following step is feature transformation, which involves the GCN applying a 208\nlearnable linear transformation on the aggregated node features. This technique is 209\nsimilar to using a Dense (or Linear) layer in a classical neural network. It allows210\nthe model to modify, combine, and enhance the aggregated data to better fit the 211\ntask at hand. Equation (17) illustrates the computation of the transformed 212\nfeature matrix: 213\nH (k) = ¯H (k)W (k) (17)\nHere, W (k) is the learnable weight matrix that determines how the aggregated 214\nfeatures are transformed. 215\n• Lastly, a non-linearity, such as ReLU, is applied element-by-element to the 216\ntransformed features, as indicated in Equation (18). This activation function 217\nintroduces nonlinearity to the model, allowing it to capture complicated patterns 218\nand relationships in the same way that hidden layers in traditional neural 219\nnetworks do. 220\nH (k) = ReLU( ¯H (k)W (k)) (18)\nThe GCN model in this study consists of two graph convolutional layers followed by a 221\nmean pooling layer, which aggregates node features by averaging them to produce a 222\nfixed-size graph-level embedding that captures the overall structure and feature 223\ndistribution. 224\nTo enhance the embeddings’ discriminative power, the GCN is trained using a 225\nFebruary 15, 2026 10/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\ncontrastive loss function that pulls embeddings of similar graphs closer while pushing 226\ndissimilar ones apart. The loss is defined as follows: 227\nLoss = 1\n2 yD2 + 1\n2(1−y ) (max(0, m− D))2 (19)\nwhere D = ∥ei − ej∥2 is the Euclidean distance between embeddings ei and ej, and 228\ny indicates pair similarity (y = 1 for similar pairs—both binding or non-binding—and 229\ny = 0 for dissimilar pairs). The margin m sets the minimum distance for negative pairs. 230\nThe first term pulls positive pairs together, while the second pushes negative pairs apart 231\nif closer than m. 232\n0.8 Decision Fusion 233\nIn this study, we provide a decision fusion strategy that improves the detection of 234\nbinding and non-binding protein conformations while also enhancing the AI/ML 235\nmodel’s capacity to distinguish between them. We propose a method that leverages the 236\npredictive power of four machine learning techniques which operate on embeddings 237\nproduced by a unique dual Graph Convolution Network (GCN) model. This dual GCN 238\nmodel consists of two separately trained GCN on a dataset that contains local spatial 239\nand global linear relationship respectively, allowing it to capture a more complete 240\ndescription of the data. 241\nLet us define the prediction results obtained from the four machine learning 242\nmodels—Gaussian Na¨ ıve Bayes (GB), K-Nearest Neighbor (KNN), Random Forest 243\n(RF), and Support Vector Machine (SVM)—applied to GCN embeddings of local and 244\nglobal descriptors as follows: LGB, LKN N, LRF and LSV M represent the predictions 245\nobtained using embeddings of GCN model trained on local descriptors and GGB, 246\nGKN N, GRF and GSV M represent the predictions obtained using embeddings of GCN 247\nmodel trained on global descriptors. After obtaining the individual predictions from 248\neach dataset, the cumulative prediction score is computed by summing all the 249\nFebruary 15, 2026 11/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\npredictions, as defined by the following equation (19). 250\nT otalP rediction = LGB + LKN N + LRF\n+ LSV M + GGB + GKN N\n+ GRF + GSV M\n(20)\nwhere T otalP rediction is used to compute the final classification results. Conformations 251\nwith a cumulative score of at least 3 are allocated to Class ’1’, indicating binding 252\nprotein conformation, while those with a cumulative score of no more than 5 are 253\nassigned to Class ’0’, indicating non-binding protein conformation. 254\n0.9 Evaluation metrics 255\nThe confusion matrix and its derived metrics—such as accuracy, sensitivity, and 256\nspecificity—are widely used to evaluate the performance of machine learning classifiers. 