Journal of Chemical Information and Modeling

Structure-based drug discovery relies on the prediction of protein-bound poses of new molecule designs, the accuracy of which can impact downstream prioritization. While it is expected that crystal structures of similar molecules would provide the best template for predicting the poses of new designs, the time and cost required motivates identifying a point of diminishing returns for collecting new structures. Using 403 crystal structures of SARS-CoV-2 main protease from the open science COVID Moonshot project, we explore the tradeoff between the cost and utility of obtaining crystal structures for accurately predicting poses of designed molecules. We observe that similar reference ligands enable superior pose prediction and show that success plateaus after approximately five crystal structures per generic Bemis-Murcko scaffold, exceeding 95% for the campaign's lead series. This work provides practical recommendations for resource allocation in structure-enabled drug discovery campaigns.

G protein-coupled receptor (GPCR) allosteric modulators (AMs) offer significant therapeutic advantages over orthosteric drugs, yet structure-based virtual screening lacks validated protocols accounting for the conformational complexity of GPCR allosteric sites. We benchmark docking protocols using PDB experimental structures and structural ensembles derived from Gaussian accelerated Molecular Dynamics (GaMD) simulations across four Class A GPCRs (including the muscarinic M2 and M4 receptors, the {beta}2-adrenergic receptor, and the C-C chemokine receptor type 2) with four programs (Glide HTVS, AutoDock Vina, DOCK3.8, and Boltz-2) against experimentally validated modulator libraries and property-matched decoys. GaMD ensemble docking improved early AM enrichment across all four targets under at least one program. Glide ensemble docking was the only protocol to consistently improve early AM recovery across all four targets, ranking known actives almost exclusively within the top 0.5% of compounds at CCR2 and improving M2R active recovery nearly 9-fold relative to the PDB structure. GaMD free-energy landscape topology governed ensemble re-ranking strategy selection: population-skewed landscapes favored top binding energy ranking (BEmin) while flat, multi-populated landscapes favored average binding energy ranking (BEavg), and at targets with dominant low-energy states, a single GaMD cluster matched or exceeded full ensemble or PDB performance. Taking the union of top percentile hits identified by both ensemble re-ranking methods, BEmin / BEavg, maximizes chemical diversity at the earliest percentiles. Program-specific scaffold recovery biases further motivated a consensus BEmin / BEavg approach to maximize hit diversity. The Boltz-2 deep-learning program showed minimal sensitivity to GaMD templates and underperformed conventional docking, suggesting its affinity predictions complement rather than replace physics- and empirical-based docking approaches for GPCR AM screening.

Activity cliffs, defined as structurally similar molecules with vastly different properties represent a fundamental challenge in modern day drug discovery for property prediction models. While Graph Neural Networks (GNNs) have advanced molecular property prediction, they inherently struggle with this problem due to representation collapse and node over smoothening. In this work, we evaluate the efficacy of incorporating various contrastive learning-based loss functions, including the Supervised Contrastive (SupCon) configurations into several GNN backbones to navigate this structure-activity landscape. Tested across the MoleculeACE benchmarks, the results indicate that while contrastive losses prevent GNN mode collapse to some extent and drastically improve the property-based separation in the embedding space, their impact on absolute predictive accuracy remains highly dataset dependent. Ultimately, these methods offer critical gains in interpretability for mapping optimized lead generations, opening up new directions for models that respect these sharp pharmacological discontinuities.

