Designing small-molecule drugs that selectively bind to a target protein while avoiding unintended interactions with other proteins is essential for reducing side effects that contribute substantially to clinical attrition. Although structure-based drug design (SBDD) has advanced rapidly with deep learning, most generative models still optimize binding to a single target and disregard selectivity due to limited training data and unknown off-target interactions. Motivated by this gap, we introduce SELECTBENCH, a comprehensive benchmark assessing the selectivity of generative SBDD models. SELECTBENCH provides three curated evaluation tasks reflecting different real-world drug discovery scenarios: literature-based case studies, safety screening panels used in pharmaceutical development, and large-scale structure-based evaluation sets. We further introduce a suite of metrics to quantify selectivity and conduct a systematic evaluation of nine representative SBDD models, including methods explicitly trained for selectivity and widely-used inference-time guidance techniques. Our results show that even selectivity-trained models achieve only modest off-target discrimination and inference-time guidance provides limited gains, underscoring the need for method improvement. SELECTBENCH provides both a characterization of where current generative SBDD stands on selectivity and an open, extensible platform to accelerate progress. The benchmark is available at https://github.com/lovetaro/selectbench.
Introduction
Decreasing the high attrition rate in the drug discovery and development process is a primary goal of the pharmaceutical industry, as the current success rate of new drug candidates is less than 10% [67]. One of the main challenges is striking an appropriate balance between drug efficacy and adverse effects as early as possible in order to reduce safety-related attrition [66]. Indeed, analyses of clinical trial data from 2010 to 2017 show that lack of clinical efficacy and unmanageable toxicity are responsible for 30-50% of current clinical failures [67, 26]. The ability of small molecules to interact with multiple proteins is commonly referred to as polypharmacology. Polypharmacology of drug-like small molecules beyond their intended targets is a fact in drug discovery, as it has been estimated that over half of approved small-molecule drugs interact with more than five protein targets [37]. Polypharmacology can be beneficial for drug efficacy in some cases due to the benefits of targeting multiple signalling pathways [19]. However, unexpected interactions can equally lead to undesired and often severe side effects, and has resulted in the withdrawal of several drugs from the market in the past [29]. Directed polypharmacology has recently been promoted as a new paradigm for drug discovery, particularly in oncology [64]. Controlling the polypharmacology of small-molecules implies appropriate tuning of their binding selectivity: we want to rationally select which target(s) they bind and do not bind to.
The field of SBDD, which aims to find small-molecule ligands that bind to specific three-dimensional (3D) sites in proteins with high affinity and selectivity, has seen rapid progress driven by recent deep learning (DL) breakthroughs in protein structure prediction [1, 39].
In parallel, a growing body of generative SBDD methods has emerged, leveraging structural information of protein–ligand complexes to propose novel drug candidates [31, 35, 38, 49, 57, 59, 73]. However, most of these approaches focus on optimizing small molecules binding affinity against a single target, and do not consider the crucial selectivity aspect of the drug development process. Although recent efforts have begun to incorporate selectivity into their modeling objectives [16, 74, 71], a systematic evaluation of how well generative SBDD models achieve selective binding remains to be conducted and is crucial to understand the real-world utility of their outputs.
Here, we introduce SELECTBENCH: the first comprehensive evaluation of the selectivity of SBDD generative models (See Figure 1). SELECTBENCH aims to address the following research questions:
Figure 1. Off-target interactions are a major consideration in real-world drug design (top left) and often not accounted for in current generative SBDD (top right). SELECTBENCH provides tasks, metrics, and baselines to assess selectivity in generative SBDD (bottom).
Research Questions
How do we align in silico assessment of generative SBDD models with the selectivity requirements in real-world drug discovery scenarios?
Can state-of-the-art SBDD generative models design selective molecules, despite not necessarily being trained to do so? Can inference-time techniques improve this?
To facilitate the investigation of these questions, we provide the following contributions in this work:
Contributions
Task curation: We design and curate three new evaluation tasks to assess the selectivity of small molecules. These tasks cover a range of design scenarios, including real-world drug discovery use cases. We provide an easy-to-use, open framework for users to enrich SELECTBENCH and design their own selectivity case studies.
Establishing metrics: We provide a suite of metrics to quantify molecular selectivity.
Establishing baselines: We benchmark nine representative SBDD generative models to establish reference performance levels, as well as two inference-time steering techniques, providing a comparative framework for future works.
Related works
Selectivity in drug design. Beyond potency, selectivity is a fundamental requirement in drug discovery, both in the traditional “one drug, one target” paradigm and in polypharmacological settings [36]. It plays a key role in reducing off-target toxicity, caused by interactions with unintended proteins, as well as on-target toxicity, which can arise when the intended protein is inhibited in healthy contexts [61, 3]. Designing for selectivity, however, is substantially more challenging than optimizing affinity for a single target. Prior work has identified several structure-based principles that can improve selectivity, including binding site shape complementarity, electrostatic complementarity, and conformational plasticity [36]. In addition, allosteric modulation (binding to a site different from the active site) has been an alternative strategy to achieve more selective control.
SBDD generative models. Deep learning approaches for SBDD span several modeling classes, including autoregressive, flow-, and diffusion-based methods. Autoregressive models generate molecules atom by atom, predicting their types and positions conditioned on a protein pocket [48, 57]. More recently, flow- and diffusion-based methods have come to dominate the SBDD landscape [30, 31, 63], although they typically focus on single-target generation. There has been growing interest in incorporating binding specificity into generative models. Prior work explores this through contrastive objectives over targets [27], physics-informed energy-guided formulations that promote stable binding [17], or implicit priors that bias generation toward pocket-specific chemistry [71].
Evaluation metrics in generative modeling for SBDD. Evaluation in generative SBDD is largely centered on single-target objectives. Common metrics include docking scores (e.g., Vina) as empirical estimates of binding affinity [21], alongside ligand-centric measures of drug-likeness (QED) [10], synthesizability (SA score) [24, 55], and adherence to medicinal chemistry heuristics (e.g. Lipinski’s Rule of Five [45]). More recent work has expanded evaluation to include the quality of generated 3D binding poses [32, 15]. However, only a limited number of studies explicitly consider selectivity, typically via docking-based comparisons across targets [27, 71].
Evaluation tasks for selective SBDD generative models. Evaluation of selectivity in SBDD generative models remains limited. Prior work typically relies on single case studies [63, 74], subsets of CrossDocked2020 [27, 25, 49], or randomly selected off-targets from the Protein Data Bank (PDB) with low sequence similarity [71]. While case studies lack statistical robustness, random target selection often lacks biological relevance. Although recent efforts have explored biologically grounded test sets (e.g., kinase panels) [16], such resources are not publicly available. These limitations highlight the need for standardized, biologically relevant benchmarks for evaluating selectivity in generative SBDD models.
Evaluation tasks
Overview
We propose three evaluation tasks: two manually-curated and based on literature-backed drug discovery use-cases, and a third to provide a larger-scale evaluation setting for structure-based small-molecule selectivity (see Figure 2 for a graphical summary and Sec. A.1 for licenses).
Figure 2. SELECTBENCH task summary. SELECTBENCH-Case focuses on case studies, SELECTBENCH-Safety models safety screening, and SELECTBENCH-XL enables large-scale evaluation.
SELECTBENCH-Case: Five selectivity tasks based on on-/off-target pairs from the same protein family or subfamily, drawn from established literature case studies.
SelectBench-Safety: A pharmaceutical safety panel of 85 off-targets linked to adverse drug reactions (ADRs) based on secondary pharmacology assay recommendations, forming 5 selectivity tasks with one on-target and up to 85 off-targets.
SelectBench-XL: A large-scale benchmark from CrossDocked2020 with 9,900 synthetic on-/off-target pairs, grouped into difficulty categories based on structural similarity.
To ensure accessibility and standardization for DL applications, we provide Croissant files [2] describing dataset metadata, structure, and semantics. We describe each task in detail below; additional curation details are provided in Sec. B.1.
SelectBench-Case: Selectivity assessment from literature-based case studies
We identified five protein pocket pairs from the literature to study small-molecule selectivity. The pairs were chosen based on application domain and prior use as benchmarks in computational studies. The case studies are summarized in Table 1, with additional details for each pair in Appendix § 10.§ 2.
Table 1. Summary of case studies in SelectBench-Case.
SelectBench-Safety: Targeted selectivity for improving safety-related drug attrition
Secondary pharmacology evaluates off-target effects unrelated to a compound’s intended therapeutic action. Brennan et al. introduced Brennan-77, a panel linking drug targets to adverse events, which has informed widely used in vitro safety screening assays in pharmaceutical research and development [13, 12, 50, 4]. We constructed an in silico analogue of such a safety panel by curating protein structures for targets in Brennan-77 (Sec. B.3). Taking into account the assay endpoints required for a comprehensive screening, we curated multiple structures in the case where both antagonistic and agonistic binding modes need to be investigated, resulting in 85 total off-target structures. Of the 85 structures, 22 correspond to the “priority” (high-risk) panel of 17 targets, while the remaining 63 cover the other targets in Brennan-77. We denote the 22-target subset as SelectBench-Safety-Priority and the full set as SelectBench-Safety-All.
This panel enables evaluation of selectivity with respect to safety-related off-targets for any user-defined on-target protein of interest (POI). To standardize benchmarking, we selected five commonly used POIs from prior ML work: PARP1 (Poly [ADP-ribose] polymerase-1), FA7 (Coagulation factor VII), 5HT1B (5-hydroxytryptamine receptor 1B), BRAF (Serine/threonine-protein kinase B-raf), and JAK2 (Tyrosine-protein kinase JAK2) [70, 43]. This set enables direct comparison with existing benchmarks while introducing selectivity as an additional challenge. To improve reproducibility, we provide a fully specified and open-sourced protein structure preparation pipeline, addressing inconsistencies in prior work (e.g., faulty 5HT1B PDBQT structures). Finally, to avoid overlap between on- and off-targets, 5HT1B is excluded from the Brennan-77 panel in our benchmark.
SelectBench-XL: Scaling up structure-based small-molecule selectivity assessment
Most SBDD generative models are trained on a curated CrossDocked2020 dataset [25, 49], which contains 100,000 protein–ligand pairs for training and 100 proteins for testing under a 30% sequence-identity split. We built a large-scale selectivity benchmark from this test set by considering all non-identical protein pocket pairs and computing their structural similarity using ProBiS [42]. Based on the resulting Z-scores (shown in Figure A3), we stratified pairs into easy (Z<2, n=9834), medium (2≤Z<3, n=38), and hard (Z≥3, n=28) categories, with higher similarity corresponding to more challenging selectivity cases.
Figure A3. Heatmaps of structural similarity between POIs and the off-targets panel in SELECTBENCH-Safety. Each row corresponds to an on-target, each column corresponds to an off-target, and color intensity reflects the ProBiS Z-score, with darker shades indicating higher local structural similarity. ProBiS Z-score of zero or below are set to white. Priority targets of SELECTBENCH-Safety-Priority are highlighted by the yellow box, the HTR1B off-target structure excluded due to overlap with POI 5HT1B is highlighted by the blue box.
Evaluation Pipeline
Overview
We begin with an overview of the SelectBench evaluation protocol. For a given generative SBDD model, we evaluate performance across all target pairs described in § 3. For each pair, the model generates 100 valid candidate molecules conditioned on the on-target 3D protein pocket (see Sec. E for more details). Inference-time guidance and selectivity-aware methods (e.g., CMD-GEN) use off-target pocket as additional conditioning input.
All generated molecules are then scored against both the on- and off-target protein pockets using the scoring functions described in § 4.3. These scores are aggregated into selectivity-focused evaluation metrics outlined in § 4.2. The set of evaluated baselines is listed in § 5.1, and the overall pipeline is illustrated in Figure 1.
Selectivity metrics
We define our selectivity metrics in terms of a generic scoring function f, where s=f(ligand,pocket) denotes the predicted binding score for a given ligand–pocket pair. Concrete choices of scoring functions are described in § 4.3.
Delta-score Δ (per molecule). We define the delta-score as the difference between on- and off-target scores [27] from a given scoring function, with the sign chosen such that higher values indicate better selectivity:
Δ(i)={soff(i)−son(i),son(i)−soff(i),for scores where lower is better (e.g., docking energies)for scores where higher is better (e.g., predicted binding probabilities).
Under this convention, Δ(i)>0 indicates that the ligand binds more strongly to the on-target than the off-target, reflecting selectivity toward the intended pocket. However, small differences may not correspond to meaningful selectivity; we therefore also report a stricter margin criterion Δ(i)>τ. Choices of τ are explained for each scoring function. We report the mean on-target score for a population of ligands against a pocket as sˉon.
Scenario I: one on-target vs. one off-target
For SELECTBENCH-Case and SELECTBENCH-XL, each evaluation case consists of a single on-target and a single off-target. For each method, we consider the set of Nvalid generated ligands that are valid and successfully docked against both pockets. We report the following set-level statistics:
Mean Δ-score[71], Δˉ=Nvalid1∑i=1NvalidΔ(i): average selectivity gap.
Top-kΔ-score, Δˉtop−k: mean of the k largest (most selective) Δ(i) values, evaluating the best-performing candidates.
Selectivity rate[71], SRτ=Nvalid1∑i=1Nvalid1(Δ(i)>τ): fraction of ligands exhibiting any preference for the on-target by a margin of τ.
Threshold-based selectivity rate, SRτon,τoff=Nvalid1∑i=1Nvalid1(son(i)>τon∧soff(i)<τoff): fraction of ligands that achieve high-confidence binding to the on-target while avoiding the off-target, based on absolute score thresholds, if known.
Scenario II: one on-target vs. many off-targets
The SELECTBENCH-Safety setting evaluates each ligand against a panel of T anti-targets {t1,…,tT} representing common off-target liabilities (we use T=22 for the priority panel and T=84 for the full panel). For the i-th ligand, we define the delta-score with respect to anti-target tj as Δtj(i)=stj(i)−son(i), and the panel mean as sˉoff(i)=T1∑j=1Tstj(i). We then aggregate as follows:
Mean Δ-score[71], Δˉ=Nvalid1∑i=1Nvalid(sˉoff(i)−son(i)): average selectivity gap against entire panel.
Worst Δ-score per molecule, Δˉworst−mol=Nvalid1∑i=1NvalidΔworst−mol(i), where Δworst−mol(i)=minjΔtj(i): least selective (worst-case) anti-target for ligand i. This penalizes any single panel failure and is particularly relevant for safety screening.
Average selectivity rate, SRavg=Nvalid⋅T1∑i=1Nvalid∑j=1T1(Δtj(i)>τ): fraction of ligands exhibiting preference for the on-target against the anti-target tj, averaged across all anti-targets. This setting is tolerant to occasional misses.
Worst selectivity rate per molecule, SRworst−mol: fraction of ligands satisfying Δworst−mol(i)>τ; i.e., ligands that outperform all anti-targets by at least margin τ.
Top-20% ratio[71], R20: fraction of ligands whose on-target score ranks among the top 20% of all T+1 docking scores (on-target plus panel). This metric is invariant to the absolute scale and complements threshold-based rates.
Scoring functions
Docking-based evaluation
For each generated ligand, we perform rigid molecular docking against both the intended (on-target) protein and a set of competing (off-target) proteins using AutoDock Vina [21]. Given a protein structure and a defined binding site, molecular docking first predicts the most energetically favorable binding pose of a ligand in complex with that protein. This corresponds to the set of 3D coordinates of each atom in the ligand relative to the target. The molecular docking scoring function estimates the associated binding energy of this pose, returning a score which reflects how strongly the ligand is predicted to bind to the target [54]. Lower (more negative) values indicate stronger predicted binding. By comparing scores across on- and off-target proteins, we assess whether a ligand preferentially binds to the intended target. Additional details regarding docking parameters are in Sec. C.1. We report metrics for margins τ∈{0,1} kcal/mol. We select τ=1 kcal/mol based on the typical accuracy limits of state-of-the-art binding free energy prediction methods [3, 5].
Z-score normalization. One key challenge with docking scores is that they are not calibrated across different proteins due to differences in the pocket structure and chemistry, so scores may not be directly comparable between targets [46]. To remedy this, we normalize binding scores using Z-score normalization [41, 48]. For each target pocket, we compute the mean μref and standard deviation σref of docking scores over a reference set of 128 drug-like ligands from [18]. Given a generated ligand i with its binding score s(i), we define its Z-score as
Z(i)=σrefs(i)−μref.
This metric quantifies how the predicted binding strength of a generated ligand compares to the reference distribution for a given pocket, enabling better comparisons across targets. We report metrics for margins τ∈{0,1}. Setting τ=1 indicates whether the molecule is one standard deviation more favourable for the on-target than for the off-target, relative to typical molecules for each target. We denote metrics computed from Z-normalized scores with the superscript Δˉ (e.g., Δˉz, SRτ=1z). Positive control experiments that demonstrate the predictive power of docking-based Z-score normalized metrics can be found in Sec. C.2.
Predicted binding probability with Boltz-2
In addition to docking, we evaluate binding using Boltz-2 [56], a learned co-folding model that predicts protein–ligand structure and associated binding metrics from the protein sequence and ligand representation. For each ligand–target pair, Boltz-2 produces a predicted binding probability and an affinity estimate. Note that we constrain co-folding to occur with the predefined protein binding pocket. We use the predicted binding probability as our primary score, as prior work has shown that affinity predictions exhibit weaker correlation with experimental binding measurements [14].
Ranking. One challenge with Boltz-2 predicted binding probabilities is that they are not calibrated across targets and may be influenced by target-specific biases. Thus, we primarily rely on ranking-based evaluations, which was shown to have good discriminative power [14]. For each target, we convert predicted binding probabilities to percentile-normalized ranks computed over all evaluated ligands for a given target. We indicate metrics computed from ranks with the superscript rank (e.g., SRτon=0.6,τoff=0.4rank is the fraction of ligands whose percentile ranks are in the top 60% for the on-target and bottom 40% for the off-target).
Results
We apply the evaluation protocol described in § 4.1 to all models in § 5.1 across the SELECTBENCH tasks. Following the research questions outlined in § 1, we evaluate selective molecular design across tasks by analyzing (i) standard SBDD generative models, (ii) selectivity-focused models, and (iii) inference-time guidance. Results are presented through case studies (§ 5.2), safety analyses (§ 5.3), and SB-XL experiments (§ 5.4).
Model baseline selection
We benchmark a diverse set of SBDD models spanning multiple modeling types, including diffusion/flow-based methods, autoregressive models, and methods that incorporate selectivity. A summary of all baselines is provided in Table A7; detailed descriptions are provided in Sec. D, and additional RL-based methods are reported in Sec. F.1. We also evaluate two inference-time guidance strategies on compatible models: classifier-free guidance (CFG) and Feynman-Kac Correctors (FKC). Further descriptions on these guidance methods are provided in Sec. D.3.
References
[1]J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. J. Ballard, J. Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature, 630(8016):493–500, 2024.DOI
[2]M. Akhtar, O. Benjel·loun, C. Conforti, L. Foschini, P. Gijsbers, J. Giner-Miguelez, S. Goswami, N. Jain, M. Karamousadakis, S. Krishna, et al. Croissant: A metadata format for ml-ready datasets. Advances in Neural Information Processing Systems, 37:82133–82148, 2024.
[3]S. K. Albanese, J. D. Chodera, A. Volkmer, S. Keng, R. Abel, and L. Wang. Is structure-based drug design ready for selectivity optimization? Journal of chemical information and modeling, 60(12):6211–6227, 2020.DOI
[4]V. Albert, D. E. Skinner, S. Karnwal, L. N. Guli, M. Vazquez, O. Ledezma, R. Shaffer, A. Celniker, A. De La Cruz, M. Caliendo, et al. Comprehensive analysis of clinically discontinued compounds using an in vitro secondary pharmacology panel to predict potential safety risks during drug development. ACS Pharmacology & Translational Science, 8(9):3314–3333, 2025.DOI
[5]M. Aldeghi, A. Heifetz, M. J. Bodkin, S. Knapp, and P. C. Biggin. Predictions of ligand selectivity from absolute binding free energy calculations. Journal of the American Chemical Society, 139(2):946–957, 2017.DOI
[6]A. N. Aquino-Acevedo, J. A. Orengo-Orengo, M. E. Cruz-Robles, and H. I. Saavedra. Mitotic kinases are emerging therapeutic targets against metastatic breast cancer. Cell Div., 19(1):21, June 2024.
[7]A. Bateman, M.-J. Martin, S. Orchard, M. Magrane, A. Adesina, S. Ahmad, E. H. Bowler-Barnett, H. Bye-A-Jee, D. Carpentier, P. Denny, et al. Uniprot: the universal protein knowledge-base in 2025. Nucleic acids research, 53(D1):D609–D617, 2025.
[8]B. J. Bender, S. Gahbauer, A. Luttens, J. Lyu, C. M. Webb, R. M. Stein, E. A. Fink, T. E. Balijus, J. Carlsson, J. J. Irwin, et al. A practical guide to large-scale docking. Nature protocols, 16(10):4799–4832, 2021.DOI
[9]H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov, and P. E. Bourne. The protein data bank. Nucleic acids research, 28(1):235–242, 2000.
[10]G. R. Bickerton, G. V. Paolini, J. Besnard, S. Muresan, and A. L. Hopkins. Quantifying the chemical beauty of drugs. Nature chemistry, 4(2):90–98, 2012.DOI
[11]A. Bou, M. Thomas, S. Dittert, C. Navarro, M. Majewski, Y. Wang, S. Patel, G. Tresadern, M. Ahmad, V. Moens, W. Sherman, S. Sciabola, and G. De Fabritiis. ACEGEN: Reinforcement learning of generative chemical agents for drug discovery. Journal of Chemical Information and Modeling, 64(15):5900–5911, 2024.arxiv.org/abs/2405.04657
[12]J. Bowes, A. J. Brown, J. Hamon, W. Jarolimek, A. Sridhar, G. Waldron, and S. Whitebread. Reducing safety-related drug attrition: the use of in vitro pharmacological profiling. Nature reviews Drug discovery, 11(12):909–922, 2012.DOI
[13]R. J. Brennan, S. Jenkinson, A. Brown, A. Delaunois, B. Dumotier, M. Pannirselvam, M. Rao, L. Ribeiro Rosenbrier, F. Schmidt, A. Sibony, et al. The state of the art in secondary pharmacology and its impact on the safety of new medicines. Nature Reviews Drug Discovery, 23(7):525–545, 2024.
[14]G. Bret, F. Sindt, and D. Rognan. Assessing boltz-2 performance for the binding classification of docking hits. Journal of Chemical Information and Modeling, 66(3):1511–1521, 2026.
[15]M. Buttenschoen, G. M. Morris, and C. M. Deane. Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chemical Science, 15(9):3130–3139, 2024.
[16]S. Choi, H. Kim, C. Park, D. Lee, S. Lee, Y. Kim, H. Park, S. Kwon, Y. Jo, and S. Park. Controllable 3d molecular generation for structure-based drug design through bayesian flow networks and gradient integration. arXiv preprint arXiv:2508.21468, 2025.
[17]S. Choi, S. Seo, B. J. Kim, C. Park, and S. Park. Pidiff: Physics informed diffusion model for protein pocket-specific 3d molecular generation. Computers in Biology and Medicine, 180:108865, 2024.
[18]A. Comajuncosa-Creus, G. Jorba, X. Barril, and P. Aloy. Comprehensive detection and characterization of human druggable pockets through binding site descriptors. Nature communications, 15(1):7917, 2024.
[19]S.-X. Dai, W.-X. Li, G.-H. Li, and J.-F. Huang. Proteome-wide prediction of targets for aspirin: new insight into the molecular mechanism of aspirin. PeerJ, 4:e1791, 2016.
[20]R. Douc, O. Cappé, and E. Moulines. Comparison of resampling schemes for particle filtering, 2005.
[21]J. Eberhardt, D. Santos-Martins, A. F. Tillack, and S. Forli. Autodock vina 1.2. 0: new docking methods, expanded force field, and python bindings. Journal of chemical information and modeling, 61(8):3891–3898, 2021.
[22]M. Eguida and D. Rognan. Estimating the similarity between protein pockets. International Journal of Molecular Sciences, 23(20):12462, 2022.
[23]A. R. Elmore, A. Sadik, L. Paternoster, G. M. Khandaker, T. R. Gaunt, and G. Hemani. Genetic inference of on-target and off-target side-effects of antipsychotic medications. PLoS Genet., 21(7):e1011793, July 2025.
[24]P. Ertl and A. Schuffenhauer. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of cheminformatics, 1(1):8, 2009.
[25]P. G. Francoeur, T. Masuda, J. Sunseri, A. Jia, R. B. Iovanisci, I. Snyder, and D. R. Koes. Three-dimensional convolutional neural networks and a cross-docked data set for structure-based drug design. Journal of chemical information and modeling, 60(9):4200–4215, 2020.
[26]J. S. Fraser, S. Edgar, L. N. Handly, S. Kosuri, J. D. Chodera, M. Murcko, and W. P. Walters. Mapping the avoid-ome: a systematic open-science approach to predictive admet. Nature Communications, 17(1):4644, 2026.
[27]B. Gao, M. Ren, Y. Ni, Y. Huang, B. Qiang, Z.-M. Ma, W.-Y. Ma, and Y. Lan. Rethinking specificity in sbdd: Leveraging delta score and energy-guided diffusion. arXiv preprint arXiv:2403.12987, 2024.arxiv.org/abs/2403.12987
[28]J. García-Cárceles, E. Caballero, C. Gil, and A. Martínez. Kinase inhibitors as underexplored antiviral agents. J. Med. Chem., 65(2):935–954, Jan. 2022.
[29]G. Gintant, P. T. Sager, and N. Stockbridge. Evolution of strategies to improve preclinical cardiac safety testing. Nature reviews Drug discovery, 15(7):457–471, 2016.
[30]J. Guan, W. W. Qian, X. Peng, Y. Su, J. Peng, and J. Ma. 3d equivariant diffusion for target-aware molecule generation and affinity prediction. In International Conference on Learning Representations, 2023.
[31]J. Guan, X. Zhou, Y. Yang, Y. Bao, J. Peng, J. Ma, Q. Liu, L. Wang, and Q. Gu. Decompdiff: Diffusion models with decomposed priors for structure-based drug design. In International Conference on Machine Learning, pages 11827–11846. PMLR, 2023.
[32]C. Harris, K. Didi, A. R. Jamash, C. K. Joshi, S. V. Mathis, P. Lio, and T. Blundell. Benchmarking generated poses: How rational is structure-based drug design with generative models? arXiv preprint arXiv:2308.07413, 2023.arxiv.org/abs/2308.07413
[33]J. Ho and T. Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021.arxiv.org/abs/2207.12598
[34]I. House, M. Valore-Caplan, E. Maris, and G. S. Falchook. Cyclin dependent kinase 2 (CDK2) inhibitors in oncology clinical trials: A review. J. Immunother. Precis. Oncol., 8(1):47–54, Feb. 2025.DOI
[35]J. Huang and D. Zhang. Molform: multi-modal flow matching for structure-based drug design. In ICML 2025 Generative AI and Biology (GenBio) Workshop, 2025.arxiv.org/abs/2507.05503
[36]D. J. Huggins, W. Sherman, and B. Tidor. Rational approaches to improving selectivity in drug design. Journal of medicinal chemistry, 55(4):1424–1444, 2012.DOI
[37]X. Jalencas and J. Mestres. On the origins of drug polypharmacology. MedChemComm, 4(1):80–87, 2013.DOI
[38]Y. Jiang, G. Zhang, J. You, H. Zhang, R. Yao, H. Xie, L. Zhang, Z. Xia, M. Dai, Y. Wu, et al. Pocketflow is a data-and-knowledge-driven structure-based molecular generative model. Nature Machine Intelligence, 6(3):326–337, 2024.DOI
[39]J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Židek, A. Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589, 2021.DOI
[40]M. W. Karaman, S. Herrgard, D. K. Treiber, P. Gallant, C. E. Atteridge, B. T. Campbell, K. W. Chan, P. Ciceri, M. I. Davis, P. T. Edeen, et al. A quantitative analysis of kinase inhibitor selectivity. Nature biotechnology, 26(1):127–132, 2008.DOI
[41]S. S. Kim, M. L. Aprahamian, and S. Lindert. Improving inverse docking target identification with z-score selection. Chemical biology & drug design, 93(6):1105–1116, 2019.DOI
[42]J. Konc and D. Janežič. Probis algorithm for detection of structurally similar protein binding sites by local structural alignment. Bioinformatics, 26(9):1160–1168, 2010.DOI
[43]S. Lee, J. Jo, and S. J. Hwang. Exploring chemical space with score-based out-of-distribution generation. In International Conference on Machine Learning, pages 18872–18892. PMLR, 2023.arxiv.org/abs/2206.07632
[44]C. Lipinski and F. DB. Pj, 1997. experimental and computational approaches to estimate solubility and permeability in drug discovery and development settings. Adv. Drug Deliv. Rev, 23:5–25.
[45]C. A. Lipinski. Lead-and drug-like compounds: the rule-of-five revolution. Drug discovery today: Technologies, 1(4):337–341, 2004.DOI
[46]H. Liu, B. Hu, P. Chen, X. Wang, H. Wang, S. Wang, J. Wang, B. Lin, and M. Cheng. Docking score ml: Target-specific machine learning models improving docking-based virtual screening in 155 targets. Journal of Chemical Information and Modeling, 64(14):5413–5426, 2024.DOI
[47]H. H. Loeffler, J. He, A. Tibo, J. P. Janet, A. Voronov, L. H. Mervin, and O. Engkvist. Reinvent 4: Modern AI-driven generative molecule design. Journal of Cheminformatics, 16(1):20, 2024.
[48]Q. Luo, L. Zhao, J. Hu, H. Jin, Z. Liu, and L. Zhang. The scoring bias in reverse docking and the score normalization strategy to improve success rate of target fishing. PLoS One, 12(2):e0171433, 2017.DOI
[49]S. Luo, J. Guan, J. Ma, and J. Peng. A 3d generative model for structure-based drug design. Advances in Neural Information Processing Systems, 34:6229–6239, 2021.arxiv.org/abs/2203.10446
[50]J. J. Lynch III, T. R. Van Vleet, S. W. Mittelstadt, and E. A. Blomme. Potential functional and pathological side effects related to off-target pharmacological activity. Journal of pharmacological and toxicological methods, 87:108–126, 2017.DOI
[51]C. Mo, N. Wei, T. Li, M. Ahmed Bhat, M. Mohammadi, and C. Kuang. CDK9 inhibitors for the treatment of solid tumors. Biochem. Pharmacol., 229(116470):116470, Nov. 2024.DOI
[52]N. Y. Ngoi, M. M. Pham, D. S. Tan, and T. A. Yap. Targeting the replication stress response through synthetic lethal strategies in cancer medicine. Trends in cancer, 7(10):930–957, 2021.DOI
[53]M. Olivecrona, T. Blaschke, O. Engkvist, and H. Chen. Molecular de-novo design through deep reinforcement learning. Journal of Cheminformatics, 9(1):48, 2017.
[54]J. M. Paggi, A. Pandit, and R. O. Dror. The art and science of molecular docking. Annual review of biochemistry, 93(1):389–410, 2024.DOI
[55]S. M. Papidocha, A. Burger, V. Bernales, and A. Aspuru-Guzik. The elephant in the lab: synthesizability in generative small-molecule design. Current Opinion in Chemical Engineering, 51:101217, 2026.DOI
[56]S. Passaro, G. Corso, J. Wohlwend, M. Reveiz, S. Thaler, V. R. Somnath, N. Getz, T. Portnoi, J. Roy, H. Stark, et al. Boltz-2: Towards accurate and efficient binding affinity prediction. BioRxiv, 2025.DOI
[57]X. Peng, S. Luo, J. Guan, Q. Xie, J. Peng, and J. Ma. Pocket2mol: Efficient molecular sampling based on 3d protein pockets. In International conference on machine learning, pages 17644–17655. PMLR, 2022.arxiv.org/abs/2205.07249
[58]P. G. Pilié, C. M. Gay, L. A. Byers, M. J. O’Connor, and T. A. Yap. Parp inhibitors: extending benefit beyond brca-mutant cancers. Clinical Cancer Research, 25(13):3759–3771, 2019.DOI
[59]M. Ragoza, T. Masuda, and D. R. Koes. Generating 3d molecules conditional on receptor binding sites with deep generative models. Chemical science, 13(9):2701–2713, 2022.DOI
[60]V. W. Rebecca, G. M. Alicea, K. H. T. Paraiso, H. Lawrence, G. T. Gibney, and K. S. M. Smalley. Vertical inhibition of the MAPK pathway enhances therapeutic responses in NRAS-mutant melanoma. Pigment Cell Melanoma Res., 27(6):1154–1158, Nov. 2014.DOI
[61]D. G. Rudmann. On-target and off-target-based toxicologic effects. Toxicologic pathology, 41(2):310–314, 2013.DOI
[62]D. Santos-Martins, Y. He, J. Eberhardt, P. Sharma, N. Bruciaferri, M. Holcomb, M. A. Llanos, A. Hansel-Harris, A. P. Barkdull, A. F. Tillack, et al. Meeko: Molecule parametrization and software interoperability for docking and beyond. Journal of Chemical Information and Modeling, 65(24):13045–13050, 2025.DOI
[63]A. Schneuing, C. Harris, Y. Du, K. Didi, A. Jamasb, I. Igashov, W. Du, C. Gomes, T. L. Blundell, P. Lio, et al. Structure-based drug design with equivariant diffusion models. Nature Computational Science, 4(12):899–909, 2024.DOI
[64]P. Shapiro. A promiscuous kinase inhibitor reveals secrets to cancer cell survival. Journal of Biological Chemistry, 294(21):8674–8675, 2019.DOI
[65]M. Skreta, T. Akhound-Sadegh, V. Ohanesian, R. Bondesan, A. Aspuru-Guzik, A. Doucet, R. Brekelmans, A. Tong, and K. Neklyudov. Feynman-kac correctors in diffusion: Annealing, guidance, and product of experts. In Proceedings of the 42nd International Conference on Machine Learning, 2025.arxiv.org/abs/2503.02819
[66]J. L. Stevens. Future of toxicology mechanisms of toxicity and drug safety: where do we go from here? Chemical research in toxicology, 19(11):1393–1401, 2006.pubmed.ncbi.nlm.nih.gov/17112225
[67]D. Sun, W. Gao, H. Hu, and S. Zhou. Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B, 12(7):3049–3062, 2022.DOI
[68]D. R. Weiss, J. Karpiak, X.-P. Huang, M. F. Sassano, J. Lyu, B. L. Roth, and B. K. Shoichet. Selectivity challenges in docking screens for gpcr targets and antitargets. Journal of medicinal chemistry, 61(15):6830–6845, 2018.
[69]S. A. Wildman and G. M. Crippen. Prediction of physicochemical parameters by atomic contributions. Journal of chemical information and computer sciences, 39(5):868–873, 1999.DOI
[70]S. Yang, D. Hwang, S. Lee, S. Ryu, and S. J. Hwang. Hit and lead discovery with exploratory rl and fragment-based molecule generation. Advances in Neural Information Processing Systems, 34:7924–7936, 2021.arxiv.org/abs/2110.01219
[71]P. Zhang, R. Han, X. Kong, T. Chen, and J. Ma. Speclig: Energy-guided hierarchical model for target-specific 3d ligand design. bioRxiv, pages 2025–11, 2025.DOI
[72]Y. Zhang, J.-G. Yu, and W. Wen. Recent advances in representative small-molecule DRD2 inhibitors: Synthetic routes and clinical applications. Eur. J. Med. Chem., 277(116731):116731, Nov. 2024.DOI
[73]Y. Ziv, F. Imrie, B. Marsden, and C. M. Deane. Molsnapper: conditioning diffusion for structure-based drug design. Journal of Chemical Information and Modeling, 65(9):4263–4273, 2025.DOI
[74]Y. Zou, T. Guo, Z. Fu, Z. Guo, W. Bo, D. Yan, Q. Wang, J. Zeng, D. Xu, T. Wang, et al. A structure-based framework for selective inhibitor design and optimization. Communications Biology, 8(1):422, 2025.DOI