Enhanced bioactivity prediction via spatial emptiness representation of protein-ligand complexes
By integrating spatial emptiness representation and a union pocket concept with a pairwise loss function, the method enhances bioactivity prediction accuracy by considering unoccupied regions and maintaining consistent protein target information, addressing the limitations of existing methods that neglect global geometric context.
Patent Information
- Application Number
- PCT/SG2026/050288
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-08
- Filing Date
- 2026-05-06
- Publication Date
- 2026-11-12
AI Technical Summary
Existing bioactivity prediction methods focus narrowly on local interactions within protein-ligand complexes, neglecting spatial emptiness and global geometric context, which is crucial for understanding competitive inhibition and accurate bioactivity modeling.
Incorporating a geometric representation of spatial emptiness and a union pocket concept into protein-ligand complexes, along with a pairwise loss function, to enhance bioactivity prediction by considering unoccupied regions and maintaining consistent protein target information.
Improves prediction accuracy by capturing competitive binding aspects and preserving the relative order among samples, leading to more robust and context-aware bioactivity assessments.
Smart Images

Figure SG2026050288_12112026_PF_FP_ABST
Abstract
Description
ENHANCED BIOACTIVITY PREDICTION VIA SPATIAL EMPTINESS REPRESENTATION OF PROTEIN-LIGAND COMPLEXESTECHNICAL CONTRIBUTION
[0001] The present disclosure relates to bioactivity prediction of potential protein-ligand pairs. More particularly, the present disclosure relates to enhanced bioactivity prediction via spatial emptiness representation of protein-ligand complexes.BACKGROUND
[0002] Predicting the bioactivity of candidate ligands remains a central challenge in drug discovery. Ligands and endogenous substrates often compete for the same binding sites on target proteins, and the extent to which a ligand can modulate protein function depends not only on its binding but also on how effectively it occupies the relevant pocket. However, most existing methods focus narrowly on local interactions within protein-ligand complexes and neglect spatial emptiness — the unoccupied regions within the binding site that may permit endogenous molecules to engage or interfere. Such unfilled space can diminish the ligand’s functional impact, regardless of binding affinity.
[0003] Bioactivity is the measurable effect a compound exerts on a biological system, usually by modulating the functional activity of a specific molecular target (e.g., a protein or nucleic acid). It can be quantified experimentally through enzymatic kinetics (Vmax, Km, kcat / Km) or functional and binding assays that yield parameters such as EC50, IC50, Emax, and Ki. Accurate measurement — and increasingly, in-silico prediction — of these effects allows researchers in drug discovery, chemical biology, and pharmacology to identify the most promising compounds early, reducing the cost and duration of large-scale experimental screens. With the growing size of chemical libraries and biological data, there is an increasing demand for efficient and accurate computational methods to predict bioactivity. Developing robust bioactivity prediction methods has therefore become a key focus in both academic research and industrial applications.
[0004] Recently, machine learning approaches have shown great potential in modeling biological molecules and their interactions, dramatically accelerating the drug discovery process There have also been some machine learning methods for bioactivity prediction, aiming at early-stage screening of the potential active molecules. Conventional approaches rely on the intrinsic properties of small molecules as features to model ligands, aiming to uncover the relationship between molecular structure and bioactivity. A traditional strategy is to use quantitative structure-activity relationship (QSAR) models, which primarily utilize molecular descriptors derived from the physicochemical and structural properties of small molecules. While this approach is reasonable and widely adopted, it may overlook the fact that bioactivity fundamentally arises from specific biological interactions, which enable more thorough modeling and understanding bioactivity. A compound cannot function biologically without its target and environment. Furthermore, unlike other single-target prediction tasks, bioactivity is often evaluated using multiple measurement types, such as IC50, EC50, Kd, and Ki, making the task more challenging. In response to these challenges, recent studies have shifted toward developing models based on curated pocket-ligand interaction datasets that provide more comprehensive information, including abundant bound conformations with proteins and multiple bioactivity measurements. Among these methods designed for protein -ligand interaction, graph neural networks (GNNs) have become particularly prevalent, as graph-based representations offer a natural and powerful way to capture molecular properties and relational structures.
[0005] While recent studies increasingly focus on bioactivity prediction using protein information and constructing protein-ligand interaction graphs, they often concentrate on local binding regions but overlook the global geometric context of the entire pocket, missing crucial information for accurate bioactivity modeling. Geometric features are known to be important for representing binding status and have shown effectiveness in modeling molecular interactions. For instance, geometric representations in E(3)-equivariant GNNs have demonstrated strong performance in capturing local spatial relationships on protein surfaces, improving binding site prediction. However, most existing approaches emphasize detailed descriptions of local binding regions and structural complementarity between interacting molecules The existing approaches do not consider pocket-ligand interactions from the perspective of spatial emptiness within the docked conformation. This aspect, which remains underexplored, corresponds closely with the principle of competitive inhibition, where drugmolecules bind to the active site of a protein and block access for endogenous substrates For example, protein kinase inhibitors bind to the enzyme’s ATP -binding pocket, blocking ATP access and thereby suppressing kinase activity. Previous methods tend to focus on how well the ligand fits the pocket, but often overlook the issues of, once binding occurs, how much empty space remains and to what extent the ligand occupies the pocket.
[0006] Furthermore, current methods for bioactivity prediction typically focus on the local binding pocket in complex with a ligand, which limits access to global pocket information that is biologically significant. Additionally, in many prior works, each pocket-ligand pair is treated as an independent data point overlooking the realistic scenario in which multiple potential drug compounds may interact with different pockets of the same protein. Recent findings underscore the importance of evaluating bioactivity across different ligands acting on a common protein, enabling more meaningful and fair comparisons. In response, new methods and datasets have been developed to incorporate multiple ligands per protein and to perform evaluation at the protein level, allowing for more robust and context-aware assessments. However, even with a fixed protein, these approaches often involve distinct binding sites within the same protein, introducing variability that may compromise the consistency of bioactivity predictions
[0007] Additionally, a major distinction between drug discovery and other regression tasks lies in the importance of prediction order. The most common training loss, mean-square error (MSE), is symmetric and treats each data point independently. For example, given two ground-truth bioactivities, yi = 2 and 2 = 1, it gives the same loss value for the predictions (yi = 1.5, j = 1.5) and (f’y = 1.5, y- = 0.5). The order of the former is wrong, but the order of the latter aligns with the ground truths.
[0008] for these reasons, what is needed is an enhanced bioactivity prediction method that takes into account spatial emptiness representation of protein-ligand complexes, evaluation of bioactivity across different ligands acting on a common protein, and the enablement of more meaningful comparisons.SUMMARY OF INVENTION
[0009] A general embodiment of the invention provides a structural representation of spatial emptiness in each protein-ligand complex to describe the surroundings of each node in the protein-ligand graph representation. This surrounding space quantifies the emptiness within the docking conformation, incorporating a broader global perspective into the overall interaction graph representation. Quantification of emptiness holds practical significance, as it plays a pivotal role in assessing whether the pocket space is sufficiently tight to hinder the entry of other small molecules into the binding site area, thereby contributing to a more accurate prediction of the compound’s bioactivity.
[0010] In an alternative embodiment of the invention, a union pocket concept is applied to the bioactivity prediction task to ensure that the entire pocket space can be measured. When evaluating different ligands binding to the same protein, the union pocket concept encourages the model to keep the protein target information consistent and encourages the model to focus on the ligand. The ligand positioning within a union pocket can impact binding. Thus, the union pocket concept enhances the ability of the model to capture subtle conformational differences within a consistent protein context.
[0011] An alternative embodiment of the invention incorporates a pairwise loss alongside the conventional mean squared error (MSE) loss as the final training objective for bioactivity prediction. This combined loss encourages the model to preserve the relative order among samples by emphasizing pairwise differences in their bioactivity labels.
[0012] A first embodiment of the invention is a computer implemented method or system for predicting a bioactivity of a ligand with respect to a target protein. The system or method performs the steps of: (a) obtaining structural representations of the target protein and the ligand; (b) obtaining or generating a plurality of candidate binding poses of the ligand relative to the target protein; (c) for each candidate binding pose, defining a corresponding protein binding pocket of the target protein; (d) constructing, for each binding pose and corresponding binding pocket, a protein -ligand complex structural representation in which nodes represent ligand atoms and protein pocket atoms and edges represent interactions between the atoms; (e) augmenting the protein-ligand complex structural representation with a geometricrepresentation of spatial emptiness, including defining a local spherical coordinate system centered at each node, partitioning space surrounding each node into a plurality of directional cones, and introducing geometric edges encoding spatial relationships indicative of unoccupied pocket space; (f) providing the augmented protein-ligand complex structural representations as inputs to a machine-learning model; and (g) executing the machine-learning model to generate, for the ligand, a predicted bioactivity value relative to the target protein. In the system, the steps are performed by one or more processors in conjunction with memory.
