Molecular generation aims to construct valid, diverse, and novel molecular graphs by transforming noise into structured atom and bond representations through reverse diffusion. Because each reverse step relies on accurate denoising, the denoising network must recover both local chemical connectivity and global graph topology across noise levels. We propose DFDM, a dynamic fusion diffusion model for 2D molecular graph generation. DFDM combines a GINE-based spatial branch with a Chebyshev spectral branch and uses signal-to-noise-ratio-guided, layer-specific weights to adapt their contributions during reverse generation. Across three independent seeded generation-and-evaluation runs on QM9 and ZINC250k, DFDM achieved the lowest mean Frechet ChemNet Distance and the highest mean validity without correction among the compared methods. Ablations on both datasets showed that dynamic fusion provided a better overall metric balance than fixed weighting, while QM9 noise-stratified analysis revealed a systematic transition from spectral emphasis at high noise to spatial emphasis near the final denoising stage. DFDM therefore links noise-dependent denoising with molecular graph quality by integrating complementary local and global representations throughout generation.
Perfluorooctane sulfonate (PFOS), a persistent environmental pollutant, has been implicated in metabolic dysfunction-associated steatotic liver disease (MASLD), yet the underlying molecular mechanisms remain incompletely characterized. We constructed an integrated analytical framework combining population epidemiology, network toxicology, machine learning, transcriptomic analysis, single-cell mapping, molecular docking, and experimental validation. NHANES data (n = 1,834) were analyzed to assess the association between serum PFOS concentrations and FLI-defined MASLD. Machine-learning analyses using LASSO and SVM-RFE prioritized candidate genes from 874 overlapping PFOS-MASLD-associated genes. Single-cell RNA sequencing resolved cell-type-specific expression patterns, while molecular docking evaluated potential PFOS-protein interactions. A 12-week murine exposure model provided in vivo validation. Epidemiological analysis identified a nonlinear association between serum PFOS and MASLD odds (p < 0.001), with effects evident at background exposure levels (7.76 ng/mL). Convergent machine-learning analyses prioritized five candidate genes: CYP7A1, GRIA3, PHLDA1, SOCS2, and WNT5A. An exploratory five-gene model yielded an apparent AUC of 0.998 (95
Deep learning has greatly advanced de novo protein design, yet its application to rational short peptide design remains underexplored. Here, we developed SPB-Seeker (Short Peptide Binder Seeker), an integrated pipeline combining deep learning–based generative models with computational chemistry screening to discover dual-target short peptide inhibitors. Using penicillin-binding proteins PBP2b and PBP2x from drug-resistant Streptococcus pneumoniae as targets, AFDesign, RFdiffusion, and BoltzGen were employed to generate an initial library of 1101 candidate sequences. Subsequently, ESM2 was employed to extract sequence embeddings for diversity analysis, which revealed distinct algorithmic biases among the three generative models, and was then used as the feature extractor of a prediction framework for early-stage toxicity screening. Candidates were further prioritized through molecular docking, tiered molecular dynamics simulations, and MM/PB(GB)SA binding free energy calculations. Three peptides, AFD1, BG3, and RFD2, showed high binding stability, with BG3 displaying the strongest dual-target binding, achieving binding free energies of − 52.777 kcal/mol for PBP2b and − 74.071 kcal/mol for PBP2x. Interestingly, quantum chemical calculations using cluster model and the Interaction Region Indicator (IRI) method analyses indicated that BG3 adopts a stable cyclic-like conformation when bound to PBP2x, driven by proline-induced turns, intramolecular hydrogen bonds, and terminal C–H···π interactions. Overall, SPB-Seeker provides an extensible computational framework for targeted short peptide binder discovery and offers a basis for subsequent affinity optimization and stability enhancement.
KasA is an essential enzyme of Mycobacterium tuberculosis (Mtb). It plays a critical role in synthesizing long-chain mycolic acids, the major components of the bacterial cell wall, by regulating the FAS-I and FAS-II fatty acid synthesis pathways. Inhibiting KasA offers a promising strategy for treating tuberculosis (TB). This study used fragment-based drug design (FBDD) to design novel small molecules targeting KasA. Fragments from known KasA inhibitors were generated with the MacFrag tool and then combined with Fragmenstein to create potential hit compounds. A multi-tiered molecular docking approach was used to evaluate their binding affinity and interactions with KasA. Selected candidates underwent pharmacokinetic analysis and molecular dynamics (MD) simulations. Five promising molecules, namely KasA_FB1, KasA_FB2, KasA_FB3, KasA_FB4 and KasA_FB5, were identified. Their molecular docking binding energies were − 7.80, − 8.30, − 9.00, − 7.80, and − 9.00 kcal/mol, respectively, all superior to the reference co-crystal ligand TLM (− 7.20 kcal/mol). MD simulations showed that their dynamic stability was comparable to or better than TLM. MM-GBSA and free energy perturbation (FEP) analyses further confirmed their superior binding affinity for KasA. These compounds represent promising candidates for the development of new anti-TB drugs targeting KasA and warrant experimental validation.
Topoisomerase I (TOP1) is a crucial anticancer target, but the development of traditional TOP1 inhibitors suffers from long research cycles, high costs, and low success rates. Existing artificial intelligence (AI)-driven studies lack systematic comparisons of molecular fingerprints and algorithms, as well as user-friendly predictive application tools. To address these gaps, this study retrieved TOP1 inhibitor activity data from the ChEMBL database, integrated five types of molecular fingerprints (AtomPairs, MACCS, Morgan, PharmacoPFP, and RDKitDes), and constructed and compared classical machine learning (ML) models and deep learning (DL) models, resulting in a total of 40 models. The four top-performing models, SVM::Morgan, RF::Morgan, DNN::MACCS, and KNN::Morgan, achieved ROC–AUC values of 0.93–0.94 under random splitting. Y-scrambling supported that the models learned non-random structure–activity relationships, while SHAP analysis identified key molecular features. The URL of the developed web application is http://drugpred.top:5000 , and this application enables the prediction of TOP1 inhibitory activity via SMILES (Simplified Molecular-Input Line-Entry System) or molecular structure drawing. Additionally, standalone desktop applications (.exe) for offline prediction are freely available at https://github.com/zenghuang8006/TOP1-inhibitor-prediction . Screening of 189,554 SPECS compounds followed by in vitro validation identified AG60 and AI61 as potential TOP1 inhibitors hits, with inhibition rates of 64
Alzheimer’s disease is a multifactorial neurodegenerative disorder characterized by cholinergic dysfunction and neuroinflammation. Dual inhibition of acetylcholinesterase and monoacylglycerol lipase has emerged as a promising therapeutic approach. This study employed an integrative in silico workflow to identify potential dual acetylcholinesterase/monoacylglycerol lipase inhibitors from a molecular library derived from known inhibitors (rivastigmine, JZL-184, ABX-1431). A total of 365 compounds were screened via molecular docking, interaction-based filtering, ADME/toxicity prediction, and molecular dynamics simulations. Among them, compound H34 demonstrated a comparatively favorable overall computational profile, supported by MM/GBSA binding-energy estimates (ΔGbind = − 30.96 and − 37.34 kcal/mol) and comparatively favorable structural stability metrics (RMSD, RMSF, Rg, and SASA) in the MD simulations. Further ProLIF interaction mapping and free energy landscape analysis supported the persistent interaction profile and conformational behavior of the H34-protein complexes. Additionally, H34 displayed favorable pharmacokinetic properties and low predicted acute toxicity. These results highlight H34 as a promising dual-target candidate for Alzheimer’s disease therapy and illustrate the effectiveness of integrated computational strategies in early-stage drug discovery.
Predicting chemosensory attributes from chemical structures is a fundamental task in cheminformatics and molecular modeling. When applied to flavor specifically, this task becomes particularly challenging due to the structural diversity of flavor molecules and the complex, multi-label nature of human sensory perception. Traditional machine learning methods often rely on one-dimensional fingerprints, which inadequately capture high-dimensional topological and geometric features. In this study, we introduce the FlavorGraph Predictive Network (FlavorGPN), a novel graph neural network (GNN) framework for multi-label flavor prediction. FlavorGPN leverages the pretrained 2D graph encoder from GraphMVP, whose parameters are learned through 3D-informed pretraining, to enhance molecular graph representations. Notably, no explicit 3D conformers or atomic coordinates are used during downstream fine-tuning or inference. Therefore, the use of 3D information in this study should be understood as 3D-supervised pretraining rather than direct 2D/3D geometric integration during inference. To mitigate the class imbalance inherent in flavor datasets, we propose ML-ROS-improved, an adaptive oversampling algorithm that integrates dynamic thresholding for minority-label identification, weighted minority-label sampling, and constrained graph augmentation. We also systematically evaluate several graph augmentation strategies. Among them, Molecular Connectivity Index (MCI)-constrained augmentation achieves the highest observed Macro-F1 and Macro AUC-ROC scores. Across the FlavorMiner and FART benchmarks, FlavorGPN achieved the highest observed Macro-F1 and Macro AUC-ROC point estimates among the evaluated baselines under the reported experimental settings. On the FART benchmark, the model achieved a Macro-F1 score of 0.8542 and a Macro AUC-ROC score of 0.9796. Literature-based contextual comparisons further indicate that the unified model performs competitively on key flavor categories, including Sweet, Bitter, and Sour. However, these comparisons do not constitute controlled head-to-head evaluations. Overall, the benchmark results demonstrate the practical value of FlavorGPN for imbalanced multi-label chemosensory prediction under the evaluated settings and suggest its potential to support computational screening and molecular-level analyses of flavor-associated chemical properties.
Accurate identification of protein–small molecule binding sites is a fundamental problem in computational biology and drug discovery. Existing sequence-based methods lack explicit spatial awareness, while structure-based approaches often struggle to integrate long-range functional dependencies and semantic information, leading to limited generalization on low-similarity or sparsely annotated proteins. To address these challenges, we propose DSC-BSite, a dynamic–static collaborative multimodal graph learning framework for residue-level binding site prediction. First, a Static Global Sequence Encoding module captures multi-scale local patterns and long-range contextual dependencies from protein sequences. Second, a Gated Dual-Graph Dynamic Propagation (GDDP) module jointly models spatial geometric interactions and sequence-derived functional correlations using a dynamic spatial graph and an attention-guided sequence graph, enabling adaptive residue interaction modeling. Third, a PPI-guided Structural–Semantic Alignment (PSSA) pre-training strategy aligns structural representations with function-aware semantic embeddings, enhancing the biological expressiveness of structural features without requiring PPI information during inference. Experimental results on the UniProtSMB and SJC benchmark datasets demonstrate that DSC-BSite achieves competitive performance across multiple evaluation metrics, with particularly strong results in Recall on UniProtSMB and Precision and MCC on SJC.
Endocrine-disrupting chemicals (EDCs) are widely present in the environment and consumer products and may disturb thyroid hormone homeostasis. However, the molecular mechanisms linking EDC exposure to thyroid cancer progression remain unclear. This study integrated toxicity prediction, toxicogenomics, transcriptomic analysis, machine learning, molecular simulation, and experimental validation to identify EDC-related key targets in thyroid cancer. ADMETlab 3.0 was used to evaluate the potential toxicity of BPA, PFOA, DDT, BDE-209, TCDD, and DEHP, and compound-related genes were obtained from the CTD database. By integrating thyroid cancer-related genes from multiple disease databases, 1113 shared EDC-thyroid cancer targets were identified and were mainly enriched in PI3K-Akt, FoxO, and AGE-RAGE signaling pathways. Combined with differential expression analysis, machine learning identified a six-gene diagnostic model consisting of FN1, BCL2, CD44, CDKN1A, CTNNB1, and JUN. The Lasso + LDA model achieved an average AUC of 0.976 across the training cohort and three external validation cohorts. CD44 showed robust diagnostic performance, with AUC values of 0.950, 0.801, 0.878, and 0.938 in the training set, GSE27155, GSE29265, and GSE153659, respectively, and had the highest contribution in SHAP analysis. Immune infiltration, TCGA survival, and single-cell analyses indicated that CD44 was associated with the tumor immune microenvironment, cellular state changes, and prognosis. Molecular docking and 200 ns molecular dynamics simulations generated plausible docking poses of BPA, DEHP, and PFOA on CD44, with PFOA showing the most favorable predicted docking score. Experimental validation showed higher CD44 expression in thyroid cancer tissues and cells and increased CD44 expression following EDC exposure. CD44 knockdown attenuated EDC-associated increases in proliferation, colony formation, and migration. These findings identify CD44 as a candidate molecule associated with EDC-responsive malignant phenotypes in thyroid cancer.
Cyclin-dependent kinases 4 and 6 (CDK4/6) are pivotal regulators of the G1-to-S phase transition, and their dysregulation is a hallmark of numerous malignancies. Despite the clinical success of existing CDK4/6 inhibitors, there remains a persistent need for chemically diverse scaffolds with potent dual-target affinity. In this study, we developed and implemented a virtual screening workflow that synergistically integrates ligand-based machine learning with structure-based molecular docking. By benchmarking multiple ML algorithms against curated ChEMBL datasets (265 CDK4 inhibitors and 402 CDK6 inhibitors), a Bayesian Ridge regressor utilizing ECFP4 fingerprints was identified as the most predictive model, achieving cross-validated R2 values of 0.731 ± 0.022 for CDK4 and 0.721 ± 0.070 for CDK6. This optimized ML filter was deployed to prioritize a 22,823-compound library, followed by rigorous dual-target docking refinement. This strategy prioritized three candidate hits for biochemical evaluation, among which HY-18,623 showed potent dual inhibitory activity, with IC50 values of 3.5 nM against CDK4 and 17.4 nM against CDK6. Extensive 200-ns molecular dynamics simulations and binding free energy analyses elucidated that HY-18,623 achieves high-affinity binding through persistent hydrogen bonds with hinge residues Val96 (CDK4) and Val101 (CDK6). These findings demonstrate that our integrated computational funnel is a highly efficient tool for discovering potent kinase inhibitors and position HY-18,623 as a promising lead candidate for further therapeutic development in oncology.
Excitotoxicity is a core pathological mechanism underlying various neurological disorders, including depression. It is primarily driven by the overactivation of NMDA receptors and disruption of calcium homeostasis. However, effective therapeutic interventions targeting this process remain limited. In this study, aloe-emodin was shown to reverse NMDA-induced reduction in cell viability, apoptosis, and calcium overload in a dose-dependent manner, while also attenuating NMDA-induced autophosphorylation of CaMKII. In a chronic unpredictable mild stress (CUMS) model, aloe-emodin significantly ameliorated depression-like behaviors, suppressed inflammatory responses and oxidative stress, and downregulated the expression of cleaved-PARP. By conducting biotin pull-down coupled with liquid chromatography-tandem mass spectrometry (LC–MS/MS) analysis, cellular thermal shift assay (CETSA), drug affinity responsive target stability (DARTS) analysis, and surface plasmon resonance (SPR) analysis, RANBP9 was identified as a direct molecular target of aloe-emodin, with a binding affinity (KD) of 716 nM. Molecular docking and molecular dynamics simulations revealed that aloe-emodin selectively binds to the His332 residue of RANBP9. Aloe-emodin promoted the degradation of RANBP9 via the ubiquitin–proteasome pathway by facilitating the interaction between RANBP9 and the E3 ligase CHIP. After RANBP9 was knocked down, the protective effects of aloe-emodin on cell viability and apoptosis were significantly weakened, confirming that the anti-excitotoxic activity of aloe-emodin is RANBP9-dependent. These findings collectively demonstrate that aloe-emodin is a novel and potent RANBP9-targeting compound with antidepressant activity and suggest that RANBP9 may serve as a promising target for developing antidepressant drugs.
The recurring Nipah virus outbreaks and the lack of effective antiviral therapies, emphasize the urgent need for effective therapeutic interventions. Given the sporadic and unpredictable nature of NiV outbreaks, drug repurposing offers a time-efficient alternative to the development of novel antivirals. In this study, we leveraged machine learning (ML) techniques to accelerate the process of identifying potential therapeutic candidates. Several supervised ML models such as Support Vector Machines, Random Forest, Logistic Regression, Decision Tree, k-Nearest Neighbors, Artificial Neural Networks, and Ridge Classifier were implemented using publicly available NiV inhibitor datasets (viz. Anti-Nipah, NVIK, PubChem) as well as literature review (n = 211 compounds). Among these, the Random Forest model demonstrated highest predictive performance, achieving high accuracy on both training (95
The serotonin transporter (SERT) plays a pivotal role in inflammatory responses and is a central therapeutic target for depression. Consequently, identifying potent SERT inhibitors remains a high priority in early-stage drug discovery. In this study, we evaluated twelve regression models, among which LightGBM, Random Forest, and XGBoost exhibited superior predictive performance, yielding R2 values of 0.7138, 0.7001, and 0.6920, respectively. Leveraging these optimized machine learning models, we conducted a large-scale virtual screening of over 11.5 million compounds, identifying 24 promising candidates. Subsequent molecular dynamics (MD) simulations and MM/GBSA binding free energy calculations supported the structural stability and strong predicted binding affinities of three lead molecules: Z2215663922, 19,835,875, and Z310319934. Furthermore, ADMET profiling indicated generally acceptable pharmacokinetic properties, although potential hERG-related liabilