Ligand-agnostic off-target site prediction for early toxicity screening by leveraging point cloud-based protein cavity analysis

Abstract

Adverse drug reactions caused by molecules binding unintended targets are a major concern in drug discovery. Early identification of such interactions during drug design and development minimizes risks and enhances therapeutic efficacy. While experimental approaches are time-consuming and resource-intensive, in silico virtual screening offers a faster, cost-effective strategy to anticipate off-target effects early in drug design. Here, we present a point cloud-based virtual screening method, designed to perform off-target identification based solely on the chemical composition of the primary drug binding site. By screening multidimensional point clouds representing all potential human binding sites, this approach identifies alternative targets based on shape and physicochemical properties. Notably, it operates independently of the protein’s overall structure or sequence. This focus on the binding site broadens the search space to structurally unrelated proteins and enables screening without requiring lead molecule information. Using an experimentally validated test set, we demonstrated the method’s ability to identify alternative targets across protein families and predict drug promiscuity, achieving a Top-10 recall of 24% for validated, strongly modulated targets. While ligand-agnostic sequence- (BLASTP) and structure-based (Foldseek) methods show higher overall recall, point cloud-based screening uniquely recovers 13.9% (79.4% of its total recovered hits) of known off-targets with <30% sequence identity at Top-100, which are missed by both BLASTP and Foldseek. Beyond known targets, we found several high-ranking candidates not yet annotated as drug–targets but showing notable cavity similarity despite being structurally unrelated, which we present as testable hypotheses for follow-up.

Introduction

The discovery of off-target interactions in drug development is critical for minimizing side effects and identifying new therapeutic uses for existing drugs (i.e. drug repositioning). Despite advances in drug design and development, unexpected side effects often arise from unintended drug–target interactions, which may only become apparent during late-stage clinical trials [1, 2]. Such occurrences pose significant risks to patient safety and the commercial success of new drugs. Indeed, it is estimated that ~30% of drugs fail clinical trials due to safety concerns, with off-target toxicity being one of the leading causes of failure [2]. This underscores the importance of early identification of potential off-target interactions in the drug development process.

Traditionally, off-target interactions are detected through experimental methods, which can be costly and time-consuming, like in vitro polypharmacological screens or in vivo models [3]. Standard toxicity screening assays, which are used to prioritize lead molecules, usually cover fewer than 60 protein targets [4]. While these are selected based on their relevance for known side-effects, they do not extensively cover the whole off-target spectrum possible, which is a contributing factor to drug attrition in late stages of development. Virtual screening methods facilitate early identification of off-target interactions, thereby minimizing the need for costly downstream experiments. Consequently, there has been an immense interest in computational off-target or side-effect prediction tools [5, 6], and many of the available tools for off-target prediction rely on chemical information of the ligand [7–9]. Ligand-based approaches utilize chemical similarity to known ligands of other targets to infer off-target interactions. They are dependent on and constrained by the availability of well-characterized ligands and chemical libraries with predefined ligand-target relationships. Recent advances in in silico structure prediction have revolutionized structural biology, enabling the modeling of protein structures directly from sequences, even in the absence of experimental data [10]. Artificial intelligence (AI)-driven tools such as AlphaFold [11] have significantly enhanced the efficiency of structure prediction, making structural models available for a larger fraction of the sequence space than ever before. This advancement opens new abilities for drug discovery and off-target prediction. Structure- or sequence-based approaches for off-target prediction, e.g. via powerful search engines like Foldseek [12], Basic Local Alignment Search Tool for Proteins (BLASTP) [13], Pfam [14], or PROSITE [15] searches, rely on the sequential or structural similarity of alternative targets to the intended target. Methods focusing on overall protein structures and sequences might miss subtle differences in drug-binding sites critical for accurate off-target predictions and fail to identify targets across structurally diverse proteins [16]. Drug promiscuity, also known as polypharmacology, demonstrates that structurally diverse proteins, beyond those similar to the intended target, can also interact with a given drug. This phenomenon underscores the limitation of relying solely on structural or sequence similarity.

Cavity-based off-target prediction overcomes some constraints of existing methods by operating in a ligand-agnostic manner, relying on binding-site-specific structural and physicochemical features rather than ligand or overall protein similarity. Previously, the most comprehensive structural database of all human proteins was generated within a partnership of Innophore and NVIDIA using up to four different modeling strategies per sequence followed by structural refinement to improve quality [17]. Independent energy minimization and integration of ligands enhanced the representation of dynamics, especially upon ligand binding. Based on this structural dataset together with human protein structures stored in the Protein Data Bank (PDB) [18], Innophore has calculated 849 355 potential binding-sites in the human proteome and represents them as multidimensional, searchable point-clouds, so called Catalophores™, which intrinsically encode up to 19 different physicochemical properties that are projected by the amino acids in protein pockets into their surrounding [19]. Serving as a virtual screening kit, a physics-based comparison of a drug-binding site (primary target) represented as a Catalophore point-cloud with the human “cavitome” results in the identification of chemically similar pockets in the human organism, which might present potential off-targets for drug candidates designed to bind to the primary target. This ligand-agnostic, structure-, and sequence-independent approach broadens the scope of off-target predictions, particularly for recognizing targets across structurally different proteins and for early-stage drug discovery where ligand candidates are not yet fully characterized. By identifying functionally relevant binding sites on the surface of a protein, this method enhances virtual screening capabilities for off-target prediction, which can help identify potential adverse effects and support repositioning strategies of existing drugs. Several studies have documented and validated the utility of Catalophore cavity matchings across various applications. These studies demonstrate that similar binding site cavities across different proteins indicate their potential to bind the same molecule [20] or share enzymatic function [21, 22].