[0013] In an alternative embodiment, the steps of the method or system further comprise constructing a unified binding pocket representation for the target protein by aggregating atoms from a plurality of protein binding pockets associated with the target protein to form a union pocket, and providing the union pocket together with each augmented protein-ligand complex structural representation as inputs to the machine-learning model.
[0014] In an alternative embodiment, the machine-learning model is trained using a loss function that incorporates a pairwise loss term comprising pairwise comparisons between bioactivity values of different ligands associated with the target protein.
[0015] Bioactivity prediction is a crucial task in computational drug discovery. Embodiments of the invention introduce a geometric representation of spatial emptiness in protein-ligand complexes (referred to, for convenience, as “GeoREC”), and the union pocket, a unified structural context for multiple pockets, to enhance bioactivity prediction in drug discovery. By measuring unoccupied space and integrating global pocket information, embodiments of the invention capture aspects of competitive binding that were previously overlooked. Additionally, a pairwise loss function is employed to better preserve the relative order among samples. Extensive experiments across diverse datasets and bioactivity measurements demonstrate that use of embodiments of the invention improves prediction accuracy.BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Embodiments of the present disclosure are described herein with reference to the drawings in which:FIG. 1 illustrates a target, a first ligand candidate, a second ligand candidate, and an endogenous substrate.FIGs. 2A and 2B illustrate a first interaction between a target and a first ligand candidate.FIGs.3A and 3B illustrate a second interaction between a target and a second ligand candidate.FIGs.4A and 4B are side-by-side comparisons of the second interaction and a third interaction.FIG. 5 illustrates a spherical coordinate system with a directional bin indicated.FIGs. 6A and 6B illustrate example configuration of the spherical coordinate system.FIGs. 7A to 7C illustrate a union pocket in the target in FIG. 7A.FIGs. 8A and 8B illustrate the location of a ligand binding to a first region of the target in FIG.8A and the location of the same ligand alternatively binding to an adjacent second region of the target in FIG. 8B.FIG. 9 is a flowchart of the steps performed in an embodiment of the invention.FIGs. 10 to 24 include Tables 1-1 to 2-15 and Tables 2-1 to 2-3.DETAILED DESCRIPTION
[0017] In the following detailed description, reference is made to the accompanying drawings, which form a part hereof. The illustrative embodiments described in the detailed description, drawings and claims are not meant to be limiting. Other embodiments can be utilized, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein. Unless specified otherwise, the terms “comprising,” “comprise,” “including” and “include” used herein, and grammatical variants thereof, are intended to represent “open” or “inclusive” language such that they include recited elements but also permit inclusion of additional, un-recited elements.
[0018] As used herein, a controller and its associated memory may be included in the hardware of a personal computer, a laptop, and / or a server. The controller may comprise one or more microcontrollers and / or graphics processing units. A “server” may be implemented within a stand-alone server, multiple dedicated servers, and / or a virtual server running on a larger network of servers and / or a cloud-based service.
[0019] FIG. 1 illustrates a target 01, a first ligand candidate 02A, a second ligand candidate 02B, and an endogenous substrate 03.
[0020] FIGs.2A and 2B detail a first interaction 10A between the target 01 and the first ligand candidate 02A. FIG. 2A is an expanded view of the first interaction 10A, while FIG. 2B is an expanded view of the first interaction 10A.
[0021] FIGs. 3A and 3B detail a second interaction 10B between the target 01 and the second ligand candidate 02B. FIG. 3A is an expanded view of the second interaction 10B, while FIG.3B is an expanded view of the second interaction 10B.
[0022] FIGs.4A and 4B are side-by-side comparisons of the second interaction 10B and a third interaction 10C. The second interaction 10B is between the target 01 and the second ligand candidate 02B. The third interaction 10C is between the target 01 and an endogenous substrate 03. This comparison demonstrates that the second interaction 10B results in more vacant space in the pocket than in the third interaction 10C. As a result, the second interaction 10B betweenthe target 01 and the second ligand candidate 02B is likely to be unsuccessful for inhibition purposes.
[0023] FIG. 5 illustrates a spherical coordinate system 30 with a directional bin 31 indicated.
[0024] FIG. 6A and 6B illustrate example configuration of the spherical coordinate system. In FIG. 6A, the i-th atom of a ligand is the origin point of the spherical coordinate system. Distance e(ji) geo is measured between the j-th atom of the target to the i-th atom of the ligand, by analysis of directional bin n1. Distance e(ki) geo is measured between the k-th atom of the target to the i-th atom of the ligand, by analysis of directional bin n2. Hence, these distances are determined by examination of separate directional bins.
[0025] In FIG. 6B, the i-th atom of a ligand is the origin point of the spherical coordinate system. Distance e(ji) geo is measured between the j-th atom of the ligand to the i-th atom of the target, by analysis of directional bin n1. Distance e(ki) geo is measured between the k-th atom of the ligand to the i-th atom of the target, measured by analysis of directional bin n2. Hence, these distances are determined by examination of separate directional bins
[0026] FIG. 7A to 7C illustrate a union pocket in the target 01 in FIG. 7A, wherein the union pocket is defined by a first data point 40A (see FIG. 7B for expanded view) and a second data point 40B (see FIG. 7C for expanded view). Both data points help to identify a union pocket 40C of a third ligand 40D within the target 01. FIG. 7B illustrates the third ligand 40D binding to a first region of the target 01 to indicate a first data point 40A. FIG. 7C illustrates the third ligand 40D binding to a second region of the target 01 to indicate a second data point 40B. The represented third ligand 40D is the same ligand in both of FIGs. 7B and 7C.
[0027] The third ligand 40D, thus, can bind to different regions within the target 01. FIG. 8A illustrates the location of the third ligand 40D binding to a first region of the target 01 at a first data point 40A within the union pocket 40C. FIG. 8B illustrates the location of the third ligand 40D binding to a second region of the target 01 at a second data point 40B within the union pocket 40C.
[0028] FIG. 9 is a flowchart of the steps performed in an embodiment of the invention. These method steps include:
[0029] A general embodiment of the invention includes the preparation of a structural representation of the local spatial emptiness around each atom in a protein-ligand graph. This structural representation can be augmented by the construction of a unified binding pocket that is the union of all local pockets sampled from multiple ligand docking poses on the same protein. This structural representation can be further augmented by formulation of a pairwise loss to penalize incorrect ordering between sample pairs in the context of real -valued bioactivity prediction.
[0030] A first embodiment of the invention is a computer implemented method or system for predicting a bioactivity of a ligand with respect to a target protein The system or method performs the steps of: (a) obtaining structural representations of the target protein and the ligand; (b) obtaining or generating a plurality of candidate binding poses of the ligand relative to the target protein; (c) for each candidate binding pose, defining a corresponding protein binding pocket of the target protein; (d) constructing, for each binding pose and corresponding binding pocket, a protein-ligand complex structural representation in which nodes represent ligand atoms and protein pocket atoms and edges represent interactions between the atoms; (e) augmenting the protein-ligand complex structural representation with a geometric representation of spatial emptiness, including defining a local spherical coordinate system centered at each node, partitioning space surrounding each node into a plurality of directional cones, and introducing geometric edges encoding spatial relationships indicative of unoccupied pocket space; (f) providing the augmented protein-ligand complex structural representations as inputs to a machine learning model; and (g) executing the machine learning model to generate, for the ligand, a predicted bioactivity value relative to the target protein. In the system, the steps are performed by one or more processors in conjunction with memory.
[0031] In an alternative embodiment, the plurality of candidate binding poses for the ligand is represented as a pose set Xi in accordance with Equation (1-1):where xl,irepresents the i-th pose of the ligand l associated with one of the available protein pockets, and d is the dimensionality of the input representation.
[0032] In an alternative embodiment, the predicted bioactivity value yi is generated in accordance with Equation (1-2):yi = fiXi. Pi).(|-2> where / is the machine-learning model, Xi is a set of candidate binding poses of the ligand / , Pi is a set of candidate protein binding pockets associated with the ligand, and yi is the predicted bioactivity value for the ligand.
[0033] Tn an alternative embodiment, wherein the predicted bioactivity value / is obtained using a permutation-invariant aggregation in accordance with Equation (1-3):where jposeis a pose-level model applied to a pose representation x / ,;and a corresponding pocket pi,i, Agg is a permutation-invariant aggregation function, and Ni is a number of candidate binding poses.
