IRIS: A Machine Learning-Based Pose Reranking Tool for RNA-Ligand Docking

Abstract

Given their fundamental roles in cellular processes and disease pathogenesis, RNA molecules are promising therapeutic targets. Predicting the 3D structure of RNA-ligand complexes using computational docking is a key element of rational, structure-based inhibitor design. However, RNA-ligand docking remains challenging, due in part to intrinsic properties of RNA such as structural flexibility and a highly charged phosphate backbone. rDock, a widely used RNA docking program, can generate ligand poses close to the experimental structure, but its scoring function frequently fails to rank these poses above less accurate alternatives. To supplement rDock, here we introduce the Intelligent RNA Interaction Scorer (IRIS), a regression model leveraging physicochemical and interaction-based features and trained on the largest data set of experimental nucleic acid–ligand complexes compiled to date for any ML-based tool designed for RNA docking (608 structures). IRIS improves rDock RNA-ligand pose ranking relative to the use of rDock scores alone. Using the best-performing rDock protocol on the RNA portion of the data set, we find that at least one of the 100 top generated poses for any given complex is within 2.0 Å RMSD of the native pose in 86.3% of test complexes. Of these 86.3%, the default rDock scoring function ranks the correct pose first in 42.7% of cases. IRIS improves this latter fraction to 59.8% and increases the success rate for selecting a near-native pose among the top five ranked poses from 64.6% to 78.0%. IRIS thus significantly enhances pose ranking accuracy and can be seamlessly integrated into docking pipelines to rerank ligand poses in RNA-targeted drug discovery.

Introduction

RNA plays diverse roles in cellular regulation and catalysis, functioning as riboswitches, ribozymes, and scaffolds for macromolecular assemblies, and has also been implicated in a wide range of human diseases, including cancers and infectious, immune, cardiovascular, and neurodegenerative disorders. − RNA targeting thus offers a promising route to expand the druggable genome space and can include genes that are noncoding or prove challenging to target at the protein level. RNA molecules are thus attractive targets for drug discovery campaigns, and RNA targeting has gained significant momentum over the past decade. ,

Computational docking is a promising approach to identifying high-affinity ligands for macromolecular targets. Docking programs can predict the preferred 3D orientation and conformation of a ligand bound to a given receptor (referred to as the pose) with an estimated receptor–ligand binding score or affinity. These programs use physics-based, knowledge-based, ML-based, or hybrid scoring functions to identify the most probable binding mode of a ligand to a receptor of interest. The accuracy of the scoring functions, however, remains a primary limitation in molecular docking applications, and especially for RNA systems. − In contrast to protein-targeted drug discovery, which has seen notable success with docking, − RNA-targeted screening presents unique challenges. , RNA molecules contain highly charged backbones with phosphate groups, leading to the participation of coordinating metal ions and water molecules to mediate ligand–receptor interactions and stabilize the binding pocket and tertiary structure folds. − Also, RNA molecules can adopt complex tertiary structures specific to their function (that include, but are not limited to, kissing loops, pseudoknots, G-quadruplexes and coaxial stacking) with molecular driving forces including electrostatic stabilization by water and metal ions and base pair stacking. Further, the structural changes induced by the presence of ligands vary considerably between different RNA systems, from highly specific conformational transitions (on/off state) in certain riboswitches (e.g., SAM-I) while shifting equilibrium toward a given conformation from multiple possible conformations in some RNA systems (e.g., Group-II intron). Capturing the RNA-ligand interactions in highly polar, flexible and low buried surface area RNA binding sites is thus challenging. − Further, generating representative conformations of RNA-binding ligands (including protonation states and tautomers) requires robust approaches as these molecules may exhibit physicochemical properties distinct from protein-binding ligands, including lower octanol–water partition coefficients (LogP), higher topological polar surface areas (TPSA), higher numbers of hydrogen bond donors and acceptors, and higher numbers of heteroatom-containing rings. Finally, another significant challenge in RNA docking is the low number of available experimental PDB structures compared to proteins. ,

Consistent with these data limitations, blind community-wide RNA structure and RNA–complex prediction benchmarks, including the RNA assessments in recent Critical Assessment of Structure Prediction (CASP) rounds, , have shown that accurately capturing the global structures of RNA-ligand complexes and recovering correct RNA–ligand interface contacts remain difficult. For example, in the RNA-Puzzles portion of CASP16, although many submissions closely captured the overall RNA fold, large deviations were frequently observed in aptamer, ribozyme, and riboswitch targets, where correct ligand-binding regions and local interfaces may be poorly reproduced. This difficulty in accurately determining local interactions at the RNA binding site has led to the development of interface-focused evaluation metrics. For example, interface interaction fidelity (I–INF) and interface root-mean-square deviation (I-RMSD) are methods that are specifically designed to assess the accuracy of predicted interfaces in RNA complexes. To our knowledge, I–INF has been primarily applied to evaluating predicted 3D structures of RNA and RNA–protein complexes, whereas I-RMSD has been used to evaluate the accuracy of RNA complexes including ligand–RNA binding interfaces. ,

Comparative assessments have been performed on a wide range of docking programs, including those originally developed for protein–ligand interactions, to evaluate their effectiveness on nucleic acid–ligand complexes. , These studies reveal that performance varies widely depending on the docking protocol and target type, and they underscore the need for tools that perform reliably in RNA-ligand docking scenarios. Among the docking programs evaluated, rDock has emerged as perhaps one of the more accurate programs for RNA-ligand pose generation and one of the few that can perform both local docking (in which the binding site is known) and global docking (in which the binding site is unknown) with reasonable accuracy. rDock has demonstrated one of the most competitive capabilities for predicting near-native conformations given sufficient sampling but arguably suffers from the lack of a scoring function that can consistently rank such poses correctly.

In addition to the standalone docking programs, alternate scoring functions that use knowledge-based, ML-based, or a hybrid of these two have been developed over the years to improve pose prediction accuracy − (e.g., RNAPosers, AnnapuRNA, and SPRank). Several of these scoring functions use rDock to generate the ligand-RNA poses for refinement and reranking. Further, direct scoring methods using data-driven strategies are increasingly being pursued to circumvent expensive docking runs but so far have not seen high success rates and are yet to be tested on a large-scale.

A common challenge historically faced by many RNA specific scoring functions is the lack of solved experimental structures of RNA-ligand complexes, both in terms of quantity and quality to be used for validation. − The development of new, improved scoring functions and programs building upon the existing data is thus a cumbersome process but can be accelerated greatly by ML methods. A recent effective example of an RNA-specific ML scoring function, to which we include a comparison in the present work, is RNAPosers, a set of random forest classifiers developed to improve pose ranking for RNA–ligand complexes using classifiers trained on experimental structures. RNAPosers represents each ligand pose using a high-dimensional “pose fingerprint” that encodes distances between all heavy atoms in the ligand and all nearby heavy atoms in the RNA receptor. Importantly, RNAPosers was trained on the structures of 80 RNA-ligand complexes and tested on two different validation sets composed of 21 and 17 RNA-ligand complexes, respectively. The use of such a restricted training and test set underscores the persistent challenge of data scarcity in ML methods for RNA-ligand pose rescoring. With a leave-one-out training and testing approach, these classifier models recovered poses that are within 2.5 Å of the native poses in ∼60% of the cases on the two separate validation sets.

The performance of ML-based methods is heavily dependent on the quality and quantity of training data. Insufficient or biased training sets can lead to inaccurate predictions and ML models perform best when there is a large, diverse set of high-quality training data available. Selecting the appropriate features to represent the receptor–ligand interactions is also extremely important, and improper feature selection can lead to poor model performance. Most previous models have been trained on interaction data extracted from experimentally solved complexes obtained from the Protein Data Bank (PDB), and have used specific sets of docking software subscores and 2D and 3D physicochemical descriptors of the ligand, the RNA binding site, and interactions between the RNA and ligand. , However, to our knowledge, most current ML-based scoring functions for RNA have not been trained on the full set of available ligand-bound structures and do not include an optimal number of features with respect to their training and validation data sets, or both. Furthermore, although DNA-ligand complexes are highly similar in structure to RNA-ligand complexes, most, if not all current RNA-based ML scoring functions do not include DNA-ligand complexes in their training sets.

To address the limitations present in current ML-based RNA-specific scoring functions, we have designed and tested an ML-based reranking tool for RNA docking that we call the Intelligent RNA Interaction Scorer (IRIS). This method uses a collection of DNA and RNA-ligand complexes in its training set. rDock was used to generate the ligand-RNA poses considering its consistent performance in being able to generate near-native poses as mentioned earlier. IRIS uses selected rDock subscores together with diverse 2D and 3D physicochemical descriptors as features. We compare the results against the rDock default scoring functions for global and local docking as well as RNAPosers. Our results emphasize the potential of approaches such as IRIS as effective predictive tools for RNA-ligand design. IRIS is freely available at https://github.com/AndrewAmburn/IRIS.