Identifying potential off-target interactions based only on drug–target information allows researchers to detect possible side effects early in drug development. This helps prioritize safer compounds from the start, increasing the chances of bringing low-risk candidates into (pre)clinical studies. By predicting off-targets in silico, researchers can design targeted in vitro screens that focus on these specific proteins, allowing for more informed drug-candidate selection. This approach helps reduce the likelihood of unforeseen toxicities in later stages [23]. Ultimately, early off-target detection decreases the number of drug candidates failing in animal testing due to unexpected side effects, leading to more efficient and ethical drug development processes. Point cloud-based cavity comparison represents a significant advancement in the field of in silico off-target prediction, providing a ligand-agnostic solution addressing limitations of current tools. In this article, we describe its development and validation and illustrate its potential for improving early therapeutic toxicity assessments.

While the underlying Catalophore point-cloud representation and matching algorithm have been described previously [19], the present study introduces and validates its first proteome-wide application to off-target prediction, following the construction of the human cavitome (849 355 Catalophores across 282 325 structures), a systematic validation against 467 experimentally confirmed drug–target pairs, and a direct comparison with sequence-, structure-, and ligand-based chemical similarity approaches. Together, these contributions establish Catalophore-based off-target screening as a validated ligand-agnostic tool for early drug discovery.

Methods

Generation of the human “cavitome” for off-target screening

The human cavitome, which serves as a virtual screening kit for off-target search, consists of multidimensional point-cloud descriptors, called Catalophores, built from protein pockets identified in a dataset of human protein structures. The structural dataset, apart from experimentally solved structures deposited in the PDB, was generated within a cooperation of Innophore and NVIDIA. The human proteome was modeled using four different modeling strategies (homology modeling using CavitOmiX Pro platform with integrated YASARA [24, 25], AlphaFold [11], OpenFold [26], and ESMFold [27]) per protein sequence and subsequent refinement of structures to increase their quality [17]. As part of this refinement, intrinsically disordered regions with low predicted confidence, as indicated by low predicted local distance difference test (pLDDT) scores, were trimmed from the structural models, retaining only well-resolved, folded protein regions for subsequent cavity detection. Independent energy minimization and integration of ligands enhanced the representation of dynamics and refined the conformation of ligand-binding amino acids. In total, this dataset provided 474 736 predicted protein structures in different stages of refinement. For the calculation of Catalophores, we selected a set of 216 907 high-quality structures (Table 1), all together representing 41 630 distinct human protein sequences. Additionally, we include 65 418 human protein structures derived from the PDB. While models generated with AI-assisted modeling algorithms (AlphaFold, OpenFold, and ESMFold) are monomeric structures, the homology models contain 4989 homodimers and 1692 models of higher oligomeric states (up to 24-mers). Nucleic acid molecules are included in 2829 homology models.

Table 1.

Structural models of human proteins used for calculating the human cavitome.

Model type	Structures with ligands	Structures without ligands	Number of Catalophores	Unique proteins represented
Homology	18 818	34 613	171 519	29 928
AlphaFold	14 973	41 688	99 558	25 574
OpenFold	15 070	41 217	98 607	26 586
ESMFold	12 855	37 673	67 613	20 876
Experimental (PDB)	37 049	28 369	412 058	11 721

This table summarizes the number of structural models used to compute the human cavitome, comprising 849 355 Catalophores. Out of 42 042 human proteins considered, 41 630 had at least one structural model with a detectable cavity. Each model type contributes a portion of the protein sequence space included in the human cavitome.

From a total of 282 325 3D structures, 849 355 multidimensional Catalophores have been calculated to represent all protein pockets in the human proteome. They were created using the Catalophore technology [19, 21], employing the LIGSITE algorithm [28] with a cutoff value (enclosure degree) of 5 and a point-cloud grid spacing of 0.375. Cavities of volumes from 30 up to 1500 Å³ were considered. Each point in the 3D point clouds encodes various physicochemical properties, such as hydrophobicity, desolvation, or electrostatic potential, which are functionally relevant. These properties result from the entire protein structure, with amino acids surrounding the pocket having the most influence. However, distal residues or structural elements (e.g. α-helices) may also contribute to the projected properties within the cavity space. The dataset used for off-target screening consists of 849 355 multidimensional Catalophores which represent potential binding sites in human proteins. Protein structure information is excluded from the search, solely focusing on physicochemical conditions inside the empty space of protein pockets.