[0034] In an alternative embodiment, a bioactivity prediction for a pose-pocket pair is generated in accordance with Equation (1-4):where is a bioactivity prediction model, xi is a representation of the ligand 1 including at least one binding pose, p is a candidate binding pocket of the target protein, and yi is a predicted bioactivity value.
[0035] In an alternative embodiment of the invention, the predicted bioactivity value / is computed according to:(1-5)where X represents a set of binding pose representations of a ligand / , and 7 / represents a subset of binding pockets of a same protein associated with the binding poses.
[0036] In an alternative embodiment, the predicted bioactivity value comprises a continuous quantitative metric selected from the group consisting of IC50, EC50, Ki, and Kd.
[0037] In an alternative embodiment, the plurality of candidate binding poses is generated using a molecular docking process or derived from experimentally determined protein-ligand structures.
[0038] In an alternative embodiment, each local spherical coordinate system is defined independently for each atom of the protein-ligand complex.
[0039] In an alternative embodiment, the geometric edges encode directional proximity information by identifying, for each directional cone, a nearest atom or an absence thereof.
[0040] In an alternative embodiment, unoccupied pocket space represents regions of the protein binding pocket not occupied by the ligand in a given binding pose.
[0041] In an alternative embodiment, the space surrounding each node is partitioned into S directional cones, and S = 32.
[0042] In an alternative embodiment, the machine-learning model comprises a graph neural network operating on the augmented protein-ligand complex structural representation.
[0043] In an alternative embodiment, modeling unoccupied pocket space enables prediction of ligand bioactivity reflective of competitive inhibition with endogenous substrates.
[0044] In an alternative embodiment, the steps of the method or system further comprise constructing a unified binding pocket representation for the target protein by aggregating atoms from a plurality of protein binding pockets associated with the target protein to form a union pocket, and providing the union pocket together with each augmented protein-ligand complex structural representation as inputs to the machine-learning model. The union pocket can be assembled from a union of atom sets from multiple local binding pockets associated with different ligand poses or different ligands bound to the same target protein. The union pocketcan be used consistently across multiple ligands associated with the target protein to enable comparative evaluation of ligand bioactivity within a shared pocket context.
[0045] The union pocket can be defined in accordance with Equation (1-6):where^4( / ?[) denotes a set of atoms of alocal binding pocket / ?', L denotes a set ofligands docked to the target protein, and Pi denotes a set of local binding pockets associated with ligand I.
[0046] Further, the predicted bioactivity value y / can be generated using the union pocket in accordance with Equation (1-7):(1-7) / ZPunioii ) •where funion is a machine-learning model configured to accept a pose set Xi and a union pocket Pmion, and v,- is the predicted bioactivity value for the ligand I. The machine-learning model can be trained using a loss function that incorporates a pairwise loss term comprising pairwise comparisons between bioactivity values of different ligands associated with the target protein.
[0047] In an alternative embodiment, the machine-learning model is trained using a loss function that incorporates a pairwise loss term comprising pairwise comparisons between bioactivity values of different ligands associated with the target protein.
[0048] The loss function can include a pointwise mean-squared error loss L2 in accordance with Equation (1-8),(1-8)where N denotes a number of training samples in a batch, yi denotes a predicted bioactivity value for an / -th sample, and yi denotes a ground-truth bioactivity value for the / -th sample.
[0049] The pairwise loss term can be computed in accordance with Equation (1-9):based on differences between pairs of samples, where N denotes a number of training samples in a batch, y, andare ground-truth bioactivity values for samples i and / , and and j) are predicted bioactivity values for samples i and j, such that the pairwise loss penalizes a discrepancy between (y, yj) and (y, yj).
[0050] The loss function can comprise a hybrid loss Lhybnd in accordance with Equation (1-10):(MO)where L2 is a pointwise mean-squared error loss, Lpairwise is the pairwise loss term, and Xi and X2 are weighting coefficients that control relative contributions of L2 and Lpairwise. These weighting coefficients can satisfy Xi = 0.8 and X2 = 0.2.
[0051] Another embodiment of the invention is a non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of the above listed embodiments of the invention.Citations
[0052] Experimental use of embodiments of the invention are detailed below. Conditions for these experiments draw on the following published background references:
[0053] Ahmet Sureyya Rifaioglu, Heval Atas, Maria Jesus Martin, Rengul Cetin-Atalay, Volkan Atalay, and Tunca Dogan. Recent applications of deep learning and machine intelligence on in silico drug discovery: methods, tools and databases. Briefings in bioinformatics, 20(5): 1878-1912, 2019.
[0054] Alexander Tropsha, Olexandr Isayev, Alexandre Varnek, Gisbert Schneider, and Artem Cherkasov. Integrating qsar modelling and deep learning in drug discovery: the emergence of deep qsar. Nature Reviews Drug Discovery, 23(2): 141-155, 2024.
[0055] Ali Vefghi, Zahed Rahmati, and Mohammad Akbari. Drug-target interaction / affmity prediction: Deep learning models and advances review. arXiv preprint arXiv:2502.15346, 2025.
[0056] Alpha A Lee, Qingyi Yang, Asser Bassyouni, Christopher R Butler, Xinjun Hou, Stephen Jenkinson, and David A Price. Ligand biological activity predicted by cleaning positive and negative chemical correlations. Proceedings of the National Academy of Sciences, 116(9):3373-3378, 2019.
[0057] Andrea Mastropietro, Giuseppe Pasculli, and Jurgen Bajorath. Learning characteristics of graph neural networks predicting protein-ligand affinities. Nature Machine Intelligence, 5(12): 1427-1436, 2023.
[0058] Anna Gaulton, Louisa J Bellis, A Patricia Bento, Jon Chambers, Mark Davies, Anne Hersey, Yvonne Light, Shaun McGlinchey, David Michalovich, Bissan Al-Lazikani, et al. Chembl: a large-scale bioactivity database for drug discovery. Nucleic acids research, 4O(D1): D1100-D1107, 2012.
[0059] Antonio Bai ci Kinetics of enzyme-modifier interactions. Selected Topics in the Theoryand Diagnosis of Inhibition and Activation Mechanisms, 1, 2015
[0060] Artem Cherkasov, Eugene N Muratov, Denis Fourches, Alexandre Varnek, Igor I Baskin, Mark Cronin, John Dearden, Paola Gramatica, Yvonne C Martin, Roberto Todeschini, et al. Qsar modeling: where have you been? where are you going to? Journal of medicinal chemistry, 57(12): 4977-5010, 2014.
[0061] Bharath Ramsundar, Peter Eastman, Pat Walters, and Vijay Pande. Deep learning for the life sciences: applying deep learning to genomics, microscopy, drug discovery, and more. O’Reilly Media, 2019.
[0062] Bin Feng, Zequn Liu, Nanlan Huang, Zhiping Xiao, Haomiao Zhang, Srbuhi Mirzoyan, Hanwen Xu, Jiaran Hao, Yinghui Xu, Ming Zhang, et al. A bioactivity foundation model using pairwise meta-learning. Nature Machine Intelligence, 6(8): 62-974, 2024.
[0063] Bonggun Shin, Sungsoo Park, Keunsoo Kang, and Joyce C Ho. Self-attention based molecule representation for predicting drug-target interaction In Machine learning for healthcare conference. PMLR, 2019.
[0064] Daniel Fernandez-Llaneza, Silas Ulander, Dea Gogishvili, Eva Nittinger, Hongtao Zhao, and Christian Tyrchan. Siamese recurrent neural network with a self-attention mechanism for bioactivity prediction. ACS omega, 6(16): 11086-11094, 2021.
[0065] David Mendez, Anna Gaulton, A Patricia Bento, Jon Chambers, Marleen De Veij, Eloy Felix, Maria Paula Magarinos, Juan F Mosquera, Prudence Mutowo, Michal Nowotka, et al. Chembl: towards direct deposition of bioassay data. Nucleic acids research, 47(D1): D93O-D940, 2019.
[0066] Dixian Zhu, Tianbao Yang, and Livnat Jerby. Gradient aligned regression via pairwise losses. arXiv preprint arXiv:2402.06104, 2024.
[0067] Elena L Caceres, Nicholas C Mew, and Michael J Keiser. Adding stochastic negative examples into machine learning improves molecular bioactivity prediction. Journal ofChemical Information and Modeling, 60(12):5957-5970, 2020.
[0068] Eric MarEchal. Measuring bioactivity: Ki, ic50 and ec50. Chemogenomics and Chemical Genetics: A User’s Introduction for Biologists, Chemists and Informaticians, pages 55-65, 2011.
