Polaris Challenge: Data-driven Priors to Improve Docking for Pose Prediction

Abstract

Accurate prediction of ligand binding poses is a central challenge in structure-based drug design. In this work, we developed a workflow for the ASAP-Polaris-OpenADMET antiviral competition, leveraging crystallographic training structures to predict time-split test ligand poses. Our approach relied on the open-source Vina-GPU software for ensemble docking, augmented with fragment-derived priors for pose generation and scoring. When fragment information was unavailable we fell back to MM/GBSA for pose scoring. Overall, our method performed competitively and achieved over 50% success in predicting ligand poses within 2 Å RMSD for both SARS-CoV-2 and MERS-CoV protease targets. Both fragment-informed rescoring of Vina-generated poses and fragment-constrained pose generation improved accuracy, underscoring the utility of crystallographic fragments as priors. Importantly, this information transferred successfully from SARS-CoV-2 to MERS-CoV, highlighting the value of conserved binding pockets in guiding pose prediction across related viral proteases. Our results demonstrate that docking-based workflows, when augmented with transferable fragment knowledge and physics-based refinements, can achieve competitive accuracy for pose prediction within protein families while remaining efficient and fully open-source.

Graphical Abstract

Introduction

The COVID-19 pandemic spurred an unprecedented global research effort aimed at discovering antiviral therapeutics.^1–4 Among the most intensively studied coronavirus targets is the main protease (Mpro),^5–8 a cysteine protease responsible for cleaving viral polyproteins into functional enzymes. Because Mpro is essential for viral replication and lacks close human homologs, it is an attractive drug target⁵ that stimulates multiple drug discovery projects,^9–23 with some eventually producing clinically approved inhibitors such as Nirmatrelvir (Paxlovid)⁹ and Ensitrelvir.¹⁰

To accelerate antiviral discovery for SARS-CoV-2 Mpro,^6,24–28 a number of open-science initiatives^3,25 produced rich experimental datasets. The COVID Moonshot consortium,^3,25 for example, reported more than two thousand biochemical measurements and over eight hundred crystal structures for inhibitor complexes. Numerous Antiviral Drug Discovery (AViDD) centers²⁹ were established in late 2021 to rapidly expand anti-virals for a range of viruses and targets that were often structurally enabled. These resources laid the groundwork for the Antiviral Drug Discovery Challenge³⁰ organized through the Polaris Hub³¹ in collaboration with ASAP Discovery³² and OpenADMET³³. Specifically, in the ligand pose prediction track, participants were provided with a chronologically curated training set of approximately 770 SARS-CoV-2 Mpro crystal structures and tasked with predicting the crystal binding poses of 98 new SARS-CoV-2 Mpro ligands and 97 MERS-CoV Mpro ligands. This prospective design, training on data available up to a fixed time point and testing on subsequent ligands, mimics real-world hit-to-lead campaigns and provides a stringent assessment of a given methods’ generalizability. During evaluations, predictions were scored using the success rate (fraction of poses within 2Åheavy-atom RMSD) across all test ligands.³⁰

Ligand pose prediction is inherently a two-component problem: candidate conformations must be generated and then evaluated using scoring functions to identify the most probable binding modes.³⁴ Traditional docking software follows this paradigm, and significant advances in both sampling strategies and scoring methodologies have led to improved screening efficiency and greater scoring accuracy.^35–39 In particular, when a large amount of existing crystal structures are available, ligand-template-based docking^37,40–42 can generate accurate results without the use of sophisticated algorithms. In parallel, thanks to the curation of high quality protein ligand structure datasets,^43–46 recent developments in machine learning docking methods,⁴⁷ particularly protein–ligand co-folding models,^48–50 have shown promise in predicting binding poses within a single inference step. Nevertheless, there remains considerable value in benchmarking and refining classical docking approaches, both to establish robust baselines and to locate the performance boundaries of traditional methods. In this work, we intentionally forgo deep learning models and instead investigate to what extent widely adopted tools like AutoDock Vina⁵¹ can be enhanced through fragment-guided sampling and scoring. By treating the Polaris challenge as a prospective benchmark for classical docking, we aim to systematically characterize the conditions under which traditional methods succeed, where they fail, and how they can be effectively incorporated into a typical drug discovery workflow.