Preparation of an evaluation dataset

The evaluation set contains experimental modulatory activity data (pAct) for 906 distinct small molecules to validated protein targets. Here, pAct represents the −log₁₀ value of reported IC50 (M), EC50 (M), Ki (M), or Kd (M) values. All known target and protein family information, along with the corresponding pAct (the highest activity value from available sources), was extracted using CLARITY® software, Version 4.0 [29]. The protein families consist of protein kinases (KC), cytochrome P450 (CP), G protein-coupled receptors (GR), transporters (TC), ion channels (IC), nuclear receptors (NR), enzymes (EC), other families (OF), and unclassified (UC). Some targets are classified into multiple target classes in the database. Therefore, each target has been assigned to a single target class in the order specified above.

Query cavities for off-target search were selected by the following criteria: (i) presence of an experimentally solved protein–ligand complex in the PDB, ensuring that the query cavity represents a reliable, ligand-refined binding site conformation, (ii) if several alternative targets, the target which showed the highest pAct, (iii) if several PDB structures were available for the selected target, the query Catalophore was selected by the largest volume of the Catalophore shaped around the respective ligand. In total, off-target screening within the human cavitome was performed with 467 ligand-binding sites. An analysis of various relevant metrics about the dataset (Supplementary Figure S1) is available in the supplementary information.

Cavity-based off-target search

Off-targets were identified by comparing a ligand-binding site (query), represented as a multidimensional 3D Catalophore, against a dataset of 849 355 precomputed Catalophores representing the human cavitome (a collection of protein pockets within the human proteome). The process of point-cloud matching within Innophore’s platform CavitOmiX Pro involves superimposing and adjusting the alignment of the cavities to minimize the matching score through Equation 1 [19].

(1)

Where i runs over all matched point pairs in the overlapping region of the two superimposed Catalophores and score_point is the per-point score quantifying the physicochemical dissimilarity between matched points. Each per-point score is the sum of per-property scores (type), where E denotes the property value, including aromatic carbon (A), nonaromatic carbon (C), electrostatic potential (ele), hydrogen bond donor (HD), nitrogen (N), nitrogen acceptor (NA), oxygen acceptor (OA), and sulfur acceptor (SA), at each grid point. The superposition is optimized using an iterative closest point algorithm that minimizes the score with respect to rigid-body rotation and translation. A score of 0 indicates identical physicochemical properties in the matched region; higher scores reflect increasing dissimilarity. More details about the point-cloud matching are outlined in the “Catalophore Halo analysis” section of the Materials and Methods in Innophore’s earlier publication [30].

The cavity matching results were filtered based on the degree of point-cloud overlap between the query and target cavities, usually requiring at least 50% overlap for ligand-bound cavities, which are reduced in size to fit the ligand-binding site only, and 30% for human cavities, along with a maximum total matching score of 0.1. The overlap percentage reflects the volume of matched points between the query and target cavities and is reported separately for each. Filtering with these overlap criteria has been found to express the highest overall hit-identification rate within the top 100 ranked off-targets, as shown in Supplementary Figure S2. The number of matching hits passing these criteria varies strongly depending on the query cavity used, spanning from 13 to 16 430 hits. Another relevant criterion for filtering matches is the total matching score, which is used for ranking the remaining hits. While we do strongly recommend considering top-ranked hits with higher total scores as well, a matching score of <0.05 typically refers to a high cavity similarity. The number of remaining hits aggregated over all 467 off-target searches when filtering by typical matching score thresholds is displayed in Supplementary Figure S3.

For the analyses presented in this manuscript, only the top-ranked hit per protein was considered for each off-target search by removing duplicate UniProt IDs in cases where multiple structural models existed for the same protein. This filtering reduced the total number of results per off-target search from 849 355 (total number of cavities available for screening) to 41 630 entries (number of unique proteins represented within the human cavitome). The top-ranked 200 predicted targets for each query ligand-binding site, together with pAct values where applicable, are provided in tabular format in the supplementary information (Supplementary Data S1).

Baseline comparison methods

To benchmark our point cloud-based off-target screening against established approaches, three baseline methods were evaluated on the same search space of 41 630 unique UniProt entries. BLASTP [13] (NCBI BLAST+ suite) was run for each of the 178 unique query UniProt sequences against all 41 630 human protein sequences which were considered for building the human cavitome. Results were filtered at E-value ≤0.05 and ranked by bit-score (descending). Sequence identity from BLASTP alignments was used to stratify off-targets into four bins (>50%, 30%–50%, 20%–30%, and <20%). Structural similarity searches were performed using Foldseek [12] (v9.427df8a) against all available protein models (282 325 structures, Table 1). The 467 PDB structures used to calculate the query Catalophores were used as protein structure queries. The search was performed with unlimited E-value threshold (−e inf) and a maximum of 500 000 hits per query. Results were ranked by bit-score (descending). Tanimoto chemical similarity was computed per query ligand from ligand SMILES [31] using Morgan fingerprints (ECFP4; radius 2, 2048-bit) via RDKit [32]. For each cavitome protein, the maximum Tanimoto coefficient (Tc) between the query ligand and all ChEMBL-annotated [33] ligands of that protein was computed. Results were ranked by Tc (descending) and filtered at Tc ≥0.4.

All methods were evaluated using the capped recovery metric. At rank K, the denominator equals the sum of min (known off-targets per query, K) across all queries, ensuring that queries with fewer known off-targets than K do not inflate the recovery rate. Self-hits were excluded and results were de-duplicated per UniProt ID. The top-ranked 200 predicted targets per method for each query ligand-binding site are provided in Supplementary Data S1.