Methods

Data Set Compilation

Our initial data set included 1326 structures of nucleic acid (NA)-ligand complexes taken from the Protein Data Bank (PDB RCSB). Given this rather limited number of available NA–ligand complexes, to maximize coverage we did not apply filters based on resolution, experimental method (e.g., X-ray, NMR, cryo-EM), sequence length, or receptor type. All entries containing both a nucleic acid chain and at least one bound ligand were included. A ligand is defined here as any chemical compound with molecular weight <2500 Da that is bound in proximity to the nucleic acid receptor and is not a receptor molecule, water molecule, protein, modified residue, or metal ion. Accordingly, the ligands in the data set are designated as heteroatom (HETATM) entities in the corresponding PDB entry.

The ligands in the data set are not restricted only to drug-like compounds, although these constitute the majority, but also include nondrug-like ligands when they are bound at a designated binding site or are the chemical compound of interest in the corresponding PDB entry. For complexes containing multiple identical copies of the same ligand (e.g., bound at symmetric or repeated sites), a single copy was selected for docking and downstream processing. In such cases, the retained ligand was the first instance listed in the PDB file or the ligand most centrally located within the receptor structure. For complexes with multiple unique ligands bound to a single receptor, each ligand–receptor pair was treated as a separate complex. For each such complex, any cocrystallized ligands other than the desired ligand were removed prior to docking to simulate the docking of ligands to NA receptors in which additional bound ligands are unknown, as is the case in many HTVS screening campaigns. The addition of such complexes expanded the original data set to 1435 structures of NA-ligand complexes.

Protein-Containing Complexes

Although the initial data set compilation step included all PDB entries containing both a nucleic acid chain and a bound ligand, we subsequently removed all complexes in which a fully folded protein was present in the structure. Protein–nucleic acid–ligand complexes were excluded to ensure that the data set reflects ligand recognition and binding-site interactions arising solely from nucleic acid environments. The presence of an intact, folded protein may introduce additional interaction regimes that can alter the effective physicochemical landscape of the nucleic acid binding site, even in cases in which the protein does not form direct contacts with the ligand. Applying this exclusion step resulted in the removal of 69 protein–RNA/DNA–ligand complexes, yielding a new data set of 1366 NA-ligand complexes.

Redundancy Filtering

To prevent data leakage between training and test sets, we applied stringent redundancy filtering based on both ligand identity and receptor sequence similarity. Specifically, a complex was considered redundant only if it shared an identical ligand (as defined by canonical isomeric SMILES) and ≥90% sequence identity in its nucleic acid receptor with another complex in the data set. Receptor sequences were extracted from their respective PDB files by mapping nucleotide residue names to one-letter base codes (A, C, G, U, T). Sequence identity between receptors was computed using global pairwise alignment without gap penalties (globalxx) as implemented in Biopython. Ligand identities were determined using canonical isomeric SMILES, generated with RDKit to ensure consistent encoding of molecular structure and stereochemistry. The redundancy removal process resulted in a total of 840 unique NA–ligand complexes.

This dual-filtering approach ensured that complexes with the same ligands but different receptor contexts, or complexes with the same or similar receptor but a different ligand, were retained, allowing for reduction of bias in testing and training sets and capturing of broader chemical and structural diversity across the receptor–ligand complexes. A complete list of redundant complex pairs and a per-complex summary of redundancy are provided in the Supporting Information.

Feature Set Selection

To characterize each nucleic acid–ligand complex comprehensively, we generated an initial set of 312 features that combined both receptor-, ligand-, and interaction-level descriptors. These included standard 2D and 3D molecular descriptors computed using RDKit, as well as custom features derived using PyMOL, SciPy, and in-house scripts that quantify electrostatic interactions, steric complementarity, spatial geometry, nucleobase composition, and distance-based metrics. Additional pose-specific features were extracted from the rDock docking output files, including energy-based subscores from the rDock scoring function (eq ), which capture distinct terms unique to each docked pose.

To reduce dimensionality and prevent multicollinearity, we eliminated features with zero, missing, or constant values across all training complexes. We then applied pairwise Spearman correlation filtering to remove highly collinear features, defined as having an absolute Spearman correlation coefficient greater than or equal to 0.8. After filtering, a final set of 184 features was retained for model training and testing. The full feature set used in the IRIS model is available in the Supporting Information.

Train/Test Splitting and Similarity Assessment

To construct a distinct and representative test set for model evaluation, we applied a 15% stratified sampling approach over the full set of 840 nonredundant NA–ligand complexes. This approach resulted in a training set of 714 complexes and a test set comprising 126 complexes. For each complex, we first applied our feature generation pipeline to the native ligand pose. Then, following the iSIM method, for each NA–ligand complex, we concatenated the feature set of molecular descriptors column-wise into a unified high-dimensional feature vector. Pairwise similarities were then computed using the Jaccard/Tanimoto (JT) coefficient, defined as

$$\mathrm{J}\mathrm{T}(A,B)=\frac{A·B}{|A{|}^2+|B{|}^2-A·B}$$ 1

where A and B are continuous feature vectors for two complexes, and A·B denotes their dot product. For each complex, we calculated its average JT similarity to all others to obtain a complementarity similarity score (comp_sim). Complexes were then sorted in ascending order of comp_sim, and the sorted list was partitioned into n = ⌊ [0.15] × N⌋ strata. Although the model is designed for RNA–ligand pose reranking, DNA–ligand complexes are included in the training set. Thus, to evaluate whether the model generalizes beyond RNA, we intentionally included approximately 25% DNA complexes in the test set to assess whether IRIS could also improve pose ranking accuracy over rDock for DNA targets. This initial test set comprised 126 complexes, including 95 RNA and 31 DNA complexes, preserving an approximate 75:25 RNA:DNA ratio.

During construction of the initial train/test split, we observed that a subset of 106 ligands were solvent-like molecules or crystallization additives. Such compounds are frequently present in nucleic acid structures to stabilize folding or facilitate crystallization during structure determination. Because these solvent-like ligands differ from most ligands in the data set in both chemical composition and binding behavior, we evaluated their impact on model performance by training random forest models on the training set with solvent-like molecules either included or excluded (Figure S1). The analysis showed that model performance was substantially improved when complexes containing solvent-like ligands were excluded. Based on these results, all complexes containing solvent-like ligands were removed from the data set. This removal resulted in a final data set of 734 complexes, comprising 608 complexes in the training set and 126 complexes in the test set. This resulting set of complexes represents the full data set we used to train and test our IRIS model. To our knowledge, this data set, available in the Supporting Information, is the largest to date for any ML-based tool designed for RNA docking.

The final data set comprises 333 RNA–ligand complexes and 401 DNA–ligand complexes. Structures were determined primarily by X-ray crystallography (533 complexes), with additional contributions from solution NMR (195 complexes), and cryo-electron microscopy (6 complexes). For structures determined by either X-ray crystallography or cryo-EM (539 complexes), reported experimental resolutions range from 0.83 to 6.09 Å, with a mean resolution of 2.21 Å (standard deviation = 0.63 Å) and a median of 2.20 Å. The nucleic acid receptors in the data set span a wide range of sizes, reflecting the diversity of RNA and DNA binding environments represented. Across all complexes, receptor lengths range from 4 to 2883 nucleotides, with a mean length of 72.4 nucleotides (standard deviation = 273.3 nucleotides) and a median of 24 nucleotides.

Although some chemical similarity among ligands is expected within the training set due to limited chemical diversity in publicly available NA-ligand complexes, we sought to ensure that our model did not contain large clusters of ligand chemotypes that might bias model training. To assess this effect, we clustered all training ligands based on Tanimoto similarity (≥0.8) using Morgan fingerprints (radius = 2; 2048 bits) from the RDKit library. Morgan fingerprints encode molecular structures by capturing local atomic environments within a defined radius, making them well-suited for quantifying molecular similarity. We then calculated the fractional representation of each ligand chemotype cluster relative to the full training set. The most frequent cluster accounted for only 2.1% of the total training data. This cluster was dominated by aminoglycoside antibiotics, including neomycin, paromomycin, and ribostamycin, which are well-known to bind RNA through polycationic scaffolds that interact favorably with the negatively charged nucleic acid backbone. , The top 10 ligand clusters collectively represented just 12.6% (Table S1) of the total training set. These results confirm that the model is likely not biased toward any dominant scaffold or ligand chemotype and that the training set exhibits substantial chemical diversity. A full description of each cluster and its constituent complexes is presented in the Supporting Information.