Here, we present a docking and rescoring workflow developed for the challenge. The process begins with ensemble docking using GPU accelerated Vina (Vina-GPU)⁵² to generate initial poses, followed by a scaffold-aware ligand fragmentation step to distinguish test ligands with training-set analogs from those without. For ligands with similar fragments, we leverage fragment alignment from related training complexes to guide pose generation similar to ligand template-based docking and apply a custom fragment-based scoring function for rescoring and selection. Ligands lacking such similarity, or failing to meet fragment-scoring thresholds, are advanced instead through a physics-based MM/GBSA refinement of generated poses. By combining fragment-informed priors with physics-based optimization, this workflow is designed to exploit the complementary strengths of data-driven and structure-based approaches to predict ligand poses of target-specific proteins.

Methodology

As depicted in Figure 1, our ligand pose prediction strategy follows a multi-step, two-branch workflow that integrates ensemble docking, fragment-based structural similarity detection, and physics-based MM/GBSA rescoring. For each of the two task targets, we first generated a diverse ensemble of protein conformations by combining clustered Monte Carlo Side Chain Ensemble (MCSCE)⁵³ conformers with clustered crystal complexes from either the organizers or the RCSB PDB.⁵⁴ As part of the standard drug discovery practice,⁵⁵ all ligands were then docked to these ensembles to produce a pool of candidate poses.

Figure 1: Schematic representation of our multi-step workflow to generate the final ligand binding poses.

The fragment based search utilizes conserved chemical fragments of training ligands to help inform a set of structural and scoring priors for pose of the test ligand. Lack of train-test fragment similarity generates a ligand pose based on standard docking and free energy scoring approaches.

Ligands sharing conserved chemical fragments with those in the training set were routed through a fragment-guided branch, where additional poses were generated by fragment-constrained conformer sampling and docking into structurally similar training-set proteins. These poses were ranked using a custom fragment-based score that quantified structural agreement with known fragment conformations. Ligands without shared fragments in the training set, or whose fragment-based scores failed to meet a quality threshold, were instead processed through the alternative refinement branch, where MM/GBSA calculations were applied to the ensemble-docked poses to select the most favorable binding mode. This workflow leveraged available training data while applying physically rigorous refinement in data-poor regimes, ensuring a balance between data-driven and physics-driven guidance.

Protein Ensemble Generation and Docking.

Since previous studies have shown that the backbone of SARS-CoV-2 Mpro and MERS-CoV Mpro remains largely rigid upon ligand binding,^23,30 we focused on sampling diverse side-chain conformations while keeping the backbone fixed. To this end, we used the Monte Carlo Side Chain Ensemble (MCSCE) method,⁵³ which proposes rotamers from backbone-dependent libraries, incrementally builds all-atom side chains with hydrogens, and accepts or rejects configurations based on force-field energies. This procedure efficiently generates physically reasonable ensembles with fixed backbone atom positions. The resulting side chain ensembles capture side-chain rearrangements that are important for induced fit, while not requiring running more expensive molecular dynamics simulations for longer timescales. The MCSCE method has been validated extensively against X-ray crystallography and NMR J-coupling data.⁵³

For each reference protein, we generated 2,000 MCSCE conformers and clustered them by side chain atom features over binding site residues identified from literature.^56–59 Specifically, since all MCSCE conformers share an identical backbone, we represented each conformer solely by the 3D coordinates of its binding site side chain atoms. These coordinates were concatenated into a feature vector for each structure, which were then clustered using DBSCAN.⁶⁰ To further expand diversity, we incorporated 770 SARS-CoV-2 Mpro structures from the challenge training set and 44 MERS-CoV Mpro structures curated from the RCSB PDB. These conformers were first aligned and then clustered on their side chain atom positions of binding site residues following the same procedure. The final protein ensemble consisted of nine representative SARS-CoV-2 Mpro structures (four from MCSCE and five from the training crystal structure set) and four representative MERS-CoV Mpro structures (one from MCSCE and three from the PDB crystal structures). All representative structures were aligned to their respective reference proteins using Cα atoms and served as docking targets for the test ligands, using Vina-GPU for pose generation. Details about the docking box information are included in the Supplementary Information. For each docking run, we retained the top four poses for SARS-CoV-2 and the top nine for MERS-CoV, ensuring a comparable total number of candidate poses between the two targets. Each ligand docking run on average takes around one second to finish on a single A-40 GPU.