257\nFor binary classification of binding vs. non-binding protein conformations, the confusion 258\nmatrix includes four outcomes [11]: 259\n•True Positive (TP): Correctly predicted binding conformations (Class 1).260\n•False Positive (FP): Non-binding conformations incorrectly predicted as binding. 261\n• False Negative (FN): Binding conformations incorrectly predicted as non-binding. 262\n•True Negative (TN): Correctly predicted non-binding conformations (Class 0). 263\nThe accuracy measures the overall proportion of correct predictions: 264\nAccuracy = T P + T N\nT P+T N+F P+F N (21)\nSensitivity quantifies the model’s ability to correctly identify binding conformations: 265\nSensitivity = T P\nT P+F N (22)\nSpecificity quantifies the model’s ability to correctly identify non-binding266\nconformations: 267\nSpecificity = T N\nT N+F P (23)\nFebruary 15, 2026 12/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n0.10 Enrichment Ratio 268\nTo validate the proposed AI/ML framework, true positive (TP) and false negative (FN) 269\npredictions were used to compute the enrichment ratio. As outlined in our previous 270\nwork [11], a baseline enrichment ratio—representing the expected performance under 271\nrandom selection—was calculated for comparison. This metric provides a benchmark to 272\nassess how effectively the AI/ML framework improves prediction performance relative 273\nto random selection of protein conformations. 274\nTable 2 summarizes the maximum Base Enrichment ratios computed for each target 275\nprotein. 276\nProtein Maximum Enrichment Ratio\nADORA2A ∼9\nADRB2 ∼5.4\nOPRD1 ∼2\nOPRK1 ∼5.5\nT able 2.Maximum Base Enrichment ratios calculated for each target proteins.\n0.11 Proposed Work 277\nThis study introduces a dual-GCN framework that integrates both local and global 278\nfeatures for protein conformation prediction. One GCN is trained on 279\npharmacophore-based local descriptors to capture spatial interactions at binding sites, 280\nwhile the other learns from global descriptors to model broader structural patterns. 281\nEach GCN is optimized using a contrastive loss to enhance the separation between 282\nbinding and non-binding conformations in the embedding space. The resulting 283\nembeddings are then fed into four traditional machine learning classifiers, whose 284\noutputs are combined through decision fusion to yield the final prediction. An 285\nenrichment ratio framework is applied to validate binding conformations in test proteins. 286\nThis approach leverages the representational strength of GCNs and the predictive power 287\nof classical models to improve protein conformation selection. An overview of the 288\nmethod is shown in Fig 1. 289\n• The summary of Framework 1, which handles the global descriptor data, is as 290\nfollows: 291\n– It begins with the original dataset, which is then refined using the feature 292\nFebruary 15, 2026 13/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nscoring method described earlier. 293\n–The ML-based feature selection and scoring framework from our prior 294\nwork [12], [13], is used to identify the most significant protein characteristics. 295\nSpecifically, four methods are used: Analysis of Variance (ANOVA) [30], 296\nMutual Information (MI) [31], Recurrence Quantification Analysis 297\n(RQA) [32], and Spearman Correlation [33]. Each method independently 298\nranks the features based on its criteria. These rankings are then integrated 299\nusing a majority voting strategy, where features that receive the highest 300\npossible vote count of 4—indicating consensus across all methods—are 301\nselected. The resulting subset of features is used to construct a new, 302\noptimized dataset for further analysis. This framework, along with the 303\nselected features for each target protein—ADORA2A, ADRB2, OPRD1, and 304\nOPRK1—has been previously described and published in [12].305\n–To address the high class imbalance observed in the proteins ADRB2, 306\nOPRK1, and OPRD1, a multi-step data re-balancing approach was 307\nimplemented as detailed in [12]. First, the Extreme Gradient Boosting 308\n(XGBoost) [34] classifier was applied to the feature-reduced dataset to 309\ndistinguish between non-binding (TN) and binding (TP) protein 310\nconformations. Next, K-Means [35]- [37] clustering was performed on the 311\nnon-binding conformations to retain representative samples while eliminating 312\nredundant data, thereby reducing dataset bias and size. Finally, Generative313\nAdversarial Networks (GANs) were employed to augment the minority class 314\nby generating additional binding conformation samples, achieving balanced 315\nclass representation without increasing the overall dataset size. For the 316\nADORA2A protein, this re-balancing step was omitted due to its low class 317\nimbalance, which did not necessitate adjustment for the GCN model. 