Molecular docking and co-folding engines are widely used to prioritize compounds for wet-lab validation, yet their accuracy is known to vary substantially across protein targets for reasons that remain only qualitatively understood. Here we benchmark six docking and co-folding engines (RevDock, DiffDock, Boltz2, AutoDock-GPU, rDock, and PandaDock) across 14 protein families, evaluating scoring power, ranking power, docking power, and physical validity. Rather than treating engine performance as protein-family-specific, we classify all 14 families into six mechanistic groups according to which of four scoring-function simplifications, rigid receptor, pairwise additivity, fixed point charges, and implicit solvent, is most severely stressed by that familys binding site. This framework helps explain, rather than simply describe, where each engine succeeds or fails: RevDocks CNN rescoring layer mitigates the pairwise additivity and fixed-charge limitations relative to physics-only scoring, achieving the highest overall pose accuracy (73.3% of poses [≤] 2.0 [A] RMSD), while Boltz2s sequence-based co-folding bypasses the rigid-receptor assumption and achieves comparable affinity correlation (mean Pearson r {approx} 0.60 for both engines). PandaDock, run with expanded conformational sampling, matches RevDock on pose accuracy (72.1% of poses [≤] 2.0 [A], lowest median RMSD at 0.96 [A]) and exceeds AutoDock-GPU on affinity correlation (mean r = 0.460), indicating that the performance of a physics-based scoring function is limited as much by search adequacy as by the scoring function itself. These results suggest that engine selection for a docking or co-folding campaign should be guided less by an engines aggregate benchmark ranking and more by which of these four structural and physical characteristics dominate the target of interest.

Current rational drug design relies predominantly on computational (CADD/AIDD) methods that model binding thermodynamics and static conformations of target proteins, primarily in their inactive states. However, the kinetic parameters that govern experimental efficacy-such as catalytic turnover and signaling potency-are determined by molecular interactions with transition states (TS), intermediate states (IS), and the entire continuum of conformations along the least free-energy activation pathway. The absence of this dynamic dimension has fundamentally limited the predictive power and success rate of conventional structure-based approaches. Here, we present a structural database that systematically maps the complete activation trajectories of pharmaceutically relevant targets, encompassing TS, IS, and all connecting conformational ensembles. This resource offers multiple strategic advantages for drug discovery: enabling rational targeting of previously "undruggable" proteins, facilitating biased agonism/antagonism design, revealing cryptic allosteric sites in inactive conformations, identifying novel transient pockets along the activation route, rationalizing the mechanisms of existing drugs, predicting mutational effects on activation barriers, and prospectively forecasting drug resistance and off-target liabilities. We demonstrate the utility of this database through representative case studies and provide implementation guidelines for integration into existing discovery pipelines. More detailed information can be found at our website: https://www.momedpamdb.com/en.

High-quality molecular datasets are essential for reliable machine learning in cheminformatics and bioinformatics, yet dataset quality is rarely assessed systematically and its relationship with downstream model performance remains poorly understood. Here, we present MolJam, an open-source framework for quantitative assessment of molecular dataset quality across five dimensions-structural integrity, data quality, experimental information quality, chemical space coverage, and data distribution-using 12 standardized metrics. Application of MolJam to 11 MoleculeNet and eight ChEMBL-derived datasets revealed widespread and heterogeneous quality issues, including undefined stereochemistry in up to 70.72% of molecules, inconsistent molecular representations, and contradictory labels. We next asked whether improving these quality metrics necessarily improves machine learning performance. Refinement of the ESOL and Lipophilicity datasets increased their MolJam quality scores but produced mixed effects on predictive performance, suggesting a competing influence of reduced dataset size. Controlled ablation experiments further demonstrated that both dataset quality and data quantity contribute to model performance and, notably, that retaining molecules with incomplete stereochemical information can outperform their removal when the resulting gain in data quantity offsets the quality penalty. Thus, molecular dataset curation cannot be reduced to maximizing data cleanliness alone but requires balancing multiple dimensions of data quality against information loss. MolJam provides a standardized framework for diagnosing molecular dataset limitations, comparing benchmark quality, and quantitatively evaluating how data curation decisions influence downstream machine learning.