Fragment Identification and Pose Generation.

To better leverage structural information from the provided training set, we incorporated a fragment-based search strategy alongside conventional docking. This approach is motivated by the observation that many test ligands share conserved chemical substructures, including core scaffolds or functional moieties, with ligands in the training set. Since these fragments often mediate key interactions with pocket residues, recognizing them enables us to bias pose generation toward geometries consistent with experimentally validated binding modes, which will improve the accuracy of binding pose predictions beyond docking alone.

To create the fragments, we implemented a scaffold-aware ligand fragmentation protocol using the training set ligands as references. The protocol begins by identifying all ring systems in the molecule using RDKit,⁶¹ where rings sharing two or more atoms are treated as a single fused system. Each fragment is then defined by the ring system plus its attached atoms, extending outwards until one bond away from any other ring system. As illustrated in Figure 2, the target ligand is partitioned into four fragments denoted by different colors. By design, this protocol keeps individual ring systems intact, preserving the local chemical context essential for protein interactions and effectively accommodating the trefoil-like structures of SARS-CoV-2 Mpro binders. Following this protocol, we identified 81 of 98 SARS-CoV-2 test ligands that shared at least one identical fragment with a training-set ligand. Given the high conservation and structural similarity between the SARS-CoV-2 Mpro and MERS-CoV Mpro binding pockets,⁵⁹ MERS-CoV test ligands were also compared against the SARS-CoV-2 training set, yielding 87 of 97 ligands eligible for downstream fragment-based processing.

Figure 2: Example of scaffold-aware ligand fragmentation.

(a) SARS_Test_Lig_21 crystal structure with binding pocket labels. (b) 3 example ligands with identical colored fragments as the test ligand SARS_Test_Lig_21 from the provided crystal structures.

For ligands with fragment matches, we sampled additional poses beyond the ensemble docking results using two strategies: (1) re-docking the test ligand into every training protein whose ligand shared a similar fragment, following the same Vina-GPU protocol, and (2) generating 10 conformers per protein in RDKit⁶¹ with the matched fragment fixed at its crystallographic position while allowing the remainder of the ligand to flexibly fit. This procedure yielded up to ~300 poses for high-similarity ligands, compared with 36 poses from ensemble docking for ligands without fragment matches. Owing to the high binding mode homology between SARS-CoV-2 and MERS-CoV as shown in Supplementary Figure S1, the same workflow was applied cross-target for MERS-CoV ligands using SARS-CoV-2 proteins, further increasing their number of candidate poses despite having no MERS-CoV Mpro structures in the training set. The expanded sampling framework remains computationally efficient. On average, docking against 8 training proteins takes approximately 8 seconds per ligand on a single A40 GPU, while the RDKit conformer generation step took less than one second on a CPU.

Custom Fragment-Based Scoring.

Traditional docking scores are largely empirical, parameterized to perform reasonably well across a wide range of targets and ligand chemotypes.³⁴ While effective for general-purpose virtual screening, they are not necessarily optimal for specialized tasks, as will be shown in the next section. Moreover, docking scores are calculated independently for each candidate pose and cannot directly incorporate prior knowledge from a training dataset. These limitations motivated us to develop a new pose-ranking metric that explicitly leverages information from the provided experimentally resolved protein-ligand complexes in the training set.