318\n– In parallel with the re-balancing procedure, a Pearson correlation matrix was 319\nconstructed to capture pairwise linear relationships among the global 320\nfeatures. This analysis provides insight into how changes in one feature relate 321\nto changes in another, enhancing understanding of feature interactions. 322\nFollowing this, a graph dataset was prepared for input into the GCN model. 323\nFebruary 15, 2026 14/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nHere, global features correspond to nodes, the Pearson correlation 324\ncoefficients serve as edge weights, and the numerical values of the features 325\nare assigned as node attributes. 326\n– The GCN model was then trained on this graph dataset using 10-fold 327\ncross-validation to ensure reliable and generalizable performance. Early 328\nstopping was incorporated to prevent overfitting by monitoring the validation 329\nloss and halting training when no further improvement was observed. The 330\nmodel with the best validation results was saved and used to generate the 331\nfinal graph-level embeddings. 332\n– These graph-level embeddings, derived from the optimized GCN model, were 333\nsubsequently fed into four traditional machine learning classifiers — 334\nGaussian Na¨ ıve Bayes (GB), K-Nearest Neighbor (KNN), Random Forest 335\n(RF), and Support Vector Machine (SVM) — to assess and compare their 336\nclassification performances. 337\n• The summary of Framework 2, which processes local descriptor (pharmacophore) 338\ndata, is as follows: 339\n– The framework begins by applying the multi-step re-balancing technique 340\ndescribed in [12]. Unlike the global descriptor dataset, the original local 341\ndescriptor data is excluded from the feature scoring process due to its high 342\nsparsity. Sparse data poses challenges for GCN training because limited and 343\nuneven feature availability hampers the aggregation of meaningful patterns 344\nacross nodes. This issue is exacerbated by conformations containing only one 345\nor two pharmacophores in close proximity. Removing such features based on 346\nrelevance risks losing critical local information necessary for capturing 347\nconformation-specific interactions. Therefore, to preserve essential local 348\ninformation, the original local descriptor dataset bypasses the feature scoring 349\nframework. 350\n–To address class imbalance, the previously described multi-step data 351\nre-balancing method is applied. Initially, the feature-reduced dataset is 352\nclassified with XGBoost to distinguish binding (TP) from non-binding (TN) 353\nconformations. K-Means clustering is then used on the TN samples to retain 354\nFebruary 15, 2026 15/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nrepresentative instances and reduce redundancy. Finally, GANs augment the 355\nminority class by generating additional binding conformations, balancing the 356\nclasses without increasing overall dataset size. 357\n– Next, a graph dataset is prepared for input to the GCN model, where local 358\nfeatures serve as nodes, and the distances between them define edge weights. 359\nEach node is further characterized by attributes including an encoded binary 360\nvector, radius, and frequency corresponding to the respective local feature. 361\n– The graph dataset is processed using the GCN model with 10-fold 362\ncross-validation and early stopping to ensure stable, reliable performance and 363\nprevent overfitting. The model exhibiting the best validation performance is 364\nselected to generate the final graph-level embeddings, which are then input365\ninto four machine learning classifiers for evaluation and classification 366\nperformance recording. 