Virtual screening promises access to tens of billions of synthetically accessible, diverse compounds, yet it is rarely used as a primary hit-discovery strategy in contemporary drug-discovery campaigns. We argue that this gap reflects the real-world underperformance of the underlying docking and scoring methods: classical docking is generalizable but limited in accuracy by simple functional forms and parsimonious parameterization, whereas recent machine-learning approaches are highly expressive but do not generalize well to novel molecules and pockets, their reported accuracy often inflated by train-test leakage. To address these challenges, we introduce DODock and DOScore, docking and scoring ML/physics hybrid frameworks that are also highly expressive, yet generalize much better out of distribution compared with the prior ML approaches. This generalization has been prospectively tested in several ways. First, DODocks blind prediction of a drug candidate binding to PCSK9 was compared to the crystal structure that was subsequently solved, recovering the binding pose to 1.2 [A] RMSD. We also used DODock and DOScore in prospective virtual screening campaigns against four therapeutic targets, spanning an ectoenzyme (CD73), a kinase (IRAK4), an extended-substrate protease (FXI), and an allosteric inhibition of protein-protein interface (IL17). These screens yielded many chemically novel, biochemically and cellularly active inhibitors. In the case of CD73, which is a historically challenging target for virtual screening, our screen resulted in a roughly hundredfold improvement in hit rate over a recent machine-learning screen. Our results indicate that the apparent ceiling in the accuracy of virtual screening that seemed to have somewhat plateaued in the last two decades is not fundamental, and that structure-based interrogation of ultralarge chemical space may eventually become a credible primary route to novel chemical matter.

Molecules generated by deep learning models are often difficult to synthesize. Their synthetic accessibility can be improved with automated retrosynthetic analysis, which allows for identifying synthons. However, synthons in a SMILES can be scattered throughout the string depending on the path taken through the molecular graph used to generate the SMILES. We tested whether the ensemble of possible SMILES for a molecule can be used to generate a concatenable ordered sequence of string fragments (blocks) from SMILES that match potential synthons obtained through automated retrosynthetic analysis. We found that exhaustively sampling the SMILES space of a molecule improves the coverage of retrosynthetic breaks. We achieved full coverage of retrosynthetic bonds in string form for 85\% of the 1.9 million molecules in the MOSES dataset. Doing so allowed us to test our block SMILES in an unconditional de novo drug design test case with MolGPT and Monte Carlo Tree Search (MCTS). We found that using blocks as an LLM's token did degrade MolGPT performance due to the curse of dimensionality. However, using the SMILES selected by our blocking algorithm with the default SMILES tokenizer improved the reproduction of physico-chemical properties of samples and also improved uniqueness, novelty, and validity. The MCTS outperforms our MolGPT models in terms of validity and novelty. However, samples generated by the MCTS had physico-chemical properties that were further away from the MOSES baseline than the samples produced by molGPT, with an improved distribution of quantitative estimation of drug-likeness (QED).

As chemical libraries expand into the trillions of molecules, Virtual SYNthon Hierarchical Enumeration Screening (V-SYNTHES) has emerged as a leading strategy for making gigascale virtual screening computationally tractable. In V-SYNTHES, a Minimal Enumeration Library (MEL) of chemical fragments is docked against a target first, and only the top-scoring fragments are expanded into full ligands for large-scale docking. However, among the large number of comparably well-docked fragments, only a small fraction can be expanded under a fixed docking budget, leaving most similarly promising fragments unexplored. General-purpose prescreening tools can be adopted to address this constraint, reallocating the same docking budget across a larger pool of fragments' enumerated full ligands by their proxy score. However, such tools are applied without accounting for target-specific pocket environments. One such method, PharmacoNet, predicts interaction hotspots from a protein structure and ranks candidates via graph matching against a fixed set of interaction-type weights. We recognize that V-SYNTHES's initial fragment-docking step, ordinarily used only for selection of best fragments for expansion, already reveals which of these hotspots and interaction types a given pocket actually favors, and we can recover this signal to fine-tune PharmacoNet accordingly. We introduce MEL-Steered PharmacoNet, a parameter-efficient adaptation framework that specializes PharmacoNet to a given target through two composable mechanisms: (i) empirical density-map steering of predicted pharmacophore hotspots, and (ii) empirical fine-tuning of interaction-type scoring weights. Across three structurally distinct GPCR targets (CB2, GPR91, 5-HT2AR), MEL-Steered PharmacoNet achieves substantial enrichment factor (EF100) gains over a random baseline, and improves EF100 over PharmacoNet by 8.94x, 6.87x, and 1.69x, respectively. The fitted per-target weights further reveal distinct, chemically interpretable interaction profiles that PharmacoNet's generic fixed weights fail to capture. These results show that fragment-docking data already generated by the standard V-SYNTHES pipeline can adapt a general-purpose pharmacophore prescreening method to an individual target, significantly improving its performance while retaining its ultra-fast screening ability, with no additional experimental data or model retraining.