In our formulation, candidate poses with fragment matches were ranked using a geometry-based metric defined as

$$S_{\text{frag}}=\sum _{\text{frag}}{\text{RMSD}}_{\text{frag}}\times \frac{N_{\text{frag}}}{N_{\text{ligand}}},$$ (1)

where RMSDfrag is the heavy-atom RMSD between the fragment in the candidate pose and the corresponding fragment from the training set, N_frag is the number of heavy atoms in the matched fragment, and N_ligand is the total number of heavy atoms in the ligand. The weighting factor N_frag/N_ligand ensures that larger fragments, which capture more of the ligand geometry, contribute proportionally more to the final score than smaller ones. Summing over all matched fragments yields a single measure of structural agreement between a candidate pose and experimentally validated binding conformations. Herein, we refer to this metric as the Custom Score.

The underlying intuition behind the design of the Custom Score parallels templated ligand docking, where chemically similar ligands are expected to adopt comparable binding geometries to preserve favorable protein–ligand interactions, particularly when the protein structure is unchanged. Lower scores indicate closer alignment of candidate fragments with training crystal structures, reflecting higher structural fidelity and a greater likelihood of predicting correct binding poses. In contrast, higher scores signify larger geometric deviations and reduced retention of key interactions. To calibrate a score cutoff for filtering unreliable matches, we created a validation set by holding out 70 random ligands from the training data and used the other 700 ligands to build the fragment library. Our analysis of the hold-out set showed a linear correlation between our Custom Score and ligand RMSD, where a score of 10 corresponds to an RMSD of 2Å. Therefore, for ligands with at least one pose below the custom score threshold of 10, we chose its lowest-scoring pose as the prediction.

Fallback MM/GBSA Scoring.

For ligands lacking structurally similar fragments in the training set, or when the fragment-based score is too high (exceeding 10) to be reliable, we employed MM/GBSA, a more rigorous physics-based approach, to rank the poses. We chose to apply MM/GBSA selectively because our analysis of the 70-ligand validation set showed that MM/GBSA rescoring actually worsened the average RMSD for poses already considered reliable by our Custom Score.

To perform MM/GBSA rescoring, each candidate pose first underwent a geometry optimization in complex with the protein, in which the positions of all protein heavy atoms were constrained. This optimization served to relieve steric clashes while preserving the overall binding geometry, a procedure shown in prior studies to improve agreement with experimental binding affinities.⁶² Then, the GBSA score was calculated using MMPBSA.py⁶³ program in AmberTools with GB-OBC⁶⁴ implicit solvation model (igb = 2). Amber ff14SB force field⁶⁵ was used for protein and GAFF2⁶⁶ along with AM1-BCC charges^67,68 was used for the ligand. The constrained geometry optimization was performed using OpenMM⁶⁹ by assigning zero mass to frozen atoms. The pose with the most favorable MM/GBSA score (i.e., the lowest predicted binding free energy) was selected as the final model, providing a robust, physics-grounded fallback when similarity-based guidance proved insufficient.

Results

In this section, we present a retrospective analysis of binding poses extracted from different stages of our workflow, evaluating how effective each stage would be if considered in isolation. This stepwise assessment allows us to identify the contributions of prior knowledge from training data and physics-based refinement to overall prediction accuracy. By examining the incremental improvements across stages, we highlight why the combined workflow leads to more reliable ligand pose prediction compared to a baseline docking protocol.

Overall Performance

As highlighted in Supplementary Figure S2, our workflow ranked 6th overall in the contest, exceeding the performance of purely ligand template-based methods but trailing behind methods that fine-tune cofolding models. To systematically evaluate the components of our workflow shown in Figure 1), we categorized our contest submissions for the SARS-CoV-2 and MERS-CoV Mpro targets into three groups based on their origin and filtering outcome. The largest group, Similar-Passed, consisted of ligands identified through similarity searches that successfully passed our Custom Score threshold. The second group, Similar-Rejected, included ligands that were initially identified by similarity but were subsequently filtered out due to poor performances in Custom Scores. The final category, Dissimilar, comprised ligands that lacked known structural analogues and were processed using ensemble-based solely physics methods. This tiered framework allows for a direct assessment of how each stage of our methodology contributed to the final prediction quality.