367\n• GEFusion (Graph Embedding Fusion), a decision-level fusion strategy, is applied 368\nto combine the classification results from Framework 1 and Framework 2. This 369\nintegration enables the unique identification of the total number of probable 370\nbinding and non-binding protein conformations. 371\n• Finally, enrichment ratios are calculated using True Positives (TP)—correctly 372\npredicted binding conformations—and False Negatives (FN)—incorrectly 373\npredicted non-binding conformations—based on the outcomes of the GEFusion 374\ndecision strategy. 375\nFig 1. Bold the figure title. Figure caption text here, please use this space for the\nfigure panel descriptions instead of using subfigure commands. A: Lorem ipsum dolor\nsit amet. B: Consectetur adipiscing elit.\nResults and Discussion 376\n0.12 Computational Evaluation on ADORA2A Dataset 377\nTable 3 presents the classification performance of the ADORA2A protein across varying 378\ntraining sizes, evaluated using the proposed GEFusion methodology. Notably, the model 379\nFebruary 15, 2026 16/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nachieves its highest performance—measured by accuracy, sensitivity, and 380\nspecificity—when trained on 40% of the dataset. The results for this training size, 381\nincluding the performance of individual classifiers applied to both local and global 382\nembeddings, are summarized in Table 4 383\nT able 3. Classification performance (in %) across different training sizes for protein\nADORA2A\nT raining Size Accuracy(%) Sensitivity(%) Specificity(%)\n0.1 90.93 88.09 92.05\n0.2 91.88 84.66 94.72\n0.3 88.52 72.90 94.53\n0.4 94.94 93.24 95.61\n0.5 89.00 70.05 96.22\n0.6 87.83 68.96 95.14\n0.7 92.00 85.48 94.48\n0.8 89.33 68.45 97.45\nT able 4.Classification performance for 40% training size for protein ADORA2A\nApproac\nh TN FP FN TP Accuracy\n(%) Sensitivit\ny (%) Sp\necificity (%)\nKNN\n(local embeddings) 303 994 69 434 40.94 86.28 23.36\nGB\n(local embeddings) 687 610 130 373 58.89 74.16 52.97\nRF\n(local embeddings) 0 1297 0 503 27.94 100.00 0.00\nSVM\n(local embeddings) 191 1106 22 481 37.33 95.63 14.73\nKNN\n(global embeddings) 1130 167 432 71 66.72 14.12 87.12\nGB\n(global embeddings) 850 447 278 225 59.72 44.73 65.54\nRF\n(global embeddings) 1297 0 503 0 72.06 0.00 100.00\nSVM\n(global embeddings) 1297 0 503 0 72.06 0.00 100.00\nGEF\nusion 1240 57 34 469 94.94 93.24 95.61\nFrom Table 4, it is evident that while the overall framework delivers strong 384\nclassification performance, some individual models exhibited extreme predictive 385\nbehavior. For instance, the Random Forest (RF) model trained on local embeddings 386\nclassified all conformations as binding, whereas both RF and Support Vector Machine 387\n(SVM) models trained on global embeddings predicted all conformations as non-binding. 388\nThese individual biases are effectively mitigated by the framework’s voting-based 389\nensemble mechanism, which combines predictions from multiple models. This ensemble 390\nstrategy reduces the impact of the outlier behavior of any single model, resulting in 391\nmore balanced and reliable classification results. 392\nAs shown in Table 5, our proposed methodology demonstrated strong performance 393\nnot only in overall model accuracy but also in achieving a high enrichment ratio.\nT able 5. Enrichment ratios of protein ADORA2A on a training size of 40%\nApproac\nh Maxima Filter %\nof Data Minima Filter %\nof Data\nGEF\nusion 13.67 Filter\nA 0.5% 13.07 Filter\nA 1.0%\n394\nFebruary 15, 2026 17/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n0.13 Computational Evaluation on ADRB2 Dataset395\nTable 6 presents the classification performance of the proposed approach across various 396\ntraining sizes for the ADRB2 protein. Among these, the 40% training size yields the 397\nbest overall performance, achieving higher accuracy, sensitivity, and specificity. Detailed 398\nresults for this optimal training size—including the performance of the proposed399\nframework and individual classifiers utilizing local and global embeddings—are provided 400\nin Table 7. 