Generative artificial intelligence is accelerating molecular design, yet the relative suitability of available models for different targets and stages of preclinical drug discovery remains unclear. Here we benchmarked 12 molecular generation and optimization methods across 176 curated protein-ligand systems spanning diverse therapeutic target classes, with experimentally validated ligands providing reference chemical space. The evaluated methods encompassed pocket-conditioned 3D generation, diffusion and flow-based modeling, autoregressive construction, reference-conditioned optimization and synthesis-aware design. Performance was assessed using operational robustness, chemical validity, uniqueness, molecular and scaffold diversity, quantitative estimate of drug-likeness, synthetic accessibility, docking, physicochemical and ADMET properties, and computational resource requirements. The results revealed architecture-dependent trade off such as receptor-conditioned methods exploited binding-pocket geometry, flow-based approaches enabled efficient sampling, reference-conditioned methods favored analogue generation, and synthesis-aware approaches improved chemical feasibility, but no method consistently optimized all criteria. To address the functional potential of generated molecules, we further developed a state-aware functional classifier (SAFC) that integrates molecular dynamics derived receptor ensembles, ensemble docking and protein ligand interaction graphs. SAFC provided dynamics-aware functional activity rankings for generated molecules that were partly complementary to docking, drug-likeness and synthetic accessibility scores. These findings support hybrid, stage specific deployment of generative models rather than reliance on any single architecture or evaluation metric. This study provides practical guidelines for generative AI based preclinical drug development processes.

Lead optimization, the systematic refinement of therapeutic compounds through iterative structural modification, faces a dual challenge in modern drug discovery: navigating astronomically vast molecular design spaces while balancing conflicting demands on potency, pharmacokinetics, and safety. We present MASCOT (Multi-Agent SearCh for molecular OpTimization), a role-specialized multi-agent framework for molecular optimization. Integrated with a chemically constrained graph-editing search, MASCOT coordinates three specialized agents: a trade-off agent that reprioritizes competing objectives, a strategy agent that adapts how molecular edits are proposed, and a reflection agent that distills lessons from previous decisions. Computational experiments showed that MASCOT achieved the best performance over competing methods on six benchmark settings. On the SARS-CoV-2 main protease task, its mean docking-score improvement was 3.6 times that of the strongest baseline. Applied to the clinically used anesthetic remimazolam (RM), MASCOT prioritized RM-1, which showed a shorter liver microsomal half-life, higher brain exposure, and a larger therapeutic index than RM. Subsequent derivative design yielded RM-7. Extensive animal studies established RM-7 as a rapid-recovery intravenous anesthetic candidate with greater potency, faster functional recovery, a wider safety margin, and preserved flumazenil reversibility. These results demonstrate that multi-agent coordination can link adaptive molecular search to medicinal chemistry and experimental pharmacology.

Three-dimensional intermolecular interactions govern molecular recognition and are fundamental to the pharmacological activity of small-molecule drugs. We propose that an interaction profile, comprising shape, electrostatics, and directional pharmacophores, is a sufficient and transferable design specification for molecular design. We introduce ShEPhERD-2, a 3D generative model that generates molecular structures conditioned on explicit interaction profiles. ShEPhERD-2 generates low-strain, drug-like molecules more efficiently and with greater fidelity to target interaction profiles than its predecessor while enabling precise control over molecular design through pharmacophore prioritization, substructure constraints, and composition of multiple interaction profiles. We demonstrate the application of a single model to bioisosteric fragment merging, dual-target design, selectivity engineering, and modality hopping without task-specific retraining. These results establish interaction profiles as a chemotype-agnostic interface between structure hypotheses and molecular design.