Analysis of binding modes with molecular docking

Selected examples of newly identified potential off-targets were further evaluated by molecular docking of the respective ligands into the binding site identified by CavitOmiX. Molecular docking simulations were carried out using YASARA [25] integrated into the CavitOmiX Pro platform. The docking region was defined with a 2 Å extension around the cavity of interest. The docking protocol utilized VINA [34] as the docking engine with the AMBER03 [35] force field and was executed in five independent runs. The resulting poses were clustered based on a root mean square deviation (RMSD) threshold of 5 Å.

Results and discussion

We report the ability of a protein cavity-centered, point cloud-based approach to correctly identify alternative protein targets (on- or off-targets) for a set of 467 ligands with available experimental activity data. For each ligand, the binding site was represented as a multidimensional Catalophore, a point cloud encoding the geometric and physicochemical properties of the cavity. This Catalophore was calculated from the experimental protein structure associated with the highest reported modulatory activity by the respective ligand. For off-target screening, the query Catalophore was then compared in a ligand-agnostic approach with 849 355 Catalophores representing the human “cavitome” to identify potentially similar binding sites, where ligands designed for targeting the query cavities may additionally bind. Identified targets referring to the same UniProt ID as the query Catalophore were excluded from the following analyses.

Evaluating the success rate of point cloud-based off-target site identification

Within the described test set of 467 ligands with known protein targets, we can determine the success rate of point cloud-based cavity-centered screening to determine labeled off-targets within top-N ranked matching hits (true positives). On the other hand, the false-positive and true-negative rate are not interpretable, as targets that have not been experimentally tested may or may not experience binding or modulatory activity in the presence of the respective ligands. Hence, we assess target recovery rates for different activity classes and identify how well our technology can recover strongly modulated over weakly modulated targets (Fig. 1). Aggregated across the matching results for 467 drug-binding sites, we were able to recover 24.0% of the validated strongly modulated targets (pAct >9) within the top 10 ranked hits, among a total of 41 630 potential human protein targets in the dataset. On the contrary, 1.9% of weakly modulated targets (pAct <6) are found within the top 10 hits. The recovery rate thus correlates positively with the highest reported modulatory activity (pAct) of ligand-target pairs. Averaged across all 467 searches, the top 10 ranked results include 3 already known targets with a documented modulation of pAct >6, while the remaining 7 are unlabeled and have yet to be experimentally characterized for their response to the respective ligand. Given the clearly visible enrichment of strongly modulated targets in the top-ranked hits, the unlabeled proteins among them emerge as high-priority candidates for follow-up. Their consistent appearance alongside known, strongly interacting targets suggests a similar pocket-level compatibility and potential biological relevance.

Figure 1.

Target recovery by activity class; the percentage of experimentally verified alternative targets found within top-N ranked cavity matching results, split into eight activity classes; pAct refers to the highest reported activity value (agonistic or antagonistic activity) in available sources for each ligand-target pair tested; pAct values ≥6 refer to high modulatory activity; the random expectation line shows the fraction of targets to be found randomly in top-N ranked hits among a total of 41 630 human protein targets considered in the screening dataset; recovery rates are aggregated across 467 ligand-binding sites used for off-target screening; identified targets referring to the same UniProt ID as the query target were excluded from the analysis.

We further display protein-family specific target recovery rates and the ability to identify off-targets across different protein families (Fig. 2), comparing the off-target identification rate between positive targets of high activity (pAct ≥6, Fig. 2a and c) with targets where relatively low binding or modulatory activity has been observed (pAct <6, Fig. 2b and d). Across all protein families, 30.7% of the labeled alternative targets with pAct ≥6 were identified within the top 100 matching results, and 23.9% within the top 10 (Fig. 2a). Notably, we observed a significant difference in the hit rates between different protein families. Targets classified as NR comprised the highest identification rate with 59.0% of all labeled alternative targets being identified within the top 100 Catalophore matching results, followed by EC other than KC and CPs, EC (47.9%), GR (38.7%), and KC (28.5%).

Figure 2.

Identification of alternative targets within top-N ranked matching results; (a and b) the percentage of alternative targets correctly identified within the top-N ranked matching results, calculated as the ratio of correctly identified targets to the number of targets that could possibly be found (either the total number of targets for the ligand or N, if N ≤ total number of targets); (c and d) the identification rate of alternative protein families, computed as the ratio of correctly identified families to the number of possible families (whichever is smaller: N or the total number of alternative families); (a and c) include targets with higher (pAct ≥6), and (b and d) include targets with lower modulatory response (pAct <6); the protein families consist of protein KC, CP, GR, TC, IC, NR, EC, OF, and UC; identified targets referring to the same UniProt ID as the query target were excluded from the analysis.