401\nT able 6.Classification performance (in %) across different training sizes for protein\nADRB2\nT raining Size Accuracy(%) Sensitivity(%) Specificity(%)\n0.1 84.11 64.08 85.42\n0.2 88.35 65.63 89.86\n0.3 88.47 64.29 90.08\n0.4 92.14 80.61 92.92\n0.5 93.06 42.68 96.50\n0.6 90.55 56.06 92.92\n0.7 86.49 66.00 87.92\n0.8 93.37 8.82 99.37\nT able 7.Classification performance for 40% training size for protein ADRB2\nApproac\nh TN FP FN TP Accuracy\n(%) Sensitivit\ny (%) Sp\necificity (%)\nKNN\n(local embeddings) 634 807 49 49 44.38 50.00 44.00\nGB\n(local embeddings) 355 1086 22 76 28.01 77.55 24.64\nRF\n(local embeddings) 348 1093 22 76 27.55 77.55 24.15\nSVM\n(local embeddings) 260 1181 17 81 22.16 82.65 18.04\nKNN\n(global embeddings) 978 463 60 38 66.02 38.78 67.87\nGB\n(global embeddings) 1000 441 62 36 67.32 36.73 69.40\nRF\n(global embeddings) 1376 65 87 11 92.06 11.22 95.49\nSVM\n(global embeddings) 1372 69 92 6 89.54 6.12 95.21\nGEF\nusion 1339 102 19 79 92.14 80.61 92.92\nTable 8 provides an overview of the derived enrichment ratios based on the decision 402\nmodel’s prediction outcomes (TP and FN), serving as a validation measure for the 403\npredictions generated by our proposed framework. 404\nT able 8. Enrichment ratios of protein ADRB2 on a training size of 40%\nApproac\nh Maxima Filter %\nof Data Minima Filter %\nof Data\nGEF\nusion 33.71 Filter\nA 0.5% 28.10 Filter\nC 1.0%\n0.14 Computational Evaluation on OPRD1 Dataset405\nTable 9 presents the classification performance of the protein OPRD1 across various406\ntraining sizes using our AI/ML-based multi-modal framework for protein conformation 407\nFebruary 15, 2026 18/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nselection and prediction. The best overall performance—measured by accuracy,408\nsensitivity, and specificity—is achieved with a training size of 10%. Table 10 further 409\ndetails the classification results of the proposed framework and the individual 410\nperformance of classifiers utilizing local and global embeddings at this optimal training 411\nsize. 412\nT able 9.Classification performance (in %) across different training sizes for protein\nOPRD1\nT raining Size Accuracy(%) Sensitivity(%) Specificity(%)\n0.1 82.64 90.91 82.44\n0.2 83.18 85.96 83.11\n0.3 93.15 58.18 94.09\n0.4 93.00 67.35 93.72\n0.5 87.74 88.64 87.71\n0.6 93.42 86.21 93.60\n0.7 96.34 39.13 97.84\n0.8 90.50 75.00 90.82\nT able 10.Classification performance for 10% training size for protein OPRD1\nApproac\nh TN FP FN TP Accuracy\n(%) Sensitivit\ny (%) Sp\necificity (%)\nKNN\n(local embeddings) 1749 887 20 46 66.43 69.70 66.35\nGB\n(local embeddings) 1783 853 26 40 67.47 60.61 67.64\nRF\n(local embeddings) 1598 1038 12 54 61.14 81.82 60.62\nSVM\n(local embeddings) 2087 549 33 33 78.46 50.00 79.17\nKNN\n(global embeddings) 1308 1328 18 48 50.19 72.73 49.62\nGB\n(global embeddings) 1225 1411 18 38 47.30 67.86 46.47\nRF\n(global embeddings) 1558 1078 23 43 59.25 65.15 59.10\nSVM\n(global embeddings) 1025 1611 11 55 39.97 83.33 38.88\nGEF\nusion 2173 463 6 60 82.64 90.91 82.44\nT able 11.Enrichment ratios of protein OPRD1 on a training size of 10%\nApproac\nh Maxima Filter %\nof Data Minima Filter %\nof Data\nGEF\nusion 38.33 Filter\nA 1.0% 32.38 Filter\nB 0.5%\nTable 11 presents the resulting enrichment ratios derived from the predictions made 413\nby the GEFusion decision strategy, further supporting the validity and effectiveness of 414\nthe proposed framework. 415\n0.15 Computational Evaluation on OPRK1 Dataset 416\nTable 12 summarizes the classification performance across various training sizes for the 417\nprotein OPRK1 using the proposed AI/ML-based multimodal framework for protein 418\nconformation selection and prediction. The highest overall performance—reflected by 419\nimproved accuracy, sensitivity, and specificity—was achieved with a training size of 30%. 420\nFebruary 15, 2026 19/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nTable 13 presents the classification results at this training size, including the 421\nperformance of individual classifiers applied to both local and global embeddings. 