Biased signaling occurs when a ligand preferentially activates a specific signaling pathway at a certain receptor within a particular cellular environment. As pharmaceuticals, functionally selective (biased) compounds can engage beneficial signaling pathways while avoiding those linked to adverse effects, resulting in safer and targeted therapeutics. However, existing experimental methods for characterizing ligand bias are often time-consuming and costly, and they hardly elucidate the structural basis of biased signaling. Here, we present microswitch-guided sampling (MGS), a computational method that uses conformational changes in molecular switches to assess ligand bias in G protein-coupled receptor systems. Using initial 2 s all-atom molecular dynamics simulations of the {beta}1-adrenergic receptor ({beta}1AR), the chemokine receptor CXCR3, and the -opioid receptor (OR) in complex with either G protein or {beta}-arrestin and known biased ligands, we identified microswitches whose conformational changes were linked to activation of a specific signaling pathway. We then sampled simulation trajectories for representative complex conformations selected at these microswitch-change time points and used them as starting points for 500 ns ligand-swapped simulations, in which we exchanged biased ligands between complexes. We observed clear differences in conformational changes between complexes bound to pathway-activating and non-activating ligands in at least two of three replicas across all examined systems, demonstrating that MGS can distinguish G protein-biased from {beta}-arrestin-biased ligands. These findings show that MGS is an effective tool for assessing signaling pathway bias in GPCR systems and suggest that this methodology can be applied to other GPCR targets to act as a fast, ns-scale computational assay to detect functional selectivity with molecular dynamics.

Traditional screening-based drug discovery is inherently limited by the astronomical scale of the chemical space. Generative modelling offers a compelling alternative to the classical search paradigm and enables rational, bottom-up design of novel and target-specific small molecules. However, its impact has been hampered by challenges in synthetic accessibility of the designed compounds and lack of large-scale experimental validation. Here, we introduce LDDM (Large Drug Discovery Model), a generative framework that supports a range of drug discovery tasks, including constrained and unconstrained docking, fragment linking and growing, and de novo design. We further introduce a programmable design algorithm that enables accurate design of synthetically accessible compounds satisfying various fine-grained objectives. We experimentally validated the designed or optimised ligands for five therapeutically relevant protein targets. In all cases, LDDM achieved high success rates, allowing us to identify molecules with confirmed binding affinity while synthesizing only a small number of generated compounds. The best designs were structurally characterised through NMR spectroscopy and X-ray crystallography, demonstrating high prediction accuracy. Overall, LDDM provides a scalable and flexible platform for the rapid and tailored design of small molecules and non-natural peptides for therapeutic applications.

The binding affinity of a ligand to a protein is influenced by the protonation and tautomeric states of both partners. However, this relationship remains under-investigated due to the limited availability of computational tools capable of considering all charge and tautomer states in a scalable manner to study clinically relevant systems. Here, we use Multi-Conformation Continuum Electrostatics (MCCE) to calculate the protonation and tautomer distributions of nine kinase domains bound to 18 FDA-approved inhibitors while considering their Boltzmann-ensemble. Our simulations show that protein net charge and proton distribution remain largely stable even upon binding charged inhibitors. Our results find that individual inhibitor charges are dynamic, frequently increasing, or decreasing upon binding a specific protein target. Kinase-inhibitor binding significantly shifts the relative probabilities of low-energy states ({Delta}G < 2.5 kcal/mol), though it does not recruit higher-energy conformers into the bound population. Our consideration of all possible charge states and tautomers enable us to identify when tautomer have significant significant free binding energy differentials (3-6kcal/mol). In turn, we find that minority species can become the dominant component in the bound state, emphasizing the necessity of considering ensemble-wide protonation and tautomer states to accurately predict protein-ligand binding energetics.