The lowest number of correctly identified targets was observed for IC, TC, OF, and UC proteins (Fig. 2a). For IC and TC, this low success rate is likely due to the poor quality of the structural data used to compute the human cavitome. These protein classes are notoriously difficult to model because of their large size, structural complexity, conformational flexibility, and the frequent presence of membrane-spanning regions [36–38]. As a result, obtaining high-resolution experimental structures is particularly challenging, which also negatively affects the performance of AI-based and homology modeling approaches. Consequently, cavity detection based on these structures may miss important features of the drug-binding site, lowering the overall hit rate in off-target identification. Improving the quality of these structural models through advances in structure prediction and experimental methods could help resolve this issue in the future. The human cavitome is a dynamic resource and will be updated continuously. Therefore, we expect the point cloud-based off-target screening to perform significantly better on challenging protein classes, such as IC and TC, in the future, as structure determination and, consequently, structure prediction methods advance.

An analysis of alternative targets, which showed lower response to the respective ligands (pAct <6), showed lower identification rates across all target classes, with only 9.4% identified within the top 10 matches, and 18.8% within the top 100 matching results (Fig. 2b). Within the top 100 matches, the identification rate decreases significantly compared to strongly modulated targets (Fig. 2a), which further indicates that the cavity-based screening can differentiate between strong and weak targeting of binding sites for all target classes. For targets with pAct <3, where the ligand shows minimal modulatory activity and which can be considered functional false positives, no targets were identified within the top 10, and 5.4% of targets were identified within the top 100 results.

All types of structural models contributed to successful target identification; however, structures deposited in the PDB showed the highest contribution to the hit rate (Supplementary Figure S2B). This likely reflects a bias in the validation setup: the query cavities used for matching (467 ligand-binding sites) were also derived from PDB structures, and 66.5% of the experimentally confirmed off-targets used for validation are likewise represented in the PDB. The sensitivity of finding alternative targets (with pAct ≥6) within the top 100 hits dropped from 30.7% to 16.3% when PDB structures were excluded from the human cavitome. The sensitivity is even higher (33.2%) when only PDB structures are considered. These findings reflect two distinct factors. First, the validation setup is intrinsically biased toward PDB-derived off-targets, as all 467 query cavities are PDB-derived by construction and 66.5% of the confirmed off-targets are likewise represented in the PDB. Recovery rates therefore may overestimate the achievable performance on novel, model-only targets. Second, beyond this validation bias, there is a dependence on structure quality and coverage depth in the validation dataset. PDB structures contribute 48.5% of all Catalophores in the human cavitome while representing only 28% of unique proteins (Table 1), resulting in ~35 Catalophores per protein compared to 3–6 for computationally predicted models. This structural redundancy provides substantially more conformational snapshots per protein and thus more opportunities for a successful cavity match. Despite this, computationally predicted models cover 72% of the human proteome currently not represented in the PDB and remain a valuable and complementary resource. The different modeling algorithms used to generate structural models, and subsequent independent minimizations ensure structural variability, especially of side chain orientations. Homology modeling accounts for the representation of pockets formed at multimer-interfaces or within nucleic acid molecules. The inclusion of ligands into structural models further increases sensitivity in cavity-based searches, as the conformation of the surrounding amino acids gets refined toward ligand binding and cryptic pockets are identified to a greater extent, which have been related to drug-binding before [39]. Notably, Catalophores also encode the influence of cofactors on the environment of binding-sites, such as heme or flavin adenine dinucleotide, which may contribute to the binding of drugs [40, 41].

Comparison with baseline methods

To contextualize the performance of cavity-based off-target prediction, we performed a systematic comparison against BLASTP (sequence-based), Foldseek (structure-based), and Tanimoto chemical similarity (ligand-based). The methods show complementary performance across sequence identity subsets (Fig. 3a). BLASTP and Foldseek achieve relatively high recovery for homologous targets (>50% identity) but decline sharply for distant targets, with BLASTP recovering 0% at <20% identity. Tanimoto similarity achieves the highest overall recovery but is highly biased toward availability of annotated ligand–target pairs in public databases and is applicable only on already existing molecule candidates for a certain query protein. Importantly, Tanimoto-based screening is inherently limited to identifying off-targets for which bioactivity data of chemically similar compounds has already been deposited in public databases, and thus cannot predict novel, previously uncharacterized off-target interactions.

Figure 3.

Comparative assessment of off-target recovery with baseline methods; (a) recovery of experimentally confirmed off-targets (pAct ≥6) as a function of the top-N ranked results, stratified by BLASTP sequence identity between query and target protein; all methods were evaluated on the same dataset (467 query ligand-target pairs); (b) Venn diagrams showing the overlap of recovered off-target pairs among our point cloud-based method, BLASTP, and Foldseek at Top-10 and Top-100 for all off-targets with pAct ≥6 (top row) and nonhomologous targets with <30% sequence identity and pAct ≥6 (bottom row); percentages are relative to the capped ground-truth denominator.

The Top-10 recall for highly modulated targets (pAct >9) is also highest for Tanimoto similarity (49.3%), followed by BLASTP (38.3%) and Foldseek (34.2%), outperforming the point cloud-based screening with a Top-10 recall of 24% (Supplementary Figure S4). However, within ligand-agnostic methods, point cloud-based screening recovers 17.5% of off-targets with <30% sequence identity at Top-100, of which 13.9% (79.4% of its recovered hits) are missed by both BLASTP and Foldseek (Fig. 3b). The union of all three methods recovers 58.4% and 25.4% of all and nonhomologous targets, respectively, at Top-100, confirming substantial complementarity.