To ensure that the test set was structurally independent from the training set, we performed a ligand-level similarity analysis. First, a binary Tanimoto similarity matrix was generated using Morgan fingerprints from the RDKit library. Ligand structures obtained from Structure Data Files (SDF) associated with test set complexes were first converted into RDKit Mol objects to ensure proper molecular parsing. Each ligand was then transformed into a 2048 bit Morgan fingerprint using a radius of 2, which encodes atomic connectivity up to two bonds away from each atom. This fingerprinting method captures both local and extended molecular features. To assess pairwise molecular similarity, the pairwise binary Tanimoto similarity coefficient (T) was computed for all ligand pairs in the test set and is defined as

$$T=\frac{c}{a+b-c}$$ 2

where a is the number of bits set to 1 in fingerprint A (training set), b is the number of bits set to 1 in fingerprint B (test set), and c is the number of bits set to 1 in both fingerprints A and B (the intersection of the two fingerprints). This metric quantifies the proportion of shared fingerprint features, with values ranging from 0 (no shared substructures) to 1 (identical molecular fingerprints). Importantly, the binary Tanimoto coefficient differs from the generalized Jaccard/Tanimoto formulation used earlier for continuous, high-dimensional descriptor vectors by being strictly defined on binary feature representations of molecular structure.

For each high-similarity pair, we computed the root-mean-square deviation (RMSD) between their maximum common substructures (MCS) using RDKit, without any structural alignment or minimization. Structural alignment was deliberately omitted to evaluate the absolute spatial positioning of the ligands within the global coordinate system of the complexes. This analysis revealed no test ligands with an MCS-RMSD ≤5 Å to any training ligand, indicating the absence of near-duplicate ligand placements or orientations between the training and test sets. For each test ligand, the training ligand yielding the lowest MCS-RMSD among those exceeding a Tanimoto similarity of 0.8 is reported in the Supporting Information. The final test set consists of 126 NA–ligand complexes, comprising 95 RNA and 31 DNA complexes.

To assess the chemical diversity of the selected test set, we applied t-distributed stochastic neighbor embedding (t-SNE) to project the complete feature set fingerprint matrix into two dimensions for each complex in the data set (Figure ). This nonlinear dimensionality reduction method transforms high-dimensional data into a lower-dimensional representation by preserving local similarities and minimizing the Kullback–Leibler divergence between the high- and low-dimensional distributions. This objective function is represented as

$$\mathrm{K}\mathrm{L}(P\parallel Q)=\sum _{i\ne j}{P}_{ij}\log \left(\frac{P_{\{ij\}}}{Q_{\{ij\}}}\right)$$ 3

where, P _ij and Q _ij denote the pairwise similarity probabilities in the high- and low-dimensional spaces, respectively. The resulting projection coordinates, labeled t-SNE1 and t-SNE2, are abstract and do not correspond to original input features, but they reflect the local structure of the feature space. The test set compounds (shown in blue) are broadly distributed across the full t-SNE embedding of the data set (shown in gray), indicating that the stratified sampling strategy effectively captured a representative range of chemical diversity within the NA–ligand interaction space.

1.

Distribution of stratified test set complexes in the t-distributed stochastic neighbor embedding (t-SNE) of the IRIS data set. High-dimensional feature vectors for all 734 nucleic acid–ligand complexes were reduced to two dimensions using t-SNE. Gray points represent the full data set, and blue points indicate the 126 held-out test complexes selected via complementarity-based stratified sampling using the Jaccard/Tanimoto similarity coefficient. Test set points are broadly distributed across the full feature space, confirming that the sampling strategy captured representative regions of the data set.

We then extracted any test set ligand complexes that exhibited a pairwise binary Tanimoto coefficient of 0.8 or higher (Figure S2). To evaluate whether these high-similarity pairs also exhibited spatial redundancy, we computed the MCS-RMSD without alignment using RDKit (Table S2). Despite Tanimoto values of 0.8–1.0, all high-similarity ligand pairs had MCS-RMSD values greater than 30 Å. These results indicate that even when chemical similarity is high, the selected test ligands occupy distinct spatial orientations and binding environments within the 3D coordinate space of the complexes. These findings validate that the test set is free of internal structural redundancy and provides a basis for evaluating pose reranking performance across a range of RNA- and DNA-ligand complexes.

Ligand Protonation and Tautomer Assignment

Assigning correct protonation and tautomeric states to ligands in nucleic acid complexes remains a fundamentally unresolved challenge and is highly nontrivial for even a single NA-ligand complex. Considerations for nucleic acid environments are complicated by intense electrostatics, such as the presence of metal ions that may or may not coordinate ligand binding or stabilize conformational states. In addition, structural data for NA–ligand complexes typically originate from crystallography, cryo-electron microscopy, or nuclear magnetic resonance (NMR) spectroscopy, where ligand confirmation, protonation and tautomeric states can be influenced by experimental conditions, such as nonphysiological pH, ionic strength, magnetic fields, and cryocooling rather than reflecting their native cellular state. −

Given these ambiguities, we adopted a standardized and widely used approach to hydrogen assignment by applying Open Babel’s command-line protonation routine (obabel -h) to all ligand and receptor structures. This tool adds explicit hydrogens so that each atom reaches its typical bonding capacity (“valence”) based on its element type and bonding geometry (“hybridization”), independent of explicit pH or local electrostatics. Notably, this same method is also used by other RNA-targeted machine learning models, including AnnapuRNA, and enables consistent preprocessing across a large and chemically diverse data set.

To maintain consistency with the experimentally determined structures, all ligands were retained in their original stereochemical and tautomeric forms as reported in the corresponding PDB entries. This approach avoids introducing artificial deviations from the native binding pose and aligns with practices used in prior RNA–ligand docking studies.

Treatment of Structural Waters, Ions, and Cofactors in Docking

In our docking workflow, structural water molecules, ions, and bound cofactors present in the crystal structures were removed prior to pose generation. Although RNA-ligand binding is heavily influenced by metal- and water-mediated interactions, removal of these elements during docking is necessary and standard protocol in RNA-targeted virtual screening due to limitations in representation of these interactions by docking software. Specifically, rDock permits inclusion of these components but by default assigns metals a uniform formal charge of +1 for example, which does not accurately capture divalent metal interactions. In testing, the addition of explicit water molecules greatly increases computational expense for all complexes. Therefore, these components may result in vastly increased compute time and poor docking performance if included.

Many recent RNA-ligand docking studies follow this practice by discarding ions, cofactors, and structural waters to simplify the docking system ,,,, and thereby avoid distortions due to poorly modeled electrostatics or poorly placed explicit solvent molecules. Importantly, although these structural elements are omitted during docking, they are reintroduced during feature extraction for the IRIS model if they are known. Specifically, features that describe these interactions, such as the minimum distance from any atom in the ligand to any metal cation or water molecule, or minimum distance from the receptor to metal cations, are computed based on the coordinates of these structural elements in the experimental structure. This approach ensures that the IRIS model incorporates the spatial context and potential interactions provided by these biologically relevant components if they are available. Although our current approach follows established conventions, establishing a benchmark of the impact of retaining versus removing these structural elements during docking across diverse NA-ligand complexes would be beneficial for future studies to investigate.

rDock Docking Methods

Ligand poses in rDock are generated through a multistage search process that combines stochastic and deterministic methods to explore the binding site. For each ligand pose, this process begins with a series of three sequential genetic algorithm stages designed to diversify the search and prevent early convergence on suboptimal solutions. After the genetic algorithm stages converge on a single pose, pose refinement continues with a low-temperature Monte Carlo simulation, which introduces controlled perturbations to escape shallow local minima while preserving favorable interactions. This step is followed by simplex minimization, which locally optimizes the pose by adjusting atomic coordinates to improve the total score.

rDock provides two methods for molecular docking: the reference ligand (RL) method and the two sphere (TS) method. In the RL method, the binding site cavity is defined based on the coordinates of a known ligand from an experimental structure. A grid is placed around the ligand within a user-defined radius. Grid points that overlap with receptor atoms or fall outside the defined region are excluded. A small spherical probe (1.0–2.0 Å radius) is used to refine the cavity by excluding areas inaccessible to the probe, ensuring that only the search space surrounding the experimental ligand is considered during docking. This method is suitable for redocking studies and cases in which the binding site is well characterized (i.e., “local docking”).

The TS method identifies potential binding cavities by placing a grid over a search space of a user-defined radius centered on specified coordinates within the receptor. The method uses two probe spheres of different user-defined sizesa large sphere (3.5–6.0 Å radius) to exclude shallow, flat, or convex regions, and a small sphere (1.0–2.0 Å) to identify deep and accessible pockets suitable for ligand binding. Grid points accessible only to the small probe are retained, and user-defined thresholds such as minimum cavity volume (e.g., 100–300 ų) control the size and shape of the cavity. This method is useful for “global” docking, i.e. when no prior knowledge of the binding site is available.