A three-dimensional pharmacophore fingerprint records the binding features a molecule can present. It is a description of a hand in search of a glove. Because it is defined by presented features instead of two-dimensional structure, it can identify pharmacophoric similarity between structurally distinct compounds, which is what scaffold hopping and non-obvious me-too design require. The descriptor has remained a niche tool because its cost is dominated by conformer generation. In the reference pipeline, generating 100 conformers requires 2.82 s of the 2.86 s needed to fingerprint one screening collection compound; the bit calculation requires 0.039 s. We therefore removed the conformational stage. PharmCast is a feedforward neural network that predicts all 10,549 bits of a PharmPrint ensemble fingerprint directly from a SMILES string. On the same machine, PharmCast generated pharmacophore fingerprints for two molecules and compared them in 0.584 ms, whereas the conventional conformer-based pipeline took 5.71 s. PharmCast version 10 was trained on 5,887,229 molecules drawn from a screening collection, activity-backed ChEMBL compounds from 142 to 1000 Da, and peptide loops excised from crystal structures. We evaluated 155,648 purchasable catalog compounds excluded from every training set, 139,700 activity-backed ChEMBL compounds not present in the version 10 training set, and 13,500 peptide loops reserved for testing. Median fingerprint error, Pearson r, and pairwise ranking accuracy were 0.008, 0.980, and 0.936 for screening collection chemistry; 0.016, 0.984, and 0.952 for loop peptides; and 0.027, 0.936, and 0.889 for activity-backed ChEMBL compounds. The reference calculation reproduces itself at an error of 0.006 and r of 0.995. Ensemble pharmacophore fingerprints can therefore be predicted from two-dimensional structure alone at a cost suitable for large-scale collection screening and virtual library exploration. Keywords: pharmacophore, fingerprint, scaffold hopping, surrogate model, virtual screening, applicability domain

Structure-based drug discovery is a widely used paradigm for the rational design of novel small molecule therapeutics. However, the benefits conferred by the use of structural information has seen limited adoption in machine learning, where ligand-only ("2D") models are still the industry standard for molecular property or binding affinity prediction. Structure-based ("3D") ML models for binding-affinity prediction promise to present a clear advantage, but have not yet overtaken existing 2D models. Here, we show that physics-based priors can improve predictive performance of structure-based models by comparing different model architectures with varying physical priors on several prediction tasks. We present the Modular Training and Evaluation of Neural Networks (mtenn) package, where we decompose affinity prediction into separate steps of embedding structure into learned representations and combining those embeddings into a predicted binding affinity. We consider both E(3)-invariant and E(3)-equivariant architectures to determine the importance of encoding roto-translational inductive biases, as well as different methods for combining learned embeddings. By first optimizing several aspects of model construction using the general purpose PDBBind dataset, we are able to improve the performance and data efficiency of structure-based models. When subsequently trained and evaluated on the COVID Moonshot small molecule drug discovery dataset, our tuned models perform on par with industry standard ligand-only models. Our decomposed model framework highlights that encoding some physical priors improves model performance, while more complex biases such as equivariance offer limited benefit. Additionally, structure-based models generalize better to an unseen target and display higher training efficiency. Overall, these results emphasize that structure-based models benefit from their ability to incorporate physics-informed constraints, giving promising directions for model architecture development. These results also suggest that the strength of these models may be in tasks specifically aimed at generalizability, providing guidelines for their use in early-stage drug discovery campaigns.