Performance in finding targets across different protein families

Our interest was primarily focused on the success of point cloud-based cavity-centered off-target screening in determining alternative targets within different protein families, which means identifying structurally and sequentially distinct proteins, which nevertheless bind the same molecule. Since ligand-binding is determined by the physicochemical properties in the binding site, such targets have a higher chance of comprising similar modulatory activity upon ligand-binding. Our cavity-centered approach successfully determined target relationships across different protein families in 13.0% of cases within the top 100 results, when considering only strongly inhibited targets (pAct ≥6, Fig. 2c). Comparing the success rate of finding alternative targets in different families which show weaker inhibitory responses to ligands (pAct <6, Fig. 2d), we found 10.2% within the top 100 results.

For example, a cavity matching across the human cavitome with the metribolone-binding site correctly identified alternative on- or off-targets without structural or sequence similarity to the intended drug–target (Fig. 4). Out of five targets that were determined to show significant metribolone-induced modulation, three are within the same target class (NR) as the query P10275 (pAct = 9.01), namely P06401, P08235, and P04150 with pAct values of 8.38, 9.27, and 7.66, respectively. The remaining two are structurally diverse proteins: the hormone-binding globulin P04278 (pAct 8.43) and the aldo-keto reductase P42330 (pAct 9.92). Using the Catalophore representing the metribolone-binding site in PDB structure 1XOW (UniProt ID: P10275) as a search query, all three NRs have been found within the top 10 matching hits, ranked first (1ZUC), second (AlphaFold model), and sixth (AlphaFold model). The alternative targets in the OF (despite the absence of overall structural or sequential similarity), share physicochemical properties of their pockets with the query Catalophore cavity, and have been ranked at position 39 and 65. The matching scores were far below 0.05 and the cavities showed significant volumetric overlap to the query. A matching score of ≤0.05 with at least 50% and 30% volumetric overlap of the query and the target Catalophore, respectively, is considered highly similar, indicating these cavities are very likely to bind the same ligand.

Figure 4.

Off-target search results for metribolone (ligand PDB: R18); (a) query Catalophore representing the binding site of metribolone in P10275 (NR), calculated from PDB structure 1XOW; all five labeled targets for metribolone were identified within the top 100 matching results, one from each protein family (hits 1, 39, and 65) are shown; total matching scores and volumetric overlaps of point clouds (query/target) are indicated; Catalophores are annotated according to hydrophobicity; (b) comparison of relevant properties of the binding sites in the query 1XOW and the identified alternative targets; point clouds represent property gradients, including hydrophobicity, desolvation potential, and charge distribution, allowing comparison of binding-site characteristics.

While the homologous steroid receptors (P06401, P04150, P08235; >50% sequence identity) are recovered by all methods (BLASTP, Foldseek, and Tanimoto similarity), P04278 (Hit #39) is recovered by point cloud-based screening and Tanimoto similarity, and P42330 (Hit #65) is found exclusively by point cloud-based screening. Among the top 10 ranked hits identified through the off-target search using the metribolone-binding site, seven targets exhibit high structural similarity around the matched binding region to the query protein P10275, suggesting a strong likelihood of modulation by metribolone. Of these, only three are experimentally verified and included in our validation set (Supplementary Data S1). While such structurally related targets represent expected alternative interactions, the top ranked, structurally unrelated targets, some of which rank higher than validated and structurally related ones, emerge as particularly intriguing off-target candidates. One of them is further evaluated below (Fig. 5), where we show two examples for such structurally unrelated protein targets, which display high binding-site similarity. These examples represent top-ranked Catalophore matching hits, performed with query cavities representing binding sites of metribolone (Fig. 5a) and methscopolamine (Fig. 5b) in their most-effected targets (maximum pAct of all targets): the androgen receptor (NR) and the muscarinic acetylcholine receptor M2 (GR), respectively. Highly similar binding sites have been identified in protein families different from these drug–targets, namely in a cGMP-specific 3′,5′-cyclic phosphodiesterase (EC) and an estrogen receptor beta (NR), comprising high volumetric compatibility and matching scores of 0.024 and 0.017. Neither BLASTP, Foldseek, nor Tanimoto similarity recover these two proteins.

Figure 5.

Potential newly identified, not yet validated alternative targets across different protein families; (a) top-ranked structurally unrelated protein target (UniProt ID: P51160) identified in an off-target search for the metribolone-binding pocket in PDB 1XOW (UniProt ID: P10275); (b) top-ranked structurally unrelated protein target (UniProt ID: Q92731) identified in an off-target search for the methscopolamine-binding pocket in PDB 5ZK8 (UniProt ID: P08172); point-clouds represent property gradients describing hydrophobicity, desolvation potential, and charge distribution; total matching scores and volumetric overlaps of point clouds (query/target) are indicated; Experimentally determined ligand-binding modes are shown in the upper structure views together with the corresponding search cavities, while the lower views show the lowest-energy binding pose determined by molecular docking into the top-ranked binding site from the proteome-wide cavity-matching.