422\nT able 12. Classification performance (in %) across different training sizes for protein\nOPRK1\nT raining Size Accuracy(%) Sensitivity(%) Specificity(%)\n0.1 83.33 64.00 84.27\n0.2 86.04 62.07 87.25\n0.3 73.08 81.19 72.67\n0.4 77.88 79.31 77.80\n0.5 84.66 70.27 85.40\n0.6 92.17 53.33 94.21\n0.7 78.67 73.47 78.97\n0.8 84.17 72.73 84.60\nT able 13.Classification performance for 30% training size for protein OPRK1\nApproac\nh TN FP FN TP Accuracy\n(%) Sensitivit\ny (%) Sp\necificity (%)\nKNN\n(local embeddings) 1005 993 52 49 50.21 48.51 50.30\nGB\n(local embeddings) 1295 703 61 40 63.60 39.60 64.81\nRF\n(local embeddings) 0 1998 0 101 4.81 100.00 0.00\nSVM\n(local embeddings) 745 1253 29 72 38.92 71.29 37.29\nKNN\n(global embeddings) 1066 932 46 55 53.41 54.46 53.35\nGB\n(global embeddings) 1317 681 52 58 60.58 48.51 65.92\nRF\n(global embeddings) 1273 725 54 47 62.89 46.53 63.71\nSVM\n(global embeddings) 987 1011 38 63 50.02 62.38 49.40\nGEF\nusion 1452 546 19 82 73.08 81.19 72.67\nAs shown in Table 14, our proposed methodology demonstrated strong performance 423\nnot only in overall model accuracy but also in achieving a high enrichment ratio. 424\nT able 14. Enrichment ratios of protein OPRK1 on a training size of 30%\nApproac\nh Maxima Filter %\nof Data Minima Filter %\nof Data\nGEF\nusion 32.91 Filter\nD 5.0% 29.74 Filter\nB 1.0%\nFrom Tables 3, 6, 9, and 12, it is observed that the model consistently performs well 425\nacross all training sizes for protein ADORA2A. However, for proteins ADRB2, OPRK1, 426\nand OPRD1, better overall performance is achieved with smaller training sizes. This 427\ntrend likely stems from higher class imbalance ratios in these datasets. At larger 428\ntraining sizes, the increased proportion of synthetic binding conformation data, coupled 429\nwith inherent data sparsity, challenges the GCN model’s ability to distinguish binding430\nfrom non-binding classes, thus affecting the framework’s predictive accuracy. 431\nThe enrichment ratios shown in Tables 5, 8, 11, and 14 demonstrate that binding 432\nconformations can be identified with significantly higher accuracy compared to random 433\nsampling, as evidenced by the baseline results in Table 2. In particular, the highest 434\nenrichment is achieved when selecting the top 0.5% to 1% of the binding conformations, 435\nFebruary 15, 2026 20/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nindicating that the most predictive and biologically relevant data points are 436\nconcentrated among the candidates ranked highest. 437\nConclusion 438\nThis work presented an AI/ML-based multi-modal framework designed to improve the 439\ndetection of both binding and non-binding protein conformations by integrating local 440\nand global structural information. The proposed approach leverages Graph 441\nConvolutional Networks (GCNs) trained with a contrastive loss function to capture 442\nmeaningful connectivity patterns within protein structures. The GCN learns 443\ndiscriminative graph embeddings that effectively group binding conformations while 444\nseparating non-binding ones. These embeddings are then processed by traditional 445\nmachine learning classifiers, and their outputs are combined using a decision fusion 446\nstrategy to enhance overall classification performance. 447\nThe study demonstrated that incorporating both global descriptors—reflecting 448\nprotein-level structural and stability features—and local descriptors—such as 449\npharmacophores that encode site-specific binding information—provides complementary 450\ninsights that significantly improve conformation classification. Additionally, the use of 451\ngraph embeddings learned via contrastive learning offers a more compact and 452\nnoise-resistant data representation than conventional graph classification methods. 453\nThese embeddings preserve essential structural relationships while facilitating a clearer 454\ndistinction between classes. 455\nOverall, the integration of multi-modal data, GCN-based embedding learning, and 456\nensemble decision-making contributes to a robust and generalizable framework capable 457\nof accurately identifying protein binding states. This methodology holds promise for 458\nadvancing structure-based protein function prediction and accelerating drug discovery 459\nworkflows. 460\nAcknowledgments 461\nThe authors gratefully acknowledge Dr. Armin Ahmadi for providing the 462\nPharmacophore dataset used in this study. 