[0069] Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni -mol: A universal 3d molecular representation learning framework. In The Eleventh International Conference on Learning Representations, 2023.
[0070] Hakime Oztiirk, Arzucan Ozgiir, and Elif Ozkirimli Olmez. Deepdta: deep drug-target binding affinity prediction. Bioinform., 2018.
[0071] Holger Gohlke, Manfred Hendlich, and Gerhard Klebe. Knowledge-based scoring function to predict protein-ligand interactions. loumal of molecular biology, 2000.
[0072] Hongming Chen, Ola Engkvist, Yinhai Wang, Marcus Olivecrona, and Thomas Blaschke. The rise of deep learning in drug discovery. Drug discovery today, 23(6): 1241-1250, 2018.
[0073] Hong-Qiang Ding, Naoki Karasawa, and William A Goddard III. Atomic level simulations on a million particles: The cell multipole method for coulomb and london nonbond interactions. The lournal of chemical physics, 97(6):4309-4315, 1992.
[0074] Hyone-Myong Eun. Enzymology primer for recombinant DNA technology. Elsevier, 1996.
[0075] IM2443682 Kapetanovic. Computer-aided drug discovery and development (caddd): in silico-chemico-biological approach. Chemico-biological interactions, 171(2): 165-176, 2008.
[0076] Izhar Wallach, Michael Dzamba, and Abraham Heifets. Atomnet: a deep convolutional neural network for bioactivity prediction in structure-based drug discovery. arXiv preprint arXiv: 1510.02855, 2015.
[0077] James B Dunbar Jr, Richard D Smith, Chao-Yie Yang, Peter Man-Un Ung, Katrina W Lexa, Nicko-lay A Khazanov, Jeanne A Stuckey, Shaomeng Wang, and Heather A Carlson. Csar benchmark exercise of 2010: selection of the protein-ligand complexes. Journal of chemical information and modeling, 51(9):2036-2046, 2011.
[0078] Jiaxian Yan, Zhaofeng Ye, Ziyi Yang, Chengqiang Lu, Shengyu Zhang, Qi Liu, and Jiezhong Qiu. Multi-task bioassay pre-training for protein-ligand binding affinity prediction. Briefings in Bioinformatics, 25(1), 2023.
[0079] Jiyun Han, Shizhuo Zhang, Mingming Guan, Qiuyu Li, Xin Gao, and Juntao Liu. Geonet enables the accurate prediction of protein-ligand binding sites through interpretable geometric deep learning. Structure, 32(12):2435-2448, 2024.
[0080] Journal of Chemical Theory and Computation, 19(15):5242-5259, 2023.
[0081] Luca Gagliardi and Walter Rocchia Siteferret: beyond simple pocket identification in proteins.
[0082] Marcel L Verdonk, Jason C Cole, Michael J Hartshorn, Christopher W Murray, and Richard D Taylor. Improved protein-ligand docking using gold. Proteins: Structure, Function, and Bioinformatics, 52(4):609-623, 2003.
[0083] Michael Tynes, Wenhao Gao, Daniel J Burrill, Enrique R Batista, Danny Perez, Ping Yang, and Nicholas Lubbers. Pairwise difference regression: a machine learning metaalgorithm for improved prediction and uncertainty quantification in chemical search. Journal of chemical information and modeling, 61(8):3846-3857, 2021.
[0084] Muhammad Ishfaq, Muhammad Aamir, Farooq Ahmad, Abdelazim M Mebed, and Sayed Elshahat. Machine learning-assisted prediction of the biological activity of aromatase inhibitors and data mining to explore similar compounds. ACS omega, 7(51) 48139-48149, 2022.
[0085] Oleg Trott and Arthur J Olson. Autodock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry, 31 (2):455-461, 2010.
[0086] Parixit Prajapati, Princy Shrivastav, Jigna Prajapati, and Bhupendra Prajapati. Deep learning approaches for predicting bioactivity of natural compounds. The Natural Products Journal, 2025.
[0087] Pedro J. Ballester and John B. O. Mitchell. A machine learning approach to predicting protein-ligand binding affinity with applications to molecular docking. Bioinform., 2010.
[0088] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and YoshuaBengio. Graph attention networks. arXiv preprint arXiv: 1710.10903, 2017.
[0089] RDKit Team Rdkit: Open-source cheminformatics. URL https: / / www.rdkit.org. Accessed: 2025-05-20.
[0090] Renxiao Wang, Luhua Lai, and Shaomeng Wang. Further development and validation of empirical scoring functions for structure-based binding affinity prediction. J. Comput. Aided Mol. Des., 2002.
[0091] Renxiao Wang, Xueliang Fang, Yipin Lu, and Shaomeng Wang. The pdbbind database: Collection of binding affinities for protein-ligand complexes with known three-dimensional structures. Journal of medicinal chemistry, 47(12):2977-2980, 2004.
[0092] Richard A Friesner, Jay L Banks, Robert B Murphy, Thomas A Halgren, Jasna J Klicic, Daniel T Mainz, Matthew P Repasky, Eric H Knoll, Mee Shelley, Jason K Perry, et al. Glide: a new approach for rapid, accurate docking and scoring. 1. method and assessment of docking accuracy. Journal of medicinal chemistry, 47(7):1739-1749, 2004.
[0093] Sarah L. Kinnings, Nina Liu, Peter J. Tonge, Richard M. Jackson, Lei Xie, and Philip E. Bourne. A machine learning-based method to improve docking scoring functions and its application to drug repurposing J. Chem. Inf. Model., 2011.
[0094] Shuangli Li, Jingbo Zhou, Tong Xu, Liang Huang, Fan Wang, Haoyi Xiong, Weili Huang, Dejing Dou, and Hui Xiong. Structure-aware interactive graph neural networks for the prediction of protein-ligand binding affinity. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 975-985, 2021.
[0095] Toshitaka Tanebe and Takashi Ishida. End-to-end learning for compound activity prediction based on binding pocket information. BMC bioinformatics, 22:1-11, 2021.
[0096] Xiangying Zhang, Haotian Gao, Haojie Wang, Zhihang Chen, Zhe Zhang, Xinchong Chen, Yan Li, Yifei Qi, and Renxiao Wang. Planet: a multi -objective graph neural network model for protein-ligand binding affinity prediction. Journal of Chemical Information and Modeling, 64(7): 2205-2220, 2023.
[0097] Yanwen Huang, Bowen Gao, Yinjun Jia, Hongbo Ma, Wei-Ying Ma, Ya-Qin Zhang, and Yanyan Lan. Redefining the task of bioactivity prediction. In The Thirteenth International Conference on Learning Representations, 2025
[0098] Yueming Yin, Haifeng Hu, Zhen Yang, Feihu Jiang, Yihe Huang, and Jiansheng Wu. Afse: towards improving model generalization of deep graph learning of ligand bioactivities targeting gpcr proteins. Briefings in Bioinformatics, 23(3):bbac077, 2022.
[0099] Yueming Yin, Haifeng Hu, Zhen Yang, Huajian Xu, and Jiansheng Wu. Realvs: toward enhancing the precision of top hits in ligand-based virtual screening of drug leads from large compound databases. Journal of chemical information and modeling, 61(10):4924-4939, 2021.
[0100] Yueming Yin, Hilbert Yuen In Lam, Yuguang Mu, Hoi-Yeung Li, and Adams Wai-Kin Kong. Advancing bioactivity prediction through molecular docking and self-attention. IEEE J. Biomed. Health Informatics, 2024.
[0101] Zechen Wang, Sheng Wang, Yangyang Li, Jingjing Guo, Yanjie Wei, Yuguang Mu, Liangzhen Zheng, and Weifeng Li. A new paradigm for applying deep learning to proteinligand interaction prediction. Briefings in bioinformatics, 25(3):bbae145, 2024.
[0102] Zhenqiao Song, Tinglin Huang, Lei Li, and Wengong Jin. Surfpro: Functional protein design based on continuous surface. arXiv preprint arXiv:2405.06693, 2024.
[0103] Zhewei Wei, Ye Yuan, Chongxuan Li, Wenbing Huang, et al. Equipocket: an e (3)-equivariant geometric graph neural network for ligand binding site prediction. In Forty-first International Conference on Machine Learning, 2024.
[0104] Ziduo Yang, Weihe Zhong, Qiujie Lv, Tiejun Dong, and Calvin Yu-Chian Chen. Geometric interaction graph neural network for predicting protein -ligand binding affinities from 3d structures (gign). The journal of physical chemistry letters, 2023.