The results shown in Figure 3 and Supplementary Table S1 demonstrate a clear performance hierarchy across these categories, validating our prioritization strategy. The Similar-Passed group achieved the highest success rates, with 51 out of 70 SARS-CoV-2 predictions and 44 out of 73 MERS-CoV predictions having an RMSD < 2Å. In contrast, the accuracy was substantially lower for the Similar-Rejected group (4/11 for SARS-CoV-2 and 4/14 for MERS-CoV). The Dissimilar group, representing the lowest-confidence predictions, had the lowest success rates, with 3/17 for SARS-CoV-2 and 3/10 for MERS-CoV. Overall, our workflow yielded greater accuracy for the SARS-CoV-2 target (59.00 ± 5.03 %) compared to the MERS-CoV target (52.49 ± 5.11 %), as determined by bootstrapping with 1000 samples.³⁰ We attribute this difference to the more extensive training data available for SARS-CoV-2. The consistently better performance of the Similar-Passed category across both targets underscores the efficacy of using similarity searches to leverage existing structural knowledge. Collectively, these results establish that our combined approach of similarity-based selection and rigorous scoring provides a robust method for distinguishing high-confidence predictions from less reliable ones, significantly improving pose prediction accuracy.

Figure 3: Box plots of RMSD distribution for ligand pose prediction filtered into different categories of similarity to the training set for SARS-CoV-2 and MERS-CoV Mpro targets.

Similar-Passed consists of ligands that pass fragment similarity searches and custom score thresholds. Similar-Rejected consists of ligands that pass fragment similarity searches but fail on the custom score threshold. Dissimilar requires a pose prediction using standard docking and MM/GBSA scoring.

Performance of Similar Test Ligands

Next we consider the test ligands identified by our scaffold-aware fragmentation method with at least one fragment identical to that in the training set. This analysis allowed us to systematically evaluate six distinct approaches: Vina Ensemble (ensemble docking with Vina Score), Vina Similar (docking to the most similar training protein with Vina Score), Custom Ensemble (ensemble docking with Custom Score), Custom Similar (similar protein docking with Custom Score), Custom Frag (fragment-constrained generation with Custom Score), and our baseline Native (retrospective docking to the test protein structure with Vina Score).

As shown in Figure 4A, ensemble docking with a Vina Score yielded inferior performance compared to Native docking for best pose selection. Docking to similar training proteins rather than pre-generated ensembles significantly improved RMSD values, although it still falls short of the Native baseline. Hence sampling of similar protein structures and ligands provides clear advantages over ensemble-based approaches, effectively better capturing the conformational specificity required for accurate pose prediction. As seen in Figure 4B, even Vina Score and using native testing proteins yielded only weak correlations between Vina Score and RMSD, indicating that the Vina Score is not well suited for reliably identifying the best poses from a large candidate pool. To address this, we applied our fragment-based Custom Score to rescore poses from both Ensemble and Similar docking. As shown in Figure 4, rescoring with the Custom Score consistently reduced RMSD values and improved score–RMSD correlations, confirming that fragment-derived information from the training set was successfully incorporated.

Figure 4: Comparison of pose prediction accuracy for ligands containing identical training fragments using different sampling and scoring approaches.

(A) the average and standard deviation of best pose RMSD values determined by each method. (B) the distribution of score–RMSD correlation. For each testing ligand, top 10 poses selected by each method were used to calculate the Spearman R between score and RMSD. A total of N=81 ligands for SARS-CoV-2 and N=87 for MERS-CoV. (C) 2D ligand structure and binding mode of SARS-CoV-2 test ligand 43. Crystal ligand poses (green) aligned with best poses from ensemble docking ranked by (D) Vina Score (cyan), (E) Custom Score (teal), and (F) best pose from fragment-constrained conformer generation ranked by Custom Score (orange). Further details are provided in Supplementary Figure S3.