463\nFebruary 15, 2026 21/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nReferences\n1. Nass, S. J. et al. Accelerating anticancer drug development — opportunities and\ntrade-offs. Nat. Rev. Clin. Oncol. 15, 777–786 (2018).\n2. Oca˜ na, A., Garc´ ıa-Alonso, S., Amir, E. & Pandiella, A. Refining early\nantitumoral drug development. Trends Pharmacol. Sci. 39, 922–925 (2018).\n3. W. Evangelista, R. L. Weir, S. R. Ellingson, J. B. Harris, K. Kapoor, J. C. Smith,\nand J. Baudry. ”Ensemble-based docking: From hit discovery to metabolism and\ntoxicity predictions.” Bioorganic & Medicinal Chemistry, vol. 24, no. 20, October\n2016, pp. 4928–4935.\n4. Anwar T., Kumar P., Khan A.U. Molecular Docking for Computer-Aided Drug\nDesign: Fundamentals, Techniques, Resources, and Applications. Elsevier;\nAmsterdam, The Netherlands: 2021. Modern Tools and Techniques in\nComputer-Aided Drug Design; pp. 1–30.\n5. Bajorath J. Deep Machine Learning for Computer-Aided Drug Design. Front.\nDrug Discov. 2022;2:829043. doi: 10.3389/fddsv.2022.829043.\n6. Ye F., Lin M., Jin J., Broussy S. Editorial: Computer-Aided Drug Design: Drug\nDiscovery, Computational Modelling, and Artificial Intelligence. Front. Chem.\n2022;10:968687. doi: 10.3389/fchem.2022.968687\n7. Oli B. Revolutionizing Drug Discovery: A Comprehensive Review of\nComputer-Aided Drug Design Approaches. Int. J. Res. Appl. Sci. Eng. Technol.\n2024;12:308–317. doi: 10.22214/ijraset.2024.63563.\n8. Amaro, R.E.; Baudry, J.; Chodera, J.; Demir, ¨O.; McCammon, J.A.; Miao, Y.;\nSmith, J.C. Ensemble Docking in Drug Discovery. Biophys. J. 2018, 114,\n2271–2278.\n9. M. Parvathaneni, A. K. Awol, M. Kumari, K. Lan, and M. Lingam. ”Application\nof artificial intelligence and machine learning in drug discovery and development.”\nJournal of Drug Delivery and Therapeutics, vol. 13, no. 1, January 2023, pp.\n151–158.\nFebruary 15, 2026 22/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n10. R. Awasthi, S. Mishra, R. Grasfield, J. Maslinski, D. Mahapatra, J. B. Cywinski,\nA. K. Khanna, K. Maheshwari, C. Dave, A. Khare, F. A. Papay, and P. Mathur.\n”Artificial intelligence in healthcare: 2023 year in review.” medRxiv (Cold Spring\nHarbor Laboratory), February 2024.\n11. Gupta S, Baudry J, Menon V. Using Big Data Analytics to “Back Engineer”\nProtein Conformational Selection Mechanisms. Molecules. 2022; 27(8):2509.\nhttps://doi.org/10.3390/molecules27082509\n12. Gupta, S., Baudry, J., & Menon, V. (2022). Big data analytics for improved\nprediction of ligand binding and conformational selection. Retrieved from\nhttps://www.frontiersin.org/articles/10.3389/fmolb.2022.953984/full\n13. S. Gupta, V. Menon and J. Baudry, ”Wavelet-based Spectral Analysis For\nProtein Conformation Selection and Prediction Using AI in Drug Discovery\nApplications,” 2022 IEEE International Conference on Bioinformatics and\nBiomedicine (BIBM), Las Vegas, NV, USA, 2022, pp. 2595-2602, doi:\n10.1109/BIBM55620.2022.9995169.\n14. A. Ahmadi, S. Gupta, V. Menon, and J. Baudry. Linking machine learning and\nbiophysical structural features in drug discovery. Frontiers in Molecular\nBiosciences, 11 (2025), 1305272. doi:10.3389/fmolb.2024.1305272.\n15. P. Schober, C. Boer, and L. A. Schwarte. Correlation Coefficients: Appropriate\nUse and Interpretation. Anesthesia & Analgesia, 126(5) (2018), 1763–1768.\ndoi:10.1213/ANE.0000000000002864.\n16.\nA. H. Jahromi and M. Taheri. A non-parametric mixture of gaussian naive bayes\nclassifiers based on local independent features. Artificial Intelligence and Signal\nProcessing Conference (AISP), IEEE (2017), 209–212.\n17. J. Ren, S. D. Lee, X. Chen, B. Kao, R. Cheng, and D. Cheung. Naive bayes\nclassification of uncertain data. Ninth IEEE International Conference on Data\nMining, IEEE (2009), 944–949.\n18. L. Wang, L. Khan, and B. Thuraisingham. An effective evidence theory based\nk-nearest neighbor (KNN) classification. IEEE/WIC/ACM International\nFebruary 15, 2026 23/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\nConference on Web Intelligence and Intelligent Agent Technology, 1 (2008),\n797–801.\n19. S. Chen. K-nearest neighbor algorithm optimization in text categorization. IOP\nConference Series: Earth and Environmental Science, 108 (2018), 052074.