[0105] Zuobai Zhang, Minghao Xu, Arian Jamasb, Vijil Vijil, Aurelie Lozano, Payel Das, and Jian Tang. Protein representation learning by geometric structure pretraining. In International Conference on Machine Learning, 2022.Tables
[0106] FIGs. 10 to 24 include Tables 1-1 to 1-15 and Tables 2-1 to 2-3.Graph representation
[0107] The chemical node and edge features adopted in the experiments are taken from RDKit [RDKit Team], The details are shown in Table 1-1 and Table 1-2. Note that for the geometric edge, the experiments use Coulomb and London dispersion forces as features based on interatomic distances, as these intermolecular interactions are generally applicable across all molecular pairs. Table 1-1 includes information of node features. Table 1-2 includes information of edge features.
[0108] For geometric calculation, the experiments uniformly partition the surrounding space into S cones for each atom Here, the experiments set S = 32. This is achieved by dividing both the azimuthal and polar angles in spherical coordinates at intervals of π. This angular partitioning allows the surrounding space of each atom to be discretized into uniform sectorsfor geometric feature extraction.
[0109] According to Equation 1-10, see below, the experiments train the models on the training set with early stopping based on validation performance, and report the final results on the test set. In the hybrid loss function,(MO)
[0110] Experiments adopt λ₁ = 0.8 and λ₂ = 0.2 in the training.Implementation of initial individual pockets
[0111] Individual pockets are defined as the set of protein residues with at least one atom located within 5 A of any ligand atom [Yang et al., 2023, Yin et al., 2024, Gagliardi and Rocchia, 2023], This definition follows a commonly used proximity-based criterion that reflects potential physical interactions.
[0112] The union pocket strategy introduces two key types of new information beyond simply aggregating existing docking poses. First, instead of using one large docking box that covers the entire protein surface, the experiments define multiple smaller docking boxes around each cluster of functional sites This approach forces ligands to explore and generate binding poses specifically within these localized regions, reducing sampling bias and ensuring more thorough and uniform coverage of all potential binding sites. Second, by taking the union of all such individual pockets, the resulting union pocket-ligand graph captures the global spatial context of the ligand, including its relative location and the overall geometry of empty regions around potential binding sites. This holistic view of binding topography is not available in any single pose or local pocket, thereby providing richer structural information for bioactivity prediction.
[0113] As shown in Figure 2c (Page 4) of SIU paper [Huang et al., 2025], the SIU dataset contains a total of 249,631 pocket-ligand pairs across 9,544 protein pockets, indicating an average of approximately 26 ligands per pocket. Although it is claimed in the “Pocket definition” section (Page 19) [Huang et al., 2025] that “For each PDB ID, a single pocket wasextracted, defined as the region centered on the co-crystal ligand within a 15 A radius”, the datasets they released via their Hugging Face repository (linked from their GitHub page, which is referenced in the paper’s abstract) contain only local pockets surrounding the individual ligand poses. In other words, the single pocket they used for docking was not fully utilized when constructing the pocket-ligand graphs in their released dataset. For each protein conformation, the experiments use the predefined global pocket-docked by all its associated ligands (e g., the union pocket).Support of using multiple complex conformations
[0114] In these experiments, multiple complex conformations can be explicitly modeled by processing each conformation independently through the graph neural network to obtain individual graph-level embeddings. These embeddings are then aggregated (using either mean or attention-based pooling) before being passed to the bioactivity prediction network (typically a multi-layer perceptron). In this way, the method supports multiple complex conformations as input while leveraging shared network weights to process each conformation consistently.Computational environment
[0115] Most experiments are run on an NVIDIA GeForce RTX 4090 GPU with 24GB of memory. For the DTIGN dataset, the experiments on the subsets containing more than 2000 atoms within the union pocket are run on an NVIDIA H100 NVL GPU with 94GB of memory. For the SIU dataset, the SIU-0.9-Ki version is run on 4 NVIDIA H100 NVL GPUs with 94GB of memory, and other versions are run on 4 NVIDIA GeForce RTX 4090 GPUs with 24GB of memory.Computational time of the geometric edge
[0116] Experiments report the computational time required to calculate the proposed geometric edge using the SIU-0.9-Kd dataset as an example. The results presented are based on singleCPU execution time; in practice, the experiments parallelize the computation using 128 CPUs to accelerate the process. The CPU model is AMD EPYC-Milan Processor. On average, for a pocket-ligand complex containing approximately 1,100 atoms, the geometric edge calculation takes 4.7 seconds per complex on a single CPU. This runtime is considered reasonable and acceptable for practical applications.DTIGN model
[0117] The Drug-Target Interaction Graph Neural Network (DTIGN) model, originally proposed in Yin et al.
[2024] , addresses the challenge of multiple possible ligand poses binding to the same protein. It takes as input multiple docking poses corresponding to a single proteinligand pair and employs a sophisticated GNN architecture combined with a multi-head selfattention mechanism to process and aggregate inter-molecular graph features for predicting bioactivity.
[0118] Distinct from earlier models for bioactivity or binding affinity prediction, DTIGN introduces several novel components. These include intramolecular bond features, as well as newly defined inter-molecular interactions based on high-order negative powers of Euclidean distances — modeled after Coulombic and London dispersion forces — to better capture physical force field characteristics indicative of binding stability [Ding et al., 1992].
[0119] Furthermore, DTIGN is the first to integrate multi-head self-attention and semisupervised learning into this context, enabling the discovery of drug-target interaction patterns from software-generated docking poses. This combination also enhances the model’s ability to focus attention on native-like structural features during training.
[0120] To assess the performance of the DTIGN model, the authors constructed a unique benchmark dataset (the DTIGN dataset used in this study) comprising protein-ligand complexes derived from molecular docking simulations. This dataset was designed to evaluate bioactivityprediction models in the context of protein-ligand interactions. DTIGN’s performance was compared against nine leading deep learning-based bioactivity prediction methods. The results demonstrated that DTIGN improved performance, with an average improvement of 27.03% over the baseline models.GIGN model
[0121] The Geometric Interaction GraphNeural Network (GIGN) model, introduced in [Yang et al., 2023], is a GNN architecture designed to predict protein-ligand binding affinities directly from 3D molecular structures GIGN constructs a joint geometric interaction graph that integrates both the atomic struc-tures of the ligand and the protein pocket, capturing intra- and inter-molecular spatial relationships in a unified representation.
[0122] A key innovation of GIGN is its explicit use of 3D spatial information through relative position encodings and distance-aware message passing. The model defines two types of graph edges: covalent edges for intra-molecular connectivity and non-covalent interaction edges based on spatial proximity across molecules. These are encoded with distances to inform the network of interaction strength.
[0123] The GIGN model was shown to outperform several baselines on PDBbind 2016 (http: / / www.pdbbind.org.cn / ) and CSAR-HiQ [Dunbar Jr et al., 2011] datasets, demonstrat-ing its ability to capture subtle geometric cues relevant for binding affinity prediction.GAT model
[0124] The Graph Attention Network (GAT), proposed in Velic'kovic' et al.
[2017] , is a graph neural network architecture that incorporates self-attention mechanisms to learn node representations by attending over their neighbors. Unlike traditional Graph Convolutional Networks (GCNs) that apply fixed weighting schemes, GAT introduces trainable attention coefficients, allowing the model to adaptively weigh the importance of neighboring nodes during message passing.
[0125] GAT is applied to molecular graphs for the task of bioactivity prediction. Atoms are represented as graph nodes, and covalent bonds as edges. The model computes attention scores between atom pairs within each molecule, enabling it to focus on chemically or functionally relevant substructures. Multiple attention heads are used to capture diverse interaction patterns, and their outputs are either concatenated or averaged to form enriched node embeddings.
[0126] While GAT does not explicitly incorporate 3D geometric information or inter-molecular context (as done in DTIGN or GIGN), its ability to assign dynamic importance to different atomic neighbors makes it a strong baseline for structure-based molecular bioactivity prediction tasks. GAT serves as a 2D graph-based baseline model for comparing the effect of incorporating geometric and docking-based features.EHIGN model
[0127] Explainable Heterogeneous Interaction Graph Neural Network (EHIGN) [Yang et al., 2024] is a graph neural network method for predicting protein-ligand binding affinity from 3D structures, based on an interaction-based inductive bias that aligns with biological binding mechanisms. The method represents protein-ligand complexes as heterogeneous graphs with two node types (ligand and protein atoms) and four edge types corresponding to ligand intramolecular, protein intramolecular, ligand-protein intermolecular, and protein-ligand intermolecular interactions. EHIGN employs four specialized graph convolution networks — two Covalent Interaction Graph Convolution networks (CIGConvs) using sum aggregation for covalent bonds and two Non-covalent Interaction Graph Convolution networks (NIGConvs) using mean aggregation for non-covalent interactions — to independently process each interaction type and prevent covalent interaction information from being overwhelmed by the more numerous non-covalent interactions.