For both the RL and TS docking methods, in the present study 100 poses per complex were generated by invoking the rbdock engine (rDock command line program for docking) with the -n 100 option. This approach performs 100 independent docking runs per ligand, in which each run begins with a randomized initial conformation within the identified binding site and optimizes the position, orientation, and torsional angles of the ligand using a genetic algorithm (GA). The GA iteratively evolves a population of candidate poses based on docking scores and selects the best-scoring pose from each run. As a result, 100 distinct docked poses are produced for each ligand–receptor complex, each representing a local minimum in S _total identified from a different randomized starting point. Importantly, because each of the 100 poses generated per complex by rDock is accompanied by a unique docking score, each of the 100 poses per complex differ in conformation to any other sampled pose. Detailed docking parameters for both RL and TS docking methods are given in the “rDock Docking Parameters” section of the Supporting Information.

rDock Scoring Function Terms

The standard rDock scoring function is a weighted sum of ligand and receptor intra- and intermolecular interaction terms and is given by eq . S _inter scores the intermolecular interactions between the ligand and receptor, S _intra accounts for the relative intramolecular strain energy of the ligand, S _site estimates the conformational flexibility of terminal OH and NH₃ groups on the receptor within the binding site, and S _restraint includes pharmacophoric restraint and cavity penalty terms by default and can also be used to bias poses based on prior knowledge.

$$S_{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l}}=S_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}}+S_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{a}}+S_{\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}}+S_{\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{i}\mathrm{n}\mathrm{t}}$$ 4

rDock uses two scoring functions, namely, dock (eq ) and dock_solv. The latter scoring function includes a desolvation term (S _solv) added to eq to represent the change in solvation energy of the ligand and the receptor docking site upon ligand binding.

IRIS Model Training Workflow

The IRIS training workflow (Figure ) was initiated with ligand-RNA docking using rDock, with preprocessed ligand SDF and corresponding receptor Mol2 files used as inputs to generate 100 poses per complex ranked by the default rDock scoring functions. Feature extraction was then performed to compute both ligand-level and pose-specific descriptors from the docking output, capturing chemical, geometric, and energetic properties of each ligand–receptor interaction. Ground truth labels were generated by calculating the RMSD of each ligand pose relative to its experimental pose. Symmetry-corrected heavy-atom RMSDs were computed with the Python package spyrmsd, which accounts for atomic equivalence and molecular symmetry. Docking poses were compared to the corresponding native poses after receptor alignment but without ligand superposition.

2.

IRIS training workflow. Illustration of initial IRIS model generation and pose reranking validation on the test set, including rDock runs to generate ligand poses, feature extraction and RMSD calculation relative to the native (input) ligand pose. The model is then trained to learn the relationship between pose-specific features and the corresponding RMSD of each docked pose. After training, we perform feature correlation analysis, model selection, and hyperparameter tuning. The final trained model predicts RMSD values for unseen ligand poses and reranks them based on lowest predicted RMSD, generating the final IRIS pose ranking. The different color shades used in the workflow diagram are used solely for visual clarity and do not encode any methodological or semantic distinction between steps.

The labeled features were then used to train IRIS to predict RMSD values for each pose and then rerank the poses by their predicted RMSD. Finally, the trained model was applied to predict the RMSD for unseen poses, allowing for reranking of ligand poses based on their predicted RMSD, producing the final ranking. Throughout this work, pose-ranking performance is evaluated based on the identification of “near-native” poses, defined as those with an RMSD <2.0 Å relative to the experimental structure.

The mean, median, and standard deviation of RMSD values for each rDock scoring function and search method combination were calculated for all NA-ligand complexes. The standard deviation (σ) is defined here as

$$\sigma =\sqrt{\frac{1}{N}\sum _{i=1}^N{(x_i-\mathrm{x̅})}^2}$$ 5

where N is the total number of RMSD values, x _i is the RMSD value for ligand pose i, and $\stackrel{\circledR }{x}$ is the mean RMSD across ligand poses.

In addition to conventional heavy-atom RMSD, we also computed the interface RMSD (I-RMSD) of each pose in the test set. In this formulation, I-RMSD is calculated over the subset of ligand heavy atoms whose minimum distance to any receptor heavy atom is within 5.0 Å in the experimental pose. The 5.0 Å cutoff is consistent with prior interface-based definitions used in docking and refinement studies. RMSD was then computed exclusively over this interaction-defined atom set.

Feature Importance

To determine the contributions of individual features to the predictions of the IRIS models, we performed a SHAP (SHapley Additive exPlanations) analysis. SHAP values quantify the contribution of each feature by distributing the difference between the actual prediction and the average prediction across all input features. The SHAP value (ϕ _i ) for a given feature i is defined as

$${\varphi }_i=\sum _{S\subseteq N\backslash \{i\}}\frac{|S|!(|N|-|S|-1)!}{|N|!}(f(S\cup \{i\})-f(S))$$ 6

where S is a subset of all features excluding i, f (S) is the model output when using the feature subset S, N is the total number of features, and f (S ∪ {i}) is the model output when adding feature i to subset S. The term $\frac{|S|!(|N|-|S|-1)!}{|N|!}$ represents the weighting factor used to distribute feature contributions fairly across all possible subsets.

Distribution of Errors

We calculated the distribution of errors between predicted and actual RMSD values across the test set. The bias, B, of the model is defined as

$$B=\frac{1}{N}\sum _{i=1}^N(\widehat{y_i}-y_i)$$ 7

where N is the number of ligand poses, $\widehat{y_i}$ is the predicted RMSD for a given pose, and y _i is the actual RMSD for the same pose. Bias quantifies the average difference between predicted and actual values, indicating whether the model systematically overestimates or underestimates RMSD across test set complexes. A bias close to 0 suggests minimal systematic error. The variance (σ²) of the prediction errors is given by

$${\sigma }^2=\frac{1}{N}\sum _{i=1}^N{(\widehat{y_i}-\mathrm{y̅})}^2$$ 8

where y̅ is the mean predicted RMSD of the complex. Variance measures the spread of individual prediction errors, with higher values indicating greater inconsistency in the ability of the model to predict RMSDs for different complexes.

RNAPosers Comparison

We evaluated the performance of RNAPosers in reproducing near-native poses for all the NA-ligand complexes used in this work. The publicly available code for RNAPosers in the master version at https://github.com/atfrank/RNAPosers was used. Receptor Mol2 files, ligand SDF files, and the rDock output file containing ligand poses were given as inputs to the program.

For each ligand atom in the corresponding SDF file, RNAPosers generates a vector (or fingerprint) by calculating the distances of the ligand atom to all RNA atoms within 20 Å. These distances are then grouped by ligand–RNA atom pair types. The final fingerprint for each pose reflects the overall spatial pattern of ligand–RNA interactions. RNAPosers also includes scoring terms from rDock, with the final model using both the fingerprints and rDock terms as input features.

Additionally, we attempted to compare IRIS rescoring with AnnapuRNA and SPRank, but only fractions of the test set could be rescored with these methods due to input format compatibility issues. Thus, to avoid fractional test set comparison, only RNAPosers was used.

Results and Discussion

We initially assessed the performance of the rDock scoring functions dock and dock_solv (see Methods) on the constructed data set of 1435 NA-ligand complexes (Table ). Both scoring functions were tested using both the RL and TS docking methods.

1. Comparison of the rDock Default Scoring Functions (dock and dock_solv) and Search Methods (Two Sphere (TS) and Reference Ligand (RL)) on Pose Prediction Accuracy .

	method	mean	median	stdev	% poses <2.0 Å RMSD
Top Ranked
0	RL_dock	4.4	3.6	3.3	29.0
1	RL_dock_solv	4.3	3.3	3.4	32.1
2	TS_dock	7.2	7.3	4.1	15.0
3	TS_dock_solv	7.1	7.3	4.2	16.4
Best of Top 5
0	RL_dock	3.1	2.5	2.5	40.8
1	RL_dock_solv	2.9	2.3	2.5	44.6
2	TS_dock	5.9	5.9	3.9	22.0
3	TS_dock_solv	5.9	5.9	4.0	22.3
Best of 100
0	RL_dock	1.5	1.3	0.9	77.0
1	RL_dock_solv	1.5	1.2	1.0	76.4
2	TS_dock	4.1	3.1	3.4	37.2
3	TS_dock_solv	4.1	3.0	3.4	37.8

^aAll mean, median and standard deviation values are in Å.

Performance was measured by first determining the mean, median, and standard deviation of RMSDs of three groups.

• (1)the top-ranked ligand pose per complex (“Top Ranked”),
• (2)the lowest RMSD pose among the top 5 ranked ligand poses per complex (“Best of Top 5”), and
• (3)the lowest RMSD ligand pose found among all 100 generated poses per complex (“Best of 100”).

Additionally, we report the percentage of poses achieving RMSD values below 2.0 Å for each group, which is a common threshold used in docking studies for acceptable prediction accuracy, , often called the success rate. RMSD distributions across these criteria are presented in Figure S3.