\n20. S. Taneja, C. Gupta, K. Goyal, and D. Gureja. An enhanced k-nearest neighbor\nalgorithm using information gain and clustering. International Conference on\nAdvanced Computing & Communication Technologies, IEEE (2014), 325–329.\n21. G. Louppe. Understanding Random Forests: From Theory to Practice. arXiv\npreprint, 2014.\n22. R. Couronn´ e, P. Probst, and A. L. Boulesteix. Random forest versus logistic\nregression: a large-scale benchmark experiment. BMC Bioinformatics, 19, 270\n(2018). https://doi.org/10.1186/s12859-018-2264-5\n23. W. S. Noble. What is a support vector machine? Nature Biotechnology, 24(12)\n(2006), 1565–1577.\n24. S. Huang, N. Cai, P. P. Pacheco, S. Narrandes, Y. Wang, and W. Xu.\nApplications of support vector machine (SVM) learning in cancer genomics.\nCancer Genomics & Proteomics, 15(1) (2018), 41–51.\n25. A. F. Rochim, K. Widyaningrum, and D. Eridani. Performance Comparison of\nSupport Vector Machine Kernel Functions in Classifying COVID-19 Sentiment.\n2021 4th International Seminar on Research of Information Technology and\nIntelligent Systems (ISRITI), Yogyakarta, Indonesia, 2021, pp. 224–228.\ndoi:10.1109/ISRITI54043.2021.9702845.\n26. B. Scholkopf et al. Comparing support vector machines with Gaussian kernels to\nradial basis function classifiers. IEEE Transactions on Signal Processing, 45(11)\n(1997), 2758–2765. doi:10.1109/78.650102.\n27. S. Zhang, H. Tong, J. Xu, et al. ”Graph convolutional networks: A\ncomprehensive review.” Computational Social Networks, vol. 6, article 11, 2019.\nhttps://doi.org/10.1186/s40649-019-0069-y\nFebruary 15, 2026 24/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n28. U. A. Bhatti, H. Tang, G. Wu, S. Marjan, and A. Hussain. Deep Learning with\nGraph Convolutional Networks: An Overview and Latest Applications in\nComputational Intelligence. International Journal of Intelligent Systems, 2023,\nArticle ID 8342104, 28 pages. https://doi.org/10.1155/2023/8342104\n29. F. Wu, A. Souza, T. Zhang, C. Fifty, T. Yu, and K. Weinberger. Simplifying\nGraph Convolutional Networks. Proceedings of the 36th International Conference\non Machine Learning, in Proceedings of Machine Learning Research 97 (2019),\n6861–6871.\n30. K. J. Johnson and R. E. Synovec. Pattern recognition of jet fuels: comprehensive\nGC×GC with ANOVA-based feature selection and principal component analysis.\nChemometrics and Intelligent Laboratory Systems, 60(1–2) (2002), 225–237.\n31. J. Song, Z. Zhu, P. M. D. Scully, and C. Price. Modified mutual\ninformation-based feature selection for intrusion detection systems in decision\ntree learning. Journal of Computers, 9(7) (2014), 1542–1546.\n32. J.-P. Eckmann, S. O. Kamphorst, and D. Ruelle. Recurrence plots of dynamical\nsystems. Europhysics Letters, 4(9) (1987), 973–977.\n33. J. Hauke and T. Kossowski. Comparison of values of Pearson’s and Spearman’s\ncorrelation coefficients on the same sets of data. Quaestiones Geographicae, 30(2)\n(2011), 87–93.\n34. T. Chen and C. Guestrin. Xgboost: A scalable tree boosting system.\narXiv:1603.02754 (2016).\n35. O. J. Oyelade, O. O. Oladipupo, and I. C. Obagbuwa. Application of k means\nclustering algorithm for prediction of students academic performance. arXiv\npreprint arXiv:1002.2425 (2010).\n36. V. S. Akondi, V. Menon, J. Baudry, and J. Whittle. ”Novel big data-driven\nmachine learning models for drug discovery application.” Molecules, vol. 27, no. 3,\nJanuary 2022, article 594.\n37. V. S. Akondi, V. Menon, J. Baudry, and J. Whittle. ”Novel K-means\nclustering-based undersampling and feature selection for drug discovery\nFebruary 15, 2026 25/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\napplications.” In Proceedings of the 2019 IEEE International Conference on\nBioinformatics and Biomedicine (BIBM), San Diego, CA, USA, 18–21 November\n2019, pp. 2771–2778.\nFebruary 15, 2026 26/26\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint \n\n.CC-BY 4.0 International licenseavailable under a \n(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 \nThe copyright holder for this preprintthis version posted February 18, 2026. ; https://doi.org/10.64898/2026.02.17.706293doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}