[0128] The model predicts binding affinity as the sum of pairwise atom-atom affinities rather than using a bottleneck architecture, which preserves 3D structural information and enhances interpretability. To mitigate the interaction hidden bias (where predictions could spuriously correlate with the number of atom pairs), EHIGN incorporates a learnable bias correction term that uses a bottleneck architecture with attention-weighted pooling to capture count-related features. Experimental results demonstrate that EHIGN improves performance across multiplebenchmark datasets (PDBbind core sets, CSAR-HiQ, and LIT-PCBA) and provides physically meaningful interpretations consistent with known binding principles, including accurate distance-affinity relationships and the ability to distinguish between different binding poses.SIGN model
[0129] Structure-aware Interactive Graph Neural Networks (SIGN) [Li et al., 2021] is a graph neural network framework for predicting protein-ligand binding affinity that addresses two challenges in structure-based drug discovery: incomplete modeling of 3D spatial structure and the loss of long-range molecular interactions. The method consists of two main components: Polar-Inspired Graph Attention Layers (PGAL) and Pairwise Interactive Pooling (PiPool). PGAL captures comprehensive 3D spatial information by establishing a polar coordinate system for each atom and preserving both distance and angle relationships through alternating node-to-edge and edge-to-node interaction layers. This approach models geometric structures more completely than distance-only methods while remaining invariant to rotation and translation, unlike Cartesian coordinate-based approaches. To recover long-range interaction information lost during graph construction (where only nearby protein atoms are retained), SIGN incorporates PiPool, which performs atomic type-aware pooling on intermolecular edges to generate a global interaction matrix representing protein-ligand contacts across all atomic type pairs. This matrix is trained with a reconstruction loss as an auxiliary self-supervised task, enabling the model to simultaneously learn local spatial geometry and global interaction patterns.MBP model
[0130] Multi-task Bioassay Pre-training (MBP) [Yan et al., 2023] is a pre-training framework designed to improve structure-based protein-ligand binding affinity (PLBA) prediction by leveraging large-scale bioassay data. The method addresses inherent challenges in utilizing extensive but noisy databases such as ChEMBL, including heterogeneous affinity measurement types, experimental variability across different bioassays, and the absence of experimentally determined 3D structures. MBP employs a multi-task learning approach that combines direct affinity regression with bioassay-specific pairwise ranking, enabling the model to learn robust structural representations while mitigating the impact of label noise and inconsistencies.DTIGN dataset details
[0131] The DTIGN dataset comprises 8 protein targets (II, 12, 13, 14, 15, El, E2, and E3), each considered as an individual subset. The original dataset is non-l. I. D. due to the manually constructed test set. To approximate an LI D. setting, the experiments randomly re-split each subset. For each subset, 20% of the data is used as the test set, while the remaining 80% is split into training and validation sets with a 9:1 ratio. The information of each subset is shown in Table 1-3. The number of atoms varies depending on the size of the original binding site for each protein. Table 1 -3 includes information of the DTIGN dataset.SIU dataset details
[0132] The SIU dataset is a recent benchmark for bioactivity prediction, featuring multiple types of bioactivity measurements. The experiments select Kd and Ki as target labels, as they complement the bioactivity types used in the DTIGN dataset. There are two versions of the SIU dataset: SIU-0.9 and SIU-06, where the numberindicates the sequence similarity threshold used to exclude test set samples that are too similar to those in the training set. For the SIU dataset, the experiments adopt the ‘distinct pocket’ defined in the dataset (identified by a single PDB ID) as the union pocket. The information of each subset the experiments use is shown in Table 1-4.
[0133] Specifically, in this study, SIU (Kd) subsets are used for evaluation of the ligand binding affinity (LBA) prediction task in the additional study. The detail is shown in Table 1-5. SIU has a much larger scale than the conventional PDBbind dataset [Wang et al., 2004]; the comparison is shown in Table 1-6.Evaluation on additional baseline models
[0134] The experiments include additional competitive baselines: the explainable heterogeneous interaction graph neural network (EHIGN), the structure-aware interactive graph neural network (SIGN), and the multi-task bioassay pre-training (MBP) approach on some subsets. Results (Table 1-7 and Table 1-8) show that the proposed method consistently improves performance over these models across most metrics (39 / 43) on representative benchmarksubsets.
[0135] Together with the existing GAT, GTGN, and DTIGN, these additional methods form a diverse and competitive baseline set that includes heterogeneous, attention-based, and multitask learning approaches. This broader comparison provides a more comprehensive assessment of the proposed model’ s effectiveness. Table 1-4 includes information on the SIU dataset. Table 1-5 includes information of SIU-0.9 and SIU-0.6 (K<j) datasets. Table 1-6 includes a comparison of SIU and PDBbind. Table 1-7 includes additional model performance on DTIGN subsets.Evaluation on ligand binding affinity prediction task
[0136] Ligand binding affinity (LBA), typically quantified by the dissociation constant (Kd), is included in the SIU dataset and serves as a relevant benchmark for evaluating virtual screening performance on docked conformations. Table 1-8 includes additional model performance on SIU subsets.
[0137] To assess efficacy, the experiments analyzed pK prediction performance on the SIU-0.9 and SIU-0.6 subsets, which are large-scale and structurally diverse datasets, sharing the same test set but differing in training data. As shown in Table 1-9, out of 24 reported metrics (RMSE, MAE, r, Spearman across all models and both subsets), 19 were improved. Table 1-9 includes model performance comparison on SIU-Xd.
[0138] In summary, the results show consistent performance gains: RMSE improved by 11.13%, MAE improved by 11.55%, Pearson correlation improved by 135.48%, Spearman correlation improved by 151.83% averaged across models and datasets, as summarized in Table 10 below. Table 1-10 provides a summary of average improvements on SIU-Kd
[0139] Additionally, the experiments evaluated the Vina scoring function on the SIU-Kd test set (same for both SIU-0.9 and SIU-0.6). For each protein-ligand pair, multiple docked conformations were scored using AutoDock Vina. These scores were converted to predicted dissociation constants using the following standard transformation of Equation (2-1):where AG is the binding free energy reported by Vina (in kcal / mol), R is the gas constant (1.987 × 10⁻³ kcal / mol·K), and T is the temperature (assumed to be the standard state condition, i.e., 298 K). For consistency, the experiments used pKd as the prediction target and averaged the pKd values across all conformations for each pair.
[0140] The models enhanced by embodiment of the invention outperform the Vina baseline across all error and correlation metrics (Table 1-9). These findings reinforce the practical value of embodiment of the invention in affinity prediction and virtual screening workflows, demonstrating improved predictive performance even when relying solely on docked poses.
[0141] Unlike traditional docking methods, which score only a single binding conformation, embodiments of the invention predict experimental affinity by integrating information from multiple plausible docking conformations. This multi-conformation strategy aligns with practices adopted in recent datasets and better reflects real-world scenarios where various binding poses collectively influence the observed affinity.
[0142] Beyond SIU dataset, the experiments also conducted supplementary experiments to validate embodiments of the invention on a redocked PDBbind v2020 datasetonly). For each PDB entry, the top three docking poses were used. The models were trained on the general subset (4,466 PDB entries, excluding all refined-set samples) and evaluated on the refined set (2,783 PDB entries). As a baseline, Vina scores were converted to pKd and averaged over the top three poses. The results, presented in Table 1-11, show that the enhanced methods of embodiments of the invention consistently outperform the baselines across all metrics, demonstrating their generalizability on the PDBbind dataset. Table 1-11 includes performance comparison on PDBbind dataset.Impact of different aggregation strategies
[0143] As indicated in Equation (1-5), the experiments aggregate multiple pocket-ligand pose embeddings to improve the accuracy of bioactivity prediction. In the experiments, models use mean aggregation unless otherwise specified. The only exception is DTIGN, which by default employs its own attention-based aggregation strategy.
[0144] To analyze the effect of different aggregation strategies, the experiments conducted controlled comparisons where both the baseline DTIGN and DTIGN+ embodiments of the invention were equipped with the same aggregation function — either mean, sum, or attention. As in other experiments, this comparison ensures that the observed improvements stem from embodiment of the invention rather than differences in aggregation. The results are shown in Table 1-12.