The results are consistent with previous analyses of rDock pose prediction accuracy. , As expected, across all metrics the RL method consistently outperforms the TS method. For the Top Ranked pose, RL_dock_solv achieves the highest accuracy, with 32.1% of poses under 2.0 Å RMSD. When considering the Best of Top 5, the RL methods again demonstrated superior performance, RL_dock_solv, for example, increasing the percentage of poses below 2.0 Å to 44.6%. For the Best of 100 across all complexes, both RL_dock and RL_dock_solv identified poses below 2.0 Å in ∼77% of cases, with RL_dock performing slightly better than RL_dock_solv.

Following the above comparative assessment of the rDock native scoring functions, we aimed to improve pose ranking accuracy using ML. All further analysis focuses on improving the pose ranking accuracy of the RL_dock method, as it provided the highest number of complexes of the 100 generated with at least one pose below 2.0 Å RMSD and thus presents the greatest potential for successful reranking.

For the machine learning analysis, the data set was split into training and test sets as detailed in the Methods section. In 84.9% of the test set complexes (i.e., 107 out of 126), RL_dock was able to generate at least one pose among the 100 poses per complex with an RMSD below 2.0 Å. (This result differs from the percentage reported for the full data set in Table , which includes both training and test complexes.) The challenge for ML, then, is to develop a method that can effectively rerank poses within these 107 complexes. Accordingly, all subsequent success rates were normalized to this subset, such that selecting the lowest-RMSD pose among the 100 generated poses for each complex corresponds to 100% success.

A comprehensive set of features was generated and calculated for each of the 100 poses per complex as described in Methods. These features served as inputs for regression models aimed at predicting the RMSD of each pose relative to the native pose. The goal was to rank ligand poses based on predicted RMSD, with lower predicted RMSD values indicating better poses. To systematically evaluate the performance of a wide variety of different regression algorithms for RNA–ligand pose reranking, we implemented a benchmarking framework using K-fold cross-validation applied to the training set. We assembled a model zoo consisting of 13 widely used regressors, selected to represent diverse modeling strategies including linear, ensemble, kernel-based, and tree-based methods. The models in the zoo included Ridge, Lasso, ElasticNet, Bayesian Ridge, Support Vector Regression (SVR), LinearSVR, K-Nearest Neighbors (KNN), Decision Tree, Random Forest, Gradient Boosting, AdaBoost, Bagging, and XGBoost. Default scikit-learn hyperparameters were used for all models in the model zoo.

In addition to the 13 models in the model zoo, we included ordinary least-squares (OLS) linear regression as a baseline prediction model. OLS linear regression is a well-established baseline in model benchmarking, as it captures only linear relationships between the input and output variables. This feature makes OLS particularly useful for performance comparisons because improvements over OLS can be directly attributed to the ability of a model to learn more complex decision boundaries beyond simple linear dependence.

To ensure robust performance estimates, we performed K-fold cross-validation using five folds on the training set only, leaving the held-out test set out of all splits to avoid information leakage. For each model, we recorded the mean absolute error (MAE) and mean squared error (MSE) across all five folds, along with their standard deviations. The model with the lowest MAE was designated as the best model in the model zoo. These loss functions are defined as

$$\mathrm{M}\mathrm{A}\mathrm{E}=\frac{1}{N}\sum _{i=1}^N|y_i-\widehat{y_i}|$$ 9

$$\mathbf{M}\mathbf{S}\mathbf{E}=\frac{\mathbf{1}}{\mathit{N}}\sum _{\mathit{i}=\mathbf{1}}^{\mathit{N}}{({\mathit{y}}_{\mathit{i}}-\widehat{{\mathit{y}}_{\mathit{i}}})}_2$$ 10

where N is the total number of ligand poses, y _i is the actual heavy atom RMSD of pose i relative to the experimental structure, $\widehat{y_i}$ is the predicted RMSD for pose i, and $\widehat{y_i}$ is the mean of actual RMSD values. MAE quantifies the average absolute difference between predicted RMSD of each pose and actual RMSD of each pose across all 100 poses per complex, with lower values indicating better performance, while MSE provides a squared penalty for large deviations, making it more sensitive to outlier predictions.

This uniform evaluation strategy enabled direct comparison across models and informed selection of the model to continue toward hyperparameter tuning. The MAE and MSE of the model zoo across all tested models are presented in Table S3.

Based on the results from the model zoo, the Random Forest Regressor was selected as the IRIS regression model. We then conducted hyperparameter tuning on the Random Forest model using K-fold cross-validation with fivefolds. A randomized search was used to optimize model hyperparameters across the hyperparameter space presented in Table S4, with the goal of minimizing the MAE across folds. The best-performing configuration was selected based on cross-validated MAE, and the final IRIS model was retrained on the full training set using these parameters. For completeness, we also calculated the MSE to characterize the distribution of residuals.

Next, we compared the ability of the IRIS model to identify near-native poses on the test set with that of the rDock native scoring function. As a negative control for pose ranking, we first evaluated a random pose-selection baseline. For this baseline, poses were sampled uniformly at random from the 100 poses generated by rDock for each complex in the test set, without using any docking score or learned model. For top-1 evaluation, a single pose per complex was randomly selected, while for top-5 evaluation five poses per complex were randomly sampled. This random sampling procedure was repeated over 1000 independent trials for each complex, and the resulting mean success rates provide an estimate of chance-level performance for pose selection.

Additionally, to determine whether a more complex model such as IRIS is necessary for improved pose ranking, we also evaluated whether the internal scoring terms computed by rDock alone are sufficient to predict pose quality. Specifically, we trained a linear regression model to predict RMSD using only the original rDock score components. This model serves as a reweighted variant of the rDock scoring function, providing an optimized linear combination of terms already available during docking. We further benchmarked IRIS against RNAPosers, a recently published RNA-specific pose reranking method. All five modelsthe default rDock scoring function, the random sampling method, the reweighted rDock model, RNAPosers, and IRISwere evaluated on the same test set (Figures –).

3.

Normalized pose-ranking performance on the full test set for IRIS (yellow), RNAPosers (blue), reweighted rDock scoring function (red), the standard rDock default scoring function (teal), and a random pose-selection baseline (gray). Performance is reported as the percentage of test set complexes for which the Top Ranked pose (ranked #1 by each method) or the Best of Top 5 poses (lowest RMSD among the top five ranked poses) achieved an RMSD <2.0 Å. Random pose selection corresponds to the mean success rate over 1000 independent trials per complex, with error bars indicating one standard deviation. All values are normalized to the maximum achievable success rate (84.9%), defined as the fraction of test set complexes for which rDock generated at least one near-native pose among the 100 sampled.

5.

Normalized performance of IRIS reranking (yellow), RNAPosers (blue), reweighted rDock scoring function (coral), the rDock default scoring function (teal), and random pose selection (gray) on the DNA subset of the test set. Performance was evaluated as the percentage of DNA test set complexes in which the Top Ranked pose (ranked #1 by each method) or the Best of Top 5 poses (lowest RMSD among top five ranked poses) achieved an RMSD <2.0 Å. All values are normalized to the maximum achievable success rate for DNA (80.65%), which reflects the percentage of DNA test set complexes for which rDock generated at least one near-native pose of the 100 poses.

The results for the full test set, which includes both RNA and DNA complexes, were evaluated (Figure ). Using the rDock default scoring function, the Top Ranked pose was below 2.0 Å in 36.4% of cases. Random pose selection achieved 29.8% (±3.7%) Top-1 success. Reweighting the rDock scoring terms increased Top-1 success to 40.2%. RNAPosers achieved 52.3%, while reranking with IRIS increased this to 54.2%. For the Best of Top 5, rDock achieved 55.1%, random sampling achieved 62.7% (±3.4%), the reweighted model reached 52.3%, RNAPosers achieved 64.5%, and IRIS achieved 74.8%. These results indicate that while the native rDock scoring terms contain some predictive signal, linear reweighting alone does not consistently improve performance and, in the Best of Top 5 setting, performs below the random baseline. In contrast, IRIS outperforms native rDock, the reweighted model, and random sampling across both metrics. Although RNAPosers approaches IRIS in Top-1 success, IRIS provides a larger improvement in the Best of Top 5 setting and yields the highest overall reranking accuracy across the full test set, approaching the theoretical upper limit defined by the quality of poses generated by rDock.