Drug resistance often emerges through mutations that reshape protein conformational ensembles, allowing inhibitor potency to be altered without directly disrupting protein-ligand contacts. Consequently, the molecular mechanisms underlying these effects remain difficult to identify, hindering the design of inhibitors with durable potency against rapidly evolving drug targets. To address this challenge, we developed ROBUST, a physics-informed machine learning framework that integrates experimental measurements, molecular dynamics simulations, and interpretable statistical modeling to identify energetic determinants of inhibitor potency. Using HIV-1 protease as a stringent model system for drug resistance, we determined 157 new inhibition constants and integrated them with prior measurements to establish a comprehensive experimental dataset of 364 protease-inhibitor pairs. Analysis of the corresponding molecular dynamics simulations revealed a compact network of energetic interactions that quantitatively explains how resistance mutations and inhibitor modifications jointly reshape molecular recognition and inhibitor potency across the multidimensional resistance landscape. Models constructed from these energetic signatures predicted binding free energies with high accuracy (RMSE < 1 kcal mol-1) and successfully generalized to unseen darunavir analogs. Importantly, ROBUST accurately prioritized unseen analogs based on their potency across the variant set, demonstrating its utility for identifying compounds with improved resistance profiles. Together, these results uncover the energetic principles through which resistance mutations and inhibitor modifications alter potency and establish ROBUST as a mechanistically interpretable framework for translating these insights into resistance-aware inhibitor optimization against rapidly evolving drug targets.

Proteolysis-targeting chimeras (PROTACs) are large, flexible molecules that frequently extend beyond conventional drug-like chemical space, including Lipinski's rule-of-five boundaries. Their membrane permeability has often been linked to molecular chameleonicity and the ability to adopt compact conformations with reduced polar surface exposure; however, the underlying conformational determinants remain unclear. Here, we investigated 20 PROTACs spanning distinct passive-permeability classes using enhanced-sampling molecular dynamics (ES-MD). All permeability classes accessed compact states with low radius of gyration (Rg) and polar surface area (PSA), indicating that compact-state accessibility alone does not distinguish permeable from non-permeable PROTACs. We therefore introduce maximum ring size (MRS), defined as the size of the largest intramolecular hydrogen-bonded ring, as a metric for identifying compact conformations stabilized by long-range intramolecular hydrogen bonding. Like the other descriptors examined, MRS neither separates nor ranks the permeability classes, but helps locate folded states in the large ensemble. Representative compact conformations from selected highly permeable, moderately permeable, and non-permeable compounds were subsequently examined by all-atom unbiased MD in water and chloroform. The highly permeable PROTAC held a closed-flat ensemble in both environments, but in two structurally distinct arrangements, whereas the other two redistributed across heterogeneous ensembles. Passive diffusion therefore appears to depend on the population, stability, and environment-dependent persistence of permeable-like states rather than on compactness alone.

Virtual screening has become central to early-stage drug discovery, and structure-based approaches have recently been reframed as a retrieval problem through contrastive learning methods such as DrugCLIP, which project protein pockets and ligands into a shared embedding space. However, what these abstract representations exactly encode, and how they relate to conventional notions of structural and chemical similarity, remains unclear. Here we present a systematic dissection of DrugCLIP's latent space. We show that its pocket embeddings, despite not being explicitly trained for the task, set a new state of the art in pocket similarity search while running over 100 times faster than existing structural descriptors, and that this embedding space is structurally coherent and robust to conformational variation. Ligand embeddings, by contrast, encode a pocket-aware notion of chemical similarity that only partially mirrors fingerprint-based measures. Using a rigorous de-leakage benchmark, we further show that DrugCLIP generalises to unseen proteins and chemistries rather than memorising training data, recovering the correct bound ligand within the top 1% of 50,000 candidates for 55-75% of novel targets. Performance nonetheless declines under increasingly realistic screening conditions, a drop attributable to sidechain reorientation across apo, holo and AlphaFold-derived structures, and to residue mismatch when using predicted pockets. These findings clarify the practical boundaries of DrugCLIP's applicability, identify pocket prediction accuracy as a key factor for improving performance, and offer a transferable framework for interpreting the latent spaces of related contrastive pocket-ligand encoders. Together, these results support the improvement of existing methods and the development of a new generation of contrastive screening approaches.