To assess the structural plausibility of these candidates, molecular docking was performed. We emphasize that these candidates are presented as cavity-similarity derived hypotheses for experimental follow-up, not as confirmed off-targets. Molecular docking revealed similar low-energy binding modes of the respective ligands into the cavity of the identified targets compared to their binding in the original experimental PDB structures. Although docking does not confirm binding, the observed similarity in predicted ligand binding poses in alternative target cavities, relative to the native binding pose in the query binding sites, suggest that these regions are structurally compatible with ligand accommodation. AutoDock Vina, the docking method used here, has previously been evaluated [42], showing that the top-scoring predicted binding pose reflects the native binding mode within 2 Å RMSD in ~90% of cases, further supporting the structural plausibility of the predicted binding modes. Experimental binding assays would be required to validate any of these predictions.

As shown in the supplementary information (Supplementary Figure S5), the number of structurally related but not yet experimentally validated targets is substantial, likely contributing to the observed sensitivity of 24% appearing lower than the actual performance of our point cloud-based off-target search approach. In contrast, structurally unrelated high-ranking targets are fewer and largely remain experimentally uncharacterized. One notable example is illustrated in Supplementary Figure S5D and E, where our method predicts heat shock protein HSP90-alpha as a potential off-target for ligands that modulate a serine/threonine kinase. Indeed, the ATP-binding conformations observed in existing structures of HSP90 (PDB ID: 7KLV) and serine/threonine KC (PDB ID: 1XR1) align well with chemical properties in the binding pockets.

Binding-site similarities as predictors of drug promiscuity

To investigate whether ligands that share protein targets can be identified through similarities in the binding cavities of their primary target sites, we evaluated the cavity similarity of ligands that either share or do not share alternative targets. This analysis included a larger dataset containing all 906 ligands and 14 866 associated targets for which significant modulatory activity was reported based on experimental data (pAct ≥6).

We assessed the structural similarity of ligand-binding pockets in experimentally resolved protein structures using point cloud-based cavity analysis. The similarity was quantified using the Catalophore matching score, filtered for cavities with volumetric overlaps of at least 50%. A Catalophore total score ≤0.05 with ≥50% overlap indicates highly similar binding pockets, suggesting that the cavities are likely to bind the same ligand. A score of 0 signifies that the overlapping regions of the cavities are identical. Matches of a cavity to itself were excluded from the analysis. The results showed significantly lower total scores for ligands that share targets, indicating that cavity matching is a reliable method for predicting shared protein targets (Fig. 6 and Table 2).

Figure 6.

Catalophore matching score versus Tanimoto similarity of ligands for distinct and shared targets; distinct targets refer to proteins only inhibited by one specific ligand, while shared targets are targets inhibited by the same ligands; all ligand-target pairs analyzed show significant modulatory effects with pAct values ≥6; (a) Catalophore matching score of ligand binding sites in experimentally solved structures deposited in the PDB; matches have been prefiltered for suitable volumetric overlaps of the matched Catalophores (≥50% each); (b) Tanimoto similarity of ligands was calculated from ligand SMILES [31] with rdkit [32] version 2024.03.5, using the GetMorganFingerprint method.

Table 2.

Catalophore matching score of binding-sites versus Tanimoto similarity of ligands to discriminate between ligands sharing same targets and ligands with unique targets.

	Catalophore matching score		Tanimoto (similarity)
	Shared targets	Distinct targets	Shared targets	Distinct targets
Min–max	0.0–0.32	0.02–0.45	0.0–1.0	0.0–0.85
Mean ± SD	0.05 ± 0.03	0.09 ± 0.05	0.13 ± 0.07	0.10 ± 0.04
25th percentile	0.02	0.06	0.09	0.08
50th percentile	0.04	0.08	0.12	0.10
75th percentile	0.07	0.11	0.16	0.10

In addition to cavity matching, we examined chemical similarity between ligands using the Tanimoto similarity index. This metric ranges from 0 (completely dissimilar) to 1 (highly similar). Our analysis revealed a slight increase in Tanimoto similarity for ligands that share targets (median = 0.12) compared to ligands that do not (median = 0.10). However, the Tanimoto score showed weaker discriminatory power compared to cavity matching, indicating that chemical similarity of ligands is less effective for predicting shared protein targets.

To sum up, cavity comparison outperforms ligand similarity in distinguishing between ligands which share targets and those which do not. This is consistent with findings by Haupt et al. [16], who reported a weak correlation between ligand properties and drug promiscuity, while highlighting stronger correlations with structural and binding site similarities of protein targets.

Conclusion

Point cloud-based comparison of protein pockets offers a ligand-agnostic approach to off-target identification that represents a strong addition to the existing ligand-, structure-, and sequence-based methodologies, improving predictive accuracy for off-target activity. By focusing directly on protein cavities, this method circumvents the limitations of ligand dependency, offering a tool for predicting potential off-targets in the earliest stages of drug development, as soon as the drug–target is known. This helps prioritize safer compounds from the start, increasing the chances that low-risk candidates will advance to (pre)clinical studies.

A key strength of this point-cloud based method lies in its foundation in the human “cavitome,” a dynamic resource that can adapt and grow with scientific advancements. The lack of experimental data on specific protein classes, like TCs and ICs, is a significant limitation for comprehensively and precisely representing all human binding sites. It is essential to keep pace with advances in protein structure prediction and to continuously update the human cavitome with newly available structures to ensure that Catalophore-based off-target searches maintain the highest possible quality.