[0145] It can be noticed that on three DTIGN subsets (I1, I2, E1), embodiments of the invention consistently improve performance across all aggregation strategies, though the extent of improvement varies In all settings, DTIGN+ embodiments of the invention outperform the original DTIGN with the same aggregation function:• With mean aggregation, embodiments of the invention achieve an average RMSE reduction of 9.62%, along with improvements of 26.43% in r and 25.42% in T.• With sum aggregation, improvements are more pronounced: 14.26% reduction in RMSE, and increases of 32.44% in r and 46.54% in T.• With attention-based aggregation, the enhanced modeling achieves the best results — 12.16% lower RMSE, and substantial gains in correlation metrics: 41.54% in r and 52.46% in r.
[0146] These results demonstrate that embodiments of the invention are robust for different aggregation choices, enhancing the baseline model regardless of the specific strategy used. Table 1-12 includes a performance comparison across different aggregation functions.Explanation of the relationship between geometric edge and empty space
[0147] GeoREC captures spatial emptiness implicitly via geometric edge construction. For each atom, the surrounding space is partitioned into K cones of equal solid angle. Each cone connects the central atom to its nearest neighbor, and the distance to the nearest atom in eachcone provides a lower bound on unoccupied volume in that direction.
[0148] Consider: A surrounding sphere of radius r is partitioned into K cones. Then the volume per cone can be represented per Equation (2-2):(2-2)
[0149] If the nearest atom in a cone lies at a distance s, the empty volume in that direction is at least, per Equation(2-3)
[0150] Thus, longer distances imply more emptiness, and this grows with s3. GeoREC encodes the distance to the nearest atom in each cone, providing a compact and physically meaningful proxy for local spatial emptiness.Evaluation metrics
[0151] Detailed definitions of the metrics adopted in this paper are provided below. The predicted and ground truth labels of bioactivity are y'i and yt, respectively; the total number of data points is N.
[0152] Root Mean Squared Error (RMSE) is defined, per below Equation (2-4), as:(2-4)
[0153] Pearson Correlation Coefficient (r) is defined, per below Equation (2-5), as:where ȳ̂ and ȳ are the means of the predicted and ground truth labels, respectivelywhere r ranges from -1 to 1, indicating the degree of linear correlation.
[0154] Kendall’s Tau Correlation Coefficient (r ) is defined, per below Equation (2-6), as:(2-6)where C and D are the numbers of concordant and discordant pairs aiming all 'A N (N - 1) possible pairs (i, j) such that i < j. A pair is concordant if (jg — _ / )(_pz~ »)>0, and discordant if (yt — y / )(yt — v / ) < 0. r measures the ordinal association between two variables, ranging from -1 (complete disagreement) to 1 (complete agreement).
[0155] Mean Absolute Error (MAE) is defined, per Equation (2-7), as:(2-7)MAE measures the average absolute difference between predicted and ground truth values, treating all errors equally regardless of their direction.
[0156] Spearman’s Rank Correlation Coefficient is defined, per Equation (2-8) as:(2-8)where d, = rank(y,) - rank()’;] is the difference between the ranks of the predicted and ground truth values for the / -th data point, p measures the strength and direction of the monotonic relationship between predicted and ground truth values It ranges from - 1 (perfect inverse rankcorrelation) to 1 (perfect rank agreement)Comparison with Uni-Mol
[0157] When compared with the large-scale pretrained model Uni-Mol, which was trained on 209 million 3D molecules and 3 million candidate pockets [Zhou et al., 2023], the enhanced baseline models achieve comparable performance without pretraining. Notably, the enhanced modeling substantially outperforms Uni-Mol in terms of r and Spearman correlation for the Kdlabel, highlighting the effectiveness of the improved pairwise loss in capturing inter-sample correlations.
[0158] Comparison of Uni-Mol with embodiments of the invention regarding the training speed, training device, and model size, shown in Table 1-13. The training speed of Uni-Mol is estimated based on the reported pretraining duration and the number of epochs used for molecular and pocket representation learning, as described in Zhou et al.
[2023] , Embodiments of the invention with GNN, exemplified by DTIGN+ enhanced modeling, exhibit a smaller model size and achieves comparable training speed. This highlights the practicality and efficiency of embodiments of the invention, demonstrating its potential for easy deployment and integration in bioactivity prediction pipelines. Table 1-13 is a comparison with Uni-Mol on the training speed, training device, and model size.Compare with the pairwise loss in other works
[0159] Several related works have adopted ranking or pairwise loss. The pairwise loss complements in embodiments of the invention provide a primary objective by explicitly modeling relative differences between samples. While the core idea of employing pairwise loss is similar, the problem settings differ. The experiments take MBP [Yan et al., 2023] and ActFound [Feng et al., 2024] as two examples. A detailed comparison with MBP and ActFound is provided below.
[0160] In MBP, they use BCE as their ranking loss. Differences between their ranking loss and the improved pairwise loss are included in Table 1-14. Table 1-14 is a comparison of loss functions with MBP.
[0161] In ActFound, the pairwise loss enables meta-learning for few-shot adaptation. E g., ligand bioactivity data from the same assay are inherently comparable, allowing the model to learn relative differences even with limited data. This loss also guides the model to refine predictions on new assays using only a few samples. The improved pairwise loss is a supplementary term in a hybrid loss for a regression task. The detailed comparison is shown in Table 1-15. Table 1-15 is a comparison of loss functions with ActFoundAnalysis of using docking conformations
[0162] Docking methods can generate inaccurate complex conformations, which may impair the performance of models trained on docked structures. Users should be aware of this potential issue and are encouraged to use biologically plausible docking poses obtained from well-established software tools. In the DTIGN dataset, ligand poses were generated using AutoDock Vina [Trott and Olson, 2010], and the top-ranked poses were selected for model training (as stated in the “Docking method” section on Page 2 and the footnote of Table 1 in [Yin et al., 2024]). In the SIU dataset, docking was performed using AutoDock Vina, Glide [Friesner et al., 2004], and GOLD [Verdonk et al., 2003], with a voting strategy applied to select representative poses (as described in the “Structural data construction via multi-software docking” section on Page 5 of [Huang et al., 2025]).
[0163] Nevertheless, predicting affinity from docking conformations offers key advantages overusing co-crystal structures. Co-crystal structures are scarce and costly to obtain, especially for novel targets, whereas docking provides scalable access to approximate protein-ligand complexes. Most training data lacks co-crystal structures, so using docked conformations enables broader applicability. Moreover, deep learning can be designed to denoise and learn meaningful patterns, making the enhanced approach practical and robust for real-world tasks like virtual screening.Trade Offs
[0164] While incorporating global pocket information can increase the number of atom nodes in the GNN, thereby raising computational resource requirements, this increase is relativelymodest and acceptable in practice, as the number of atoms in the Union-Pocket typically ranges from a few hundred to a few thousand. Furthermore, the additional computational time is negligible when compared to the much longer timescales involved in real-world drug discovery processes.Additional Experiments
[0165] Experiments first compare the baselines with embodiments of the invention, which integrates GeoREC, union pocket within the GNN, and the hybrid loss function on the DTIGN dataset. As shown in Table 2-1, use of embodiments of the invention consistently enhances the DTIGN baseline across all subsets and metrics, demonstrating robust improvements on diverse protein targets. On average, embodiments of the invention reduce RMSE by 9.93%, increase r by 18.39%, and improve T by 19.09%. On the GIGN and GAT baselines, embodiments of the invention also achieve noticeable gains on 7 out of 8 and 6 out of 8 datasets, respectively. These results highlight the consistent effectiveness of embodiments of the invention. Table 2-1 also shows that the proposed method achieves more improvement on stronger baselines.
[0166] Furthermore, the experiments evaluate the performance of embodiments of the invention on the SIU dataset, which includes a broader set of protein targets and ligands. The results are presented in Table 2-2, with the Uni-Mol [Zhou et al., 2023] benchmark taken from [Huang et al., 2025], The results demonstrate that enhanced modeling outperforms their respective baselines in 43 out of 48 dataset-metric combinations and achieves state-of-the-art performance. Notably, embodiments of the invention achieve substantial gains in some subsets. For example, in SIU-0.9 / Ki, DTIGN+ the enhanced modeling achieves r = 0.3213 and Spearman correlation of 0.3259. In contrast, the baseline methods struggle to learn meaningful relationships, as reflected by near-zero correlation scores. These results demonstrate the improved capability of embodiments of the invention on more complex datasets, further illustrating its robustness and generalizability.