The example of test ligand 43 bound to SARS-Cov-2 MPro (Fig. 4C) further highlights this point: while the best Vina-scored pose was entirely misoriented, rescoring with the Custom Score identified a near-native conformation, validating the ranking power of fragment-based scoring (Figure 4D,E). Beyond rescoring, we also tested fragment-constrained conformer generation combined with the Custom Score. This approach achieved the best RMSD values overall. As illustrated for test ligand 43, fragment-constrained generation produced a pose within 0.6 ÅRMSD of the crystal structure, faithfully recovering all key interactions, including the pyrazolopyridine-HIS163 hydrogen bond, the chlorobenzene hydrophobic contacts, and the linker amide-GLU166 hydrogen bond that was absent in the rescored pose (Fig. 4F). The benefit of this approach is particularly evident for flexible ligands, where docking alone often fails to sample correct conformations. By constraining conformer generation with fragment priors, the sampling space is reduced and the likelihood of recovering the correct pose increases.

Finally, the same performance hierarchy was observed for MERS-CoV ligands, despite all upsampling and scoring being derived from SARS-CoV-2 training fragments and protein structures. Although MERS-CoV results were generally lower in absolute terms, the relative improvement trends remained consistent, with Custom Frag achieving the best RMSD, Custom Score improving upon Vina Score, and Similar docking outperforming Ensemble docking. This transferability provides strong evidence that fragment-based approaches can leverage conserved binding site architectures to improve pose prediction across related coronavirus proteases.

Performance of MM/GBSA Rescoring

Our final analysis focuses on ligands in the low-data or low-confidence regime, which represent the most challenging prediction scenarios. These were categorized as either Similar-Rejected (ligands with fragment matches that failed our Custom Score threshold) or Dissimilar (ligands lacking any matching fragments in the training set). For these cases, we applied MM/GBSA rescoring as a physics-based refinement stage to assess whether more rigorous energy calculations could improve pose selection.

As shown in Figure 5, MM/GBSA rescoring led to modest improvements across these difficult cases. It slightly reduced average RMSDs and yielded small but consistent gains in Spearman R for the best poses sampled from different methods. The refinement is best illustrated by SARS-CoV-2 test ligand 77 (Figure 5D). In this case, the top Vina Ensemble pose in Figure 5E was poorly aligned (RMSD > 6 Å), deviating from the ground truth ligand pose. With the help of our fragment-based Custom Score, the γ-lactam scaffold was placed to form desirable interaction with GLU166 and reduced the RMSD to 3.3 Åin 5F. Furthermore, MM/GBSA helped the ligand recovered all key moieties except the solvent-exposed morpholine, yield a pose with 2.2 Å RMSD to the crystal structure as demonstrated in 5G. Overall, MM/GBSA functions as a useful optimizer that can refine poses and improve rankings to some degree, but it does not fundamentally change the outcome for cases where the initial sampling fails. Its value therefore depends critically on the quality of the starting ensemble generated by fragment-guided sampling.

Figure 5: Performance of MM/GBSA rescoring on challenging ligands without similar training fragments.

MM/GBSA provides modest improvements in (A,B) RMSD and (C) Spearman R, highlighting its role as a refinement step rather than a decisive predictor. (D) Illustration of MM/GBSA rescoring on SARS-CoV-2 test ligand 77. Best pose selected by (E) Vina score (cyan), (F) fragment-based score (magenta), and (G) MM/GBSA score (yellow), superposed with the crystal structure (green). More details are provided in Supplementary Figure S4.

To investigate the practical utility of MM/GBSA further, we performed a retrospective analysis by applying MM/GBSA rescoring to the high-confidence Similar-Passed category. As shown in Supplementary Table S2, for SARS-CoV-2, where the fragment priors were derived from abundant homologous crystal structures, MM/GBSA rescoring worsened the average RMSD of these already accurate poses. For MERS-CoV, for which no specific training data was available, MM/GBSA performed only slightly better in pose accuracy compared to our data-driven prior approach. This observation suggests that fragment-based scoring is superior when extensive, homologous training data exists, and furthermore performs comparable to a physics-based approach like MM/GBSA for related but a less-characterized target as in this case for MERS-CoV. Given that MM/GBSA is far more expensive, and performs less well on SARS-CoV-2, this analysis further supports that fragment-guided sampling and scoring offers and excellent tradeoff in computational efficiency and accuracy.