Our evaluation demonstrates that a binding site-focused, point cloud-based off-target search is a powerful tool for identifying structurally unrelated, and therefore often unexpected, alternative protein targets. Its ligand-agnostic nature makes it particularly valuable in early-stage target and drug discovery, where prior knowledge of drug candidates may be limited or unavailable.

The Catalophore descriptors are computed from static 3D protein structures. Conformational variability is accounted for only implicitly, through multiple independent modeling pipelines per protein sequence, each followed by independent energy minimization, and through the inclusion of multiple PDB entries where available. Ligands present in corresponding crystal structures were introduced into de novo generated models to refine binding-site conformations and capture cryptic pockets. Despite these mitigations, targets undergoing large-scale conformational rearrangement between apo and holo states, targets with transient cryptic pockets, or proteins with extensive hinge-like regions are expected to be under-recovered. The reduced recovery rate for IC and TC (Fig. 2) is consistent with this expectation. In practice, when applying the method to a specific drug–target, multiple query Catalophores derived from molecular dynamics snapshots, holo and apo conformations, and different crystal structures can be used to maximize off-target coverage.

Our current validation exclusively uses PDB-derived query Catalophores, because experimental binding sites provide the most reliable definition of the drug-binding cavity. For novel targets without experimental structures, off-target screening can be performed using computationally predicted models, but prediction quality will depend critically on the accuracy of the modeled binding site and in particular, the correct identification of the binding site location. In cases where no reliable binding site structure is available, ligand-based methods that do not depend on target structure may be better suited, although their applicability is inherently limited to targets for which bioactivity data of chemically similar compounds already exists in public databases.

It must be noted that disordered regions modeled by de novo structural modeling might pass pLDDT-based filtering and result in artificial pockets within the constructed human cavitome, leading to false positives in off-target screening. Nevertheless, false negatives are generally more problematic than false positives, as missing a genuine off-target may lead to relevant safety liabilities; this is why transient pockets in disordered protein regions, which display strong physicochemical similarity to a drug–target site, should not be neglected prematurely. Additionally, our point cloud-based approach does not account for cellular context such as protein expression, tissue localization, or subcellular compartmentalization. Integrating tissue-specific expression data as a postfiltering step represents a straightforward extension to increase the biological relevance of the predicted off-target list.

Looking ahead, the concept of Catalophore-based screening can be extended beyond the human proteome. While this study focuses on human targets, the concept is readily generalizable to any organism for which protein structure data are available or can be generated. This opens the door to applications such as assessing off-target effects in the gut microbiome or in environmentally relevant species, e.g. to investigate unintended interactions of agrochemicals with nontarget EC. While such use cases are beyond the current scope, they highlight the versatility and potential of point cloud-based methods for broader biological and ecological questions.

Key points

• Off-target interactions remain a leading cause of drug failure, yet current in silico prediction methods depend on ligand chemistry or protein sequence and structure similarity, limiting their applicability at the earliest stages of drug discovery and their ability to detect off-targets in sequentially and structurally unrelated protein families.

• We present a ligand-agnostic, point cloud-based method that identifies potential off-target binding sites by comparing physicochemical properties of protein cavities proteome wide, operating independently of ligand information, protein sequence, or overall protein structure, and is applicable from the earliest stages of drug discovery, before lead compounds are fully characterized.

• The human cavitome, comprising 849 355 Catalophore point-cloud descriptors covering 41 630 unique human proteins, provides a comprehensive, searchable representation of all detectable human binding sites for off-target screening.

• Validation against 467 experimentally confirmed drug–target pairs demonstrates a Top-10 recall of 24% for strongly modulated targets, with the method successfully identifying off-targets across structurally unrelated protein families.

• High-ranking unannotated candidates identified alongside confirmed targets represent hypotheses for experimental follow-up, suggesting that the reported Top-10 recall of 24% is a conservative lower bound of the method’s true predictive reach.

Author contributions

Lena Parigger prepared and evaluated data, performed structural analysis, prepared visualizations, and drafted the manuscript with input from all authors. Takafumi Takai prepared and evaluated data, and contributed to interpreting the data. Michael Hetmann contributed to data preparation, analysis, and interpreting the data. Arijit Basu contributed to interpreting the data and supervising the project. Georg Steinkellner and Christian C. Gruber designed, managed, and supervised the project. All authors edited the manuscript to its final form.

Funding

This work was supported by the Innophore GmbH and Takeda Development Center Americas, Inc.

Conflicts of interest

L.P. and M.H. report working for Innophore. G.S. and C.C.G. report being shareholders of Innophore GmbH, an enzyme and drug discovery company. Additionally, G.S. and C.C.G. report being managing directors of Innophore. T.T. and A.B. are employees of Takeda Development Center Americas, Inc. and stockholders in Takeda Pharmaceutical Company Ltd. The research described here is scientifically and financially independent of the efforts of the abovementioned companies and open science.

Data availability

The evaluation data for this manuscript was generated using CLARITY® software, Version 4.0. and is made available in the supplementary information for each of the 467 ligands analyzed, among the top-ranked 200 target predictions and additional supporting data.