[0167] To assess the contribution of each individual component in the use of embodiments of the invention, the experiments perform an ablation study by systematically removing one module at a time from the complete configuration on the DTIGN model. The setups are as follows: 1) DTIGN-GUL includes GeoREC (G), union pocket (U), and hybrid loss (L); 2)DTIGN-GU(L) includes GeoREC (G) and union pocket (U), excluding the hybrid loss (L); 3) DTIGN-(G)UL includes the union pocket (U) and hybrid loss (L), excluding GeoREC (G), and 4) DTTGN-G(U)L includes GeoREC (G) and hybrid loss (L), excluding the union pocket (U). The results of the ablation study are shown in Table 2-3. By comparing these ablation settings with the full enhanced method (DTIGN-GUL), it is evident that the geometric feature GeoREC (G) makes a substantial contribution to performance. When comparing DTIGN-(G)UL to DTIGN-GUL, observed is a reduction in RMSE by 9.3%, an increase in r by 14.4%, and an elevation in r by 19%.
[0168] Table 2-1 is a comparison of model performance on baselines and their enhanced versions with the proposed methods across multiple datasets. Table 2-2 is a comparison of model performance on SIU dataset.
[0169] The union pocket also confers consistent advantages in the majority of cases. Compared to DTIGN-G(U)L, which excludes the union pocket (U), use of the union pocket reduces the RMSE by 5.8%, increases r by 7%, and improves T by 9% on average. Notably, GeoREC and union pocket sometimes exhibit a mutually reinforcing relationship; when utilized in isolation, their influence may be restricted. For instance, on dataset El, DTIGN-(G)UL and DTIGN-G(U)L yield r of approximately 0.31 and 0.32, whereas the DTIGN-GUL boosts r to 0.5, marking a nearly 60% improvement. This demonstrates that excluding either GeoREC (G) or union pocket (U) results in substantial performance degradation, whereas their integration leads to a reinforced effect.
[0170] Integrating the hybrid loss (L) enhances the training process, particularly in terms of order consistency. This is evidenced by comparing DTIGN-GUL with DTIGN-GU(L); across 8 datasets, T improves in 7 cases. On average, compared to DTIGN-GU(L), it reduces the RMSE by 3.1%, increases r by 4%, and improves r by 4.6%. This validates the effectiveness of incorporating the pairwise loss in bioactivity prediction.
[0171] Table 2-3 includes data from an ablation study on DTIGN baseline (each variant excludes one component).
[0172] While various aspects and embodiments have been disclosed herein, it will be apparent that various other modifications and adaptations of the invention will be apparent to the person skilled in the art after reading the foregoing disclosure without departing from the spirit and scope of the invention and it is intended that all such modifications and adaptations come within the scope of the appended claims. The various aspects and embodiments disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope and spirit of the invention being indicated by the appended claims.
Claims
CLAIMS1 A computer-implemented method for predicting a bioactivity of a ligand with respect to a target protein, comprising:(a) obtaining structural representations of the target protein and the ligand;(b) obtaining or generating a plurality of candidate binding poses of the ligand relative to the target protein;(c) for each candidate binding pose, defining a corresponding protein binding pocket of the target protein;(d) constructing, for each binding pose and corresponding binding pocket, a proteinligand complex structural representation in which nodes represent ligand atoms and protein pocket atoms and edges represent interactions between the atoms;(e) augmenting the protein-ligand complex structural representation with a geometric representation of spatial emptiness, including:(i) defining a local spherical coordinate system centered at each node,(ii) partitioning space surrounding each node into a plurality of directional cones, and (iii) introducing geometric edges encoding spatial relationships indicative of unoccupied pocket space,(f) providing the augmented protein-ligand complex structural representations as inputs to a machine-learning model; and(g) executing the machine-learning model to generate, for the ligand, a predicted bioactivity value relative to the target protein.2 The method of claim 1, wherein the plurality of candidate binding poses for the ligand is represented as a pose set Xi in accordance with Equation (1-1):where xl,i represents the i-th pose of the ligand l associated with one of the available protein pockets, and d is the dimensionality of the input representation.
3. The method of claim 1, wherein the predicted bioactivity value y / is generated in accordance with Equation (1-2):where f is the machine-learning model, Xl is a set of candidate binding poses of the ligand l, Pl is a set of candidate protein binding pockets associated with the ligand, and ŷl is the predicted bioactivity value for the ligand.
4. The method of claim 1, wherein the predicted bioactivity value / is obtained using a permutation-invariant aggregation in accordance with Equation (1-3):where fpose is a pose-level model applied to a pose representation xl,i and a corresponding pocket pl,i, Agg is a permutation-invariant aggregation function, and Nl is a number of candidate binding poses5. The method of claim 1, wherein a bioactivity prediction for a pose-pocket pair is generated in accordance with Equation (1-4):where f' is a bioactivity prediction model, xl is a representation of the ligand l including at least one binding pose, p is a candidate binding pocket of the target protein, and ŷl is a predicted bioactivity value.
6. The method of claim 1, wherein the predicted bioactivity value / is computed according to:represents a set of binding pose representations of a ligand I, andP’I represents a subset of binding pockets of a same protein associated with the binding poses.7 The method of claim 1, wherein the predicted bioactivity value comprises a continuous quantitative metric selected from the group consisting of IC50, EC50, Ki, and Kd.
8. The method of claim 1, wherein the plurality of candidate binding poses is generated using a molecular docking process or derived from experimentally determined protein-ligand structures.
9. The method of claim 1, wherein each local spherical coordinate system is defined independently for each atom of the protein-ligand complex.
10. The method of claim 1, wherein the geometric edges encode directional proximity information by identifying, for each directional cone, a nearest atom or an absence thereof.
11. The method of claim 1, wherein unoccupied pocket space represents regions of the protein binding pocket not occupied by the ligand in a given binding pose.12 The method of claim 1, wherein the space surrounding each node is partitioned into S directional cones, and S = 32.
13. The method of claim 1, wherein the machine-learning model comprises a graph neural network operating on the augmented protein-ligand complex structural representation.
14. The method of claim 1, wherein modeling unoccupied pocket space enables prediction of ligand bioactivity reflective of competitive inhibition with endogenous substrates.
15. The method of claim 1, further comprising constructing a unified binding pocket representation for the target protein by aggregating atoms from a plurality of protein binding pockets associated with the target protein to form a union pocket, and providing the union pocket together with each augmented protein-ligand complex structural representation as inputs to the machine-learning model.16 The method of claim 15, wherein the union pocket is assembled from a union of atom sets from multiple local binding pockets associated with different ligand poses or different ligands bound to the same target protein.
17. The method of claim 15, wherein the union pocket is used consistently across multiple ligands associated with the target protein to enable comparative evaluation of ligand bioactivity within a shared pocket context.
18. The method of claim 15, wherein the union pocket is defined in accordance with Equation (1-6):where A(p)' denotes a set of atoms of a local binding pocket p', L denotes a set of ligands docked to the target protein, and Pi denotes a set of local binding pockets associated with ligand I.
19. The method of claim 18, wherein the predicted bioactivity value y / is generated using the union pocket in accordance with Equation (1 -7):(1-7) Ul = union V'to Punion Pwhere funionis a machine-learning model configured to accept a pose set Xland a union pocket punion, and ŷlis the predicted bioactivity value for the ligand l.
20. The method of claim 15, wherein the machine-learning model is trained using a loss function that incorporates a pairwise loss term comprising pairwise comparisons between bioactivity values of different ligands associated with the target protein.
21. The method of claim 1, wherein the machine-learning model is trained using a loss function that incorporates a pairwise loss term comprising pairwise comparisons between bioactivity values of different ligands associated with the target protein.
22. The method of claim 21, wherein the loss function includes a pointwise mean-squared error loss L2in accordance with Equation (1-8),where N denotes a number of training samples in a batch, yi denotes a predicted bioactivity value for an Z-th sample, and y? denotes a ground-truth bioactivity value for the Z-th sample.
23. The method of claim 21, wherein the pairwise loss term is computed in accordance with Equation (1-9):based on differences between pairs of samples, where N denotes a number of training samples in a batch, ŷiand ŷjare ground-truth bioactivity values for samples i and j, and ŷiand ŷjare predicted bioactivity values for samples i and j, such that the pairwise loss penalizes a discrepancy between (yi- yj) and (ŷi- ŷj).
24. The method of claim 21, wherein the loss function comprises a hybrid loss Lhybrid in accordance with Equation (1-10):(1-10)where L2 is a pointwise mean-squared error loss, Lpairmse is the pairwise loss term, and; and 2 are weighting coefficients that control relative contributions of 1.2 and Lpairwise.
25. The method of claim 24, wherein the weighting coefficients satisfy i = 0.8 and X2 = 0.2.
26. A system for predicting a bioactivity of a ligand with respect to a target protein, comprising:(a) one or more processors; and(b) one or more memories storing instructions that, when executed by the one or more processors, execute the method of any one of claims 1 to 25.
27. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform the method of any one of claims 1 to 25.