To complement the quantitative evaluation, we examined the three complexes for which IRIS performed worst at identifying the top-ranked pose, where the pose selected by IRIS as Top-1 was among the lowest-ranked poses by RMSD within the 100-pose docking ensemble. In several test complexes, including 8IJC, 1ZPH, and 3GX2, the IRIS-predicted Top-1 pose corresponded to one of the lowest-ranked poses in each complex by RMSD (IRIS-predicted ranks 96, 95, and 94, respectively). Visual inspection of these complexes revealed consistent challenges (Figure S4). In 8IJC, the ligand Pt­(NH3)­2­(2-(pyridin-4-ylmethyl)­benzo-[lmn]­[3,8]­phenanthroline-1,3,6,8­(2H,7H)-tetraone) is an organic-platinum hybrid compound bound to a DNA G-quadruplex. In 1ZPH, the ligand 1,6-dimethyl-4-(4-(4-(1-methylpyridinium-4-ylamino)­phenylcarbamoyl)­phenylamino)­quinolinium is a quinolinium quaternary salt bound to a DNA duplex minor groove. In 3GX2, the ligand is sinefungin, a SAM analog bound in a SAM-I riboswitch pocket.

Although chemically distinct, these ligands share properties that complicate pose reranking: either substantial internal flexibility (sinefungin) or rigid, extended polyaromatic frameworks that constrain the ligand to a narrow set of closely related conformations. Moreover, their binding sites support dense docking ensembles in which many alternative poses differ only subtly in placement or orientation while remaining chemically plausible. These conditions increase the number of near-degenerate poses, making it more difficult for IRIS to identify the true near-native pose reliably from among the rest of the ensemble.

To assess how performance varied by nucleic acid type, we analyzed the RNA and DNA complexes within the test set separately. Of the 126 test set complexes, 95 were RNA and 31 were DNA. For each subset, we first determined the proportion of complexes for which rDock was able to generate at least one near-native pose, representing the theoretical upper limit of reranking success. This maximum achievable success rate was 86.3% for RNA (82 of 95 complexes) and 80.7% for DNA (25 of 31 complexes). All subsequent model success rates were normalized to these respective ceilings to enable fair comparison.

IRIS consistently outperformed all baseline methods on the RNA subset (Figure ). Using the rDock default scoring function, the normalized Top Ranked success rate was 42.7%. Reweighting the rDock scoring terms increased the success rate to 47.6%. Random pose selection achieved a Top-1 success rate of 31.6%, substantially below all structure-aware methods. RNAPosers improved Top-1 performance to 57.3%, while IRIS achieved the highest Top-1 accuracy at 59.8%. For the normalized Best of Top-5 success rates, rDock achieved 64.6%, the reweighted model reached 62.2%, random pose selection achieved 64.8%, RNAPosers achieved 63.4%, and IRIS increased this to 78.1%. Taken together, these results show that IRIS provides the strongest RNA pose reranking performance, with a clear advantage in recovering near-native poses within the top five ranked candidates, while random selection performs comparably to other methods only for Top-5 recovery and fails to achieve competitive Top-1 accuracy.

4.

Normalized performance of IRIS reranking (yellow), RNAPosers (blue), reweighted rDock scoring function (coral), the rDock default scoring function (teal), and random pose selection (gray) on the RNA subset of the test set. Performance was evaluated as the percentage of RNA test set complexes in which the Top Ranked pose (ranked number 1 by each method) or the Best of Top 5 poses (lowest RMSD among top five ranked poses) achieved an RMSD <2.0 Å. All values are normalized to the maximum achievable success rate for RNA (86.32%), which reflects the percentage of RNA test set complexes for which rDock generated at least one near-native pose of the 100 poses.

Results for the DNA subset were also computed (Figure ). Using the default rDock scoring function, the normalized Top-1 success rate was 16.0%. Reweighting the rDock scoring terms decreased this rate slightly to 16.0%. Random pose selection achieved a normalized Top-1 success rate of 24.5%. RNAPosers improved to 36.0%, and IRIS further increased the success rate to 36.0%. For the normalized Best of Top 5 success rates, rDock achieved 24.0%, the reweighted model reached 20.0%, random pose selection achieved 56.5%, RNAPosers achieved 68.0%, and IRIS achieved 64.0%. These findings indicate that improvements in pose reranking for DNA complexes are more limited than for RNA, with IRIS and RNAPosers providing comparable Top-1 performance and both outperforming rDock-based scoring. For the Best of Top 5 metric, RNAPosers achieved the highest recovery of near-native poses, suggesting that RNAPosers may be the preferred choice for DNA–ligand pose reranking under the conditions evaluated here. IRIS, in contrast, is specifically designed and optimized for RNA–ligand pose reranking, where it provides its strongest and most consistent performance gains.

To address whether training on DNA- and RNA-ligand complexes provide some additional information that may influence model performance compared to training solely on RNA-ligand complexes, we retrained IRIS using only RNA complexes from the training set and evaluated it on the RNA subset of the test set. This model was compared to the RNA subset performance of the IRIS model trained on all nucleic acid complexes (RNA + DNA) (Figure S5). For the RNA + DNA-trained model, the normalized Top Ranked and Best of Top 5 success rates on the RNA subset were 59.8% and 78.0%, respectively. Restricting both training and testing to RNA complexes reduced performance to 46.3% for Top Ranked and 73.2% for Best of Top 5. The decrease in performance across both metrics indicates that the IRIS model benefits from the inclusion of DNA complexes in training, likely because these structures expand the range of nucleic acid binding site geometries and interaction patterns encountered during learning. This broader structural and chemical diversity appears to help IRIS capture generalizable features of nucleic acid–ligand recognition, improving its ability to identify near-native poses even in RNA-specific evaluations.

To assess whether improvements in pose ranking reflect more accurate recovery of local nucleic acid–ligand interactions, we further evaluated all methods using I-RMSD. For the full test set, IRIS maintained the strongest overall performance (Figure S6). All success rates reported below are normalized to the maximum achievable I-RMSD success rate of 84.1%, corresponding to the fraction of test set complexes (106 of 126) for which rDock generated at least one pose within 2.0 Å I-RMSD of the experimental structure. Using I-RMSD, IRIS achieved normalized Top-Ranked and Best of Top 5 success rates of 61.3% and 77.4%, respectively, exceeding the performance of the rDock default scoring function (44.3% and 61.3%), the reweighted rDock model (47.1% and 58.5%), and RNAPosers (51.9% and 67.0%). Random pose selection remained substantially below all structure-aware methods for Top Ranked recovery, achieving a normalized Top-1 I-RMSD success rate of 36.3%. In contrast, random sampling achieved a comparatively high normalized Best of Top 5 success rate of 69.5%, outperforming the rDock default scoring function (61.3%), the reweighted rDock model (58.5%), and RNAPosers (67.0%), while remaining below IRIS (77.4%). This result may reflect the dense sampling of interface-proximal poses within the docking ensemble.

When analysis was restricted to RNA–ligand complexes, IRIS again demonstrated the strongest overall performance for I-RMSD evaluation (Figure S7). All success rates for the RNA subset are normalized to the maximum achievable I-RMSD success rate of 85.3%, corresponding to the fraction of RNA test set complexes (81 of 95) for which rDock generated at least one pose within 2.0 Å I-RMSD of the experimental structure. Using I-RMSD, IRIS achieved normalized Top Ranked and Best of Top 5 success rates of 65.4% and 80.2%, respectively, exceeding the performance of the rDock default scoring function (50.6% and 66.7%), the reweighted rDock model (51.9% and 65.4%), and RNAPosers (60.5% and 66.7%). Random pose selection remained below all structure-aware methods for Top Ranked recovery (36.5%) but achieved a comparatively high Best of Top 5 success rate (70.5%), exceeding all methods except IRIS.

Performance on the DNA subset was lower overall under the I-RMSD evaluation (Figure S8). For DNA complexes, success rates are normalized to a maximum achievable I-RMSD success rate of 80.65%, corresponding to the fraction of DNA test set complexes (25 of 31) for which rDock generated at least one pose within 2.0 Å I-RMSD of the experimental structure. IRIS achieved the highest normalized Top Ranked success rate (48.0%), outperforming RNAPosers (24.0%), the rDock default scoring function (24.0%), the reweighted rDock model (32.0%), and random pose selection (35.5%). For the Best of Top 5, IRIS and RNAPosers achieved identical normalized success rates (68.0%), exceeding the rDock default scoring function (44.0%) and the reweighted model (36.0%), while random pose selection achieved a comparable Best of Top 5 success rate (66.6%).

Next, to determine the contributions of individual features to the predictions of the IRIS models, we performed a SHAP analysis. SHAP values quantify the contribution of each feature by distributing the difference between the actual prediction and the average prediction across all input features. Higher absolute SHAP values indicate features that strongly influence RMSD predictions. The mean SHAP value indicates the average directional impact a feature has across all predictions, showing whether a feature tends to increase or decrease the predicted RMSD. In contrast, the mean absolute SHAP value measures the average magnitude of contribution for a feature, regardless of direction, making it a reliable indicator of overall feature importance.

The top ten features in the SHAP analysis, ranked by their mean absolute SHAP values, are shown in Figure , reflecting their overall influence on model performance. Notably, several rDock scoring terms, including SCORE.INTER.norm (0.554), SCORE.RESTR (0.457), SCORE.INTER.ROT (0.287), and SCORE.INTER.VDW (0.165), rank among the most influential. Although the aggregate rDock scoring function (the sum of these terms) fails to rank poses accurately (Figures –), their inclusion alongside diverse chemical and spatial descriptors enables the IRIS model to make more accurate and interpretable RMSD predictions.

6.

Top ten most important features for the IRIS model RMSD predictions. Each dot represents the corresponding feature describing a single ligand pose in the test set. The y-axis lists the individual features, with colors describing whether the feature value is comparatively high (red) or comparatively low (blue) relative to the range of values observed for that same feature across the data set.

Grouped together, the rDock-derived features in the SHAP analysis delineate distinct energetic contributions relevant to docking. SCORE.RESTR represents the penalty term for violating the spatial constraints of the defined docking cavity. In rDock, the cavity is specified by a set of grid points defining the permissible binding volume, and SCORE.RESTR increases when any ligand atom lies outside this volume, with the penalty scaled to the distance of the violation. , SCORE.INTER.norm quantifies the ligand–receptor interaction energy, comprising van der Waals, polar (attractive and repulsive), and entropy-related terms, with the normalized form being divided by the ligand heavy atom count. , SCORE.INTER.ROT reflects rotational entropy penalties associated with ligand flexibility, penalizing poses that require substantial conformational restriction upon binding. , SCORE.INTER.VDW captures van der Waals contributions to the ligand–receptor interaction and reflects steric complementarity within the binding pocket. , As with most rDock scoring terms, lower values are generally more favorable, indicating a more energetically favorable interaction or pose. In the IRIS model, more favorable interaction energies and lower restraint and rotational penalties correspond to lower predicted RMSDs, as reflected by the negative mean SHAP values for SCORE.INTER.norm (mean SHAP = −0.130), SCORE.RESTR (mean SHAP = −0.065), and SCORE.INTER.ROT (mean SHAP = −0.110). In contrast, SCORE.INTER.VDW exhibits a near-zero mean SHAP value (mean SHAP = 0.0038) despite a moderate mean absolute SHAP value, indicating that while steric interactions strongly influence individual predictions, their directional contribution varies across docking scenarios. These patterns underscore the ability of the model to interpret and contextualize energetic terms beyond what any single raw scoring component can achieve on its own.

FpDensityMorgan1 (mean absolute SHAP = 0.106) measures the local density of circular substructures within a Morgan fingerprint of radius 1, capturing fine-grained atomic environment complexity. Higher values correlated with lower RMSD predictions (mean SHAP = −0.049), indicating that more densely featured local environments aid in pose recognition. MinEStateIndex, an electronic descriptor reflecting localized extrema in atomic electrotopological state, showed moderate importance (mean absolute SHAP = 0.133) and trended negatively with RMSD (mean SHAP = −0.013), suggesting that ligands with favorable localized electronic environments are more consistently placed during docking. fr_NH1 (mean absolute SHAP = 0.214) reflects the number of primary amine groups capable of forming hydrogen bonds, consistent with known preferences in RNA-ligand recognition. Its low mean SHAP value (−0.020) suggests variable impact across complexes.

Surface-area–based descriptors further contribute to model interpretability. PEOE_VSA8 (mean absolute SHAP = 0.094) represents the van der Waals surface area associated with atoms in a near-neutral partial-charge range. , Its negative mean SHAP value (−0.051) indicates that specific distributions of near-neutral surface area in the ligand are associated with lower predicted RMSDs. SMR_VSA7 (mean absolute SHAP = 0.075) represents the surface area contribution of atoms weighted by molar refractivity, capturing aspects of ligand polarizability and steric bulk. , The low negative mean SHAP value (−0.024) suggests that ligand polarizability is associated with more consistent docking poses. Finally, com_N–N_mindist (mean absolute SHAP = 0.079) represents the minimum distance, in angstroms, between any nitrogen atom in the ligand and any nitrogen atom in the receptor. This descriptor captures close-contact geometry between heteroatoms that are frequently involved in hydrogen bonding or electrostatic interactions in RNA–ligand complexes. In the IRIS model, com_N–N_mindist exhibits a small positive mean SHAP value (mean SHAP = 0.025), indicating that larger minimum N–N separations are, on average, associated with higher predicted RMSDs.

A complete list of feature definitions, sources, and calculation methods is provided in the Supporting Information, including detailed descriptions of all included energetic, geometric, and chemical descriptors.

Features in the SHAP analysis align with known determinants of RNA–ligand recognition. FpDensityMorgan1, which measures local atomic environment complexity, captures the higher heteroatom content and functional group density common in RNA-binding ligands. The fr_NH1 descriptor highlights the contribution of primary amines, which are frequently enriched in RNA binders due to their capacity to form hydrogen bonds and electrostatic contacts.

MinEStateIndex captures regions of extreme electronic character within a ligand and may influence nucleic acid binding through modulation of hydrogen-bonding propensity or electrostatic complementarity. PEOE_VSA8 captures the spatial extent of weakly polar ligand surface that can contribute to favorable shape complementarity and stacking interactions with nucleobases. , SMR_VSA7, a molar refractivity–weighted surface area descriptor, reflects ligand polarizability and aromatic surface contributions that have been associated with dispersion and π-stacking interactions in nucleic acid–ligand complexes. , Finally, com_N–N_mindist captures close-contact nitrogen atom geometry that is relevant for hydrogen bonding with nitrogenous bases. Taken together, these electronic, distance-, and surface-based descriptors support the notion that RNA-binding ligands with favorable charge distribution, polarizable surface area, and well-positioned heteroatoms achieve more consistent and reproducible docking poses through improved local complementarity at the RNA–ligand interface.

To assess the aggregate influence of related descriptors, features from the SHAP analysis were grouped into broader categories based on their origin and functional relevance: RDKit molecular descriptors, rDock subscores, ligand–receptor distances and geometries, binding site interactions, and RDKit fragment counts. The summed mean absolute SHAP values for each category were normalized to the total SHAP magnitude across all features to yield their relative contributions (Figure S9). rDock subscores contributed the largest share of the total SHAP impact (43.2%), indicating that docking-derived energetic terms dominate the feature contribution to the RMSD predictions of the model. RDKit molecular descriptors accounted for 29.4%, indicating the role of general ligand physicochemical properties. Ligand–receptor distance and geometry metrics contributed 14.6%, reflecting the importance of spatial fit and distance-based features in determining ranking accuracy. Binding site interaction features and RDKit fragment counts contributed smaller but comparable fractions (6.6% and 6.3%, respectively), indicating more limited but non-negligible influence. This category-level analysis shows that IRIS primarily leverages docking scores while still integrating chemically and spatially diverse descriptor classes, supporting the benefit of combining rDock-derived terms with ligand- and geometry-based features. A detailed description of the features included in each category is provided in the Supporting Information.

Next, we assessed the sensitivity of the IRIS model to training data availability by constructing a learning curve using the IRIS training set (Figure S10). To ensure fair evaluation at each training size, we applied K-fold cross-validation with 5 folds. For each fold, the model was trained on a subsample of the training data and evaluated on the held-out fold, and this was repeated across ten incremental training set sizes. In this setup, 20% of the data is held out in each fold for validation, so even at the largest plotted training size the model is trained on only 80% of the available training data. This means the training MAE at the rightmost point of the curve is based on a 20% smaller data set than the full training set used in fitting of the true IRIS model, which may contribute to the observed gap between training and validation errors. The shaded regions represent the standard deviation across folds and highlight variability due to data partitioning.

The resulting learning curve shows that training error remains low and highly stable throughout the incremental addition of data. Meanwhile, the validation error exhibits a steady, slowing decline with increasing training set size. This downward trajectory, and the absence of an upward inflection in validation error, suggests that the model avoids overfitting, where performance on new data degrades as it memorizes training noise. However, a persistent gap between the training and validation performance remains. This gap indicates that the IRIS model possesses high capacity, effectively capturing patterns within the training complexes, but is currently limited by the structural diversity available in the training set. That the validation curve has not yet fully reached a horizontal asymptote suggests that the model remains in a data-limited regime. Further improvements in generalization of the IRIS model would likely be achieved by incorporating additional, structurally diverse NA-ligand complexes that better represent the broader chemical space.

To assess the consistency of IRIS predictions across ligand poses, we calculated the MAE between predicted and actual RMSD values for each complex in the test set (Figure ). Each complex includes approximately 100 poses, allowing us to evaluate how prediction errors vary across the ligand pose ensemble. Complexes with high MAE exhibit larger deviations between predicted and actual RMSD values across poses, reflecting inconsistent or poor ranking performance. For example, complex 1ZPH (complex 1 in Figure ) displays the highest average prediction error at 5.62 Å, with a standard deviation of 0.87 Å, indicating high error magnitude despite relatively low variability across poses. In contrast, complex 6VUI (complex 126 in Figure ) shows the lowest MAE at 0.27 Å with a standard deviation of 0.41 Å, suggesting consistently accurate predictions across all poses.

7.

MAE (Å) in prediction accuracy across test set complexes. The x-axis represents test set complexes sorted in descending order of MAE, while the y-axis shows the MAE of predicted RMSD values for all ligand poses within each complex. Error bars represent one standard deviation above and below the mean MAE, in some cases plotted asymmetrically to prevent values below zero, reflecting the variability in prediction accuracy across poses within each complex.

To characterize molecular factors associated with prediction error, we compared ligand features between the top 10 complexes with the highest and lowest per-complex MAE (Table ). For each group, the mean and standard deviation of molecular weight, number of rotatable bonds, and cavity volume are reported.

2. Comparison of Average Ligand Features for Test Complexes with the Top 10 Highest vs. Top 10 Lowest MAE (Å) Values .

descriptor	high MAE (mean)	high MAE (stdev)	low MAE (mean)	low MAE (stdev)
molecular weight (Da)	557.3	292.3	331.9	192.4
number of rotatable bonds	7.5	5.9	3.0	3.25
cavity volume (Å3)	1339.6	515.3	841.3	580.2

^aFor each group, the mean and standard deviation of each ligand feature are reported.

Complexes with the highest MAE values tended to contain larger and more flexible ligands than those with the lowest MAE values, as reflected by higher mean molecular weight (557.3 Da vs 331.9 Da) and greater numbers of rotatable bonds (7.5 vs 3.0). These ligands were also bound in substantially larger cavities on average (1339.6 ų vs 580.2 ų), providing greater conformational space and potentially increasing the difficulty of accurately predicting pose RMSD. In contrast, low-MAE complexes generally involved smaller, more rigid ligands in more spatially constrained binding sites, which may facilitate more consistent pose prediction.

We then visually analyzed the top three complexes in the highest and lowest MAE groups. To this end, we used PyMOL to identify similarities or differences that may contribute to their respective error ranges in pose prediction (Figures S11 and S12). For clarity, each visualization retains only the nucleic acid receptor (RNA or DNA) bound to the ligand and the ligand itself in the experimentally resolved pose, with all other structural components omitted to focus on the nucleic-acid binding pocket.

We note that the complexes highlighted here do not directly reflect the ability of IRIS to identify the correct top-ranked pose, but rather its ability to predict RMSD values consistently across all poses for a given complex. A complex may exhibit a high per-complex MAE because its ensemble of poses spans a wide structural range, even if the top-ranked pose is close to the native structure, while a low per-complex MAE indicates that predicted RMSDs are consistently accurate across poses with minimal variability. This distinction is essential when interpreting MAE-based comparisons, as they evaluate prediction consistency rather than pose-ranking success alone.

The three lowest-MAE complexes contain ligands and binding modes that restrict the set of plausible docking geometries, yielding tight pose ensembles with limited variability. 6VUI (MAE = 0.27 Å ± 0.41 Å) features 7-deaza-7-aminomethyl-guanine, a rigid nucleobase analog bound within a preorganized riboswitch pocket that seems to enforce a single dominant binding orientation. 8OM0 (MAE = 0.38 Å ± 0.30 Å) contains 3,4,5-trihydroxybenzoic acid, a compact planar ligand bound within the folded core of a group II intron RNA, where the geometry of the binding pocket and the ligand’s polar functional groups tightly localize its binding position. 278D (MAE = 0.45 Å ± 0.54 Å) contains 2′-bromo-4′-epidaunorubicin, a bulky anthracycline analog whose polycyclic core is conformationally preorganized and binds DNA through intercalation, which strongly limits the range of viable orientations and translations. In all three systems, geometric constraints imposed by the binding site and/or ligand preorganization reduce structural dispersion across the docking ensemble, resulting in consistently low RMSD prediction errors.

The three highest-MAE complexes contain ligands and binding contexts that support broad, near-degenerate pose ensembles. 1ZPH (MAE = 5.62 Å ± 0.87 Å) features 1,6-dimethyl-4-(4-(4-(1-methylpyridinium-4-ylamino)­phenylcarbamoyl)­phenylamino)­quinolinium, an elongated cationic quinolinium ligand bound in the DNA minor groove. 1EEL (MAE = 5.55 Å ± 0.77 Å) contains 2,5-bis-[4-[cyclopenta-1,3-dien-5-ylamino-1-aminomethyl]-phen-1-yl]­furan, a long, aromatic ligand that similarly binds along the DNA minor groove. In both systems, many docked poses remain within the native groove but differ primarily by head-to-tail reversal in orientation, producing large RMSD differences despite occupying the same binding site. 6E8U (MAE = 5.26 Å ± 3.16 Å) instead contains TO1-Biotin bound to an RNA aptamer, where the fluorophore head is anchored within the binding pocket while the biotin tail is connected by a long, flexible linker. Variability in the orientation and conformation of this linker substantially increases RMSD across the pose ensemble even when the ligand head remains correctly positioned.

Conclusions

Building on previous ML-based methods − for improving pose accuracy in RNA-ligand docking, we present here IRIS, an ML-based framework designed to improve RNA-ligand docking pose ranking by reranking rDock-generated poses. In the test set constructed, IRIS consistently ranks a greater proportion of near-native poses below 2.0 Å RMSD than the standard rDock scoring function, a linear regression model that uses the rDock scoring function terms, and the RNAPosers scoring function. SHAP analysis shows that both rDock scoring function terms and ligand-specific physicochemical properties are important for pose ranking, suggesting that future ML-based scoring approaches should focus on selection and weighting of these descriptors.

The downward trend in the learning curve (Figure S10) indicates that the scale and diversity of training data influence the effectiveness of the ML-based rescoring method. Therefore, including a larger and more diverse set of nucleic acid–ligand complexes likely contributed to the improved ranking accuracy in IRIS compared to previous methods. This finding is consistent with previous studies, which have demonstrated that the composition of the training data set influences the performance of ML models used for molecular property prediction. ,

By identifying poses that are structurally closer to the experimentally relevant binding mode, IRIS may be useful in increasing the probability that subsequent binding affinity calculationswhether through scoring functions, molecular dynamics simulations, or free energy perturbation methodsare based on physically meaningful conformations rather than misaligned or nonphysiological poses. Our pose reranking tool is particularly valuable in high-throughput virtual screening (HTVS), where docking algorithms generate a set of poses for each ligand, but their scoring functions do not always rank the near-native pose at the binding site correctly. Without effective pose filtering, inaccurate poses can negatively impact the accuracy of affinity predictions, as illustrated by the influence of the ligand binding pose on protein affinity predictions. By prioritizing near-native poses, IRIS helps focus computational efforts toward ligands with a higher probability of genuine binding, thereby improving the overall efficiency and accuracy of lead selection in RNA-targeted drug discovery.

The limited availability of large data sets remains a challenge for many ML methods aimed at predicting chemical properties. The accuracy of IRIS is limited by the size of the training data set, and additional experimentally determined RNA-ligand complexes are likely to improve generalization. Model accuracy is also found to decrease for ligands with higher molecular weight, more rotatable bonds, and larger cavity volume, indicating challenges in ranking highly flexible and bulky ligands. This is consistent with previous studies which observed that larger ligands with more rotatable bonds tend to correspond with greater inconsistency and decreased accuracy in protein–ligand docking pose predictions. Other limitations of IRIS are that it is strictly a reranking tool, meaning its performance depends on the generation of near-native poses by rDock, and it does not perform pose refinement.

Considerations for future work include extending IRIS beyond pose reranking to predict binding affinities. This goal might be achieved by leveraging current IRIS features but integrating experimental binding data and training the model on affinity measurements, such as dissociation constants, for complexes for which such values have been experimentally determined. Moreover, expanding IRIS to function independently of rDock search constraints would enhance its ease of use, allowing it to generalize across diverse docking pipelines. Adapting IRIS for use with other docking programs is straightforward and would further broaden its applicability.

Data and Software Availability The IRIS software is freely available at https://github.com/AndrewAmburn/IRIS. The data that support the findings of this study are available at 10.5281/zenodo.15597919.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c11891.

• Model and data set comparisons and analyses (Table S1–S5 and Figures S1–S8), and rDock parameters (PDF)

• PDB IDs and corresponding ligand CCD codes for all NA–ligand complexes used in the present study (XLSX), Redundant ligand pairs and their corresponding sequence similarities (XLSX), Training set ligand clusters with high Tanimoto similarity (XLSX) Test and training set splits with corresponding tSNE values (XLSX) (XLSX)

• A list of each feature and the category in which the feature is included (CSV), The full feature set used in the IRIS models accompanied by the generation method and a brief description of each feature (CSV) (CSV)

The authors declare no competing financial interest.