Discussion and Conclusion

The ligand pose prediction problem is inherently two-step: candidate poses must first be sampled, and then accurately scored and ranked to obtain the best pose. In this work, we present a docking-based workflow that achieves over 50% success in predicting ligand poses within 2 Å RMSD for both SARS-CoV-2 Mpro and MERS-CoV Mpro. While Vina docking and scoring alone performed poorly, their combination with fragment-derived priors and robust rescoring proved effective, highlighting the importance of incorporating prior knowledge to guide the selection stage.

Within this framework, fragment-informed scoring and fragment-constrained sampling provided the largest performance gains, steering pose selection toward geometries consistent with experimental observations. Instead of focusing on maximum common substructure in template-based docking, our fragment-constrained sampling method yields a more diverse set of candidate pose while retaining key protein-ligand interactions, leading to more effective prediction when coupled with a scoring function. Interestingly, even though our fragment information was derived exclusively from SARS-CoV-2, it transferred effectively to MERS-CoV, suggesting that conserved pockets across related proteins can provide valuable priors for pose prediction. While for ligands without suitable fragment matches, the more expensive MM/GBSA rescoring served as a moderately effective fallback, often refining initially suboptimal poses into more native-like conformations, although its value depends critically on the quality of the starting ensemble. As shown in our retrospective analysis, the physics-based refinement worsened the results for SARS-CoV-2 and only provided small improvements for MERS-CoV from the fragment transfers. Together, these findings emphasize that physics-based docking engines can generate sufficient diversity at the sampling stage, but their success depends on tailored scoring strategies that integrate domain knowledge.

We also observed an orthogonality between sampling and scoring. The score–RMSD correlation was strongest when sampling with Vina and scoring with the fragment-inspired Custom Score, even though fragment-constrained generation achieved lower RMSD values overall. By contrast, combining fragment-driven sampling with fragment-informed scoring yielded weaker correlations. This suggests that mixing sampling and scoring strategies, rather than tightly coupling them, may improve discrimination among candidate poses. An interesting direction for future study would be to score fragment-constrained poses with Vina or apply Custom Score to constrained docking results to further differentiate pose quality. These findings highlight not only the utility of fragments in scoring but also the broader advantages of docking-based sampling for generating diverse candidate ensembles.

Compared with end-to-end co-folding models, physics-based docking retains additional advantages. AutoDock Vina generated a wide range of plausible poses without any learned scoring augmentation, preserving structural diversity that may be lost in co-folding models⁷⁰. For new proteins not represented in the co-folding training data, or for proteins with multiple binding sites, this pose diversity can benefit downstream rescoring, hybrid methods, or expert evaluation.

These findings suggest that physics-based docking remains a valuable tool for ligand pose prediction, particularly for targets with limited training data, multiple pockets, or less-conserved binding environments. By integrating fragment-derived priors when available, and falling back on physics-based rescoring when not, such workflows can balance diversity in sampling with selectivity in pose ranking. Moving forward, hybrid approaches that combine the sampling power of docking with the discriminatory strength of learned or knowledge-based scoring functions may offer a robust and generalizable route to high-accuracy pose prediction.

Supplementary Material

Details of the data formats and usage notes are provided in the Supplementary Information.

Synopsis:

A computational workflow that leverages ensemble docking, training fragment priors, and MM/GBSA for the ASAP-Polaris-OpenADMET ligand pose prediction challenge.

Data and Software Availability

All the codes for our multi-step workflow to ASAP-Polaris-OpenADMET ligand pose challenge are provided in a publicly accessible GitHub repository: https://github.com/KSUN63/Polaris-Submission. The MCSCE code can be found at: https://github.com/THGLab/MCSCE. The data of the ASAP-Polaris-OpenADMET challenge are available at: https://fs.polarishub.io/2025-01-asap-discovery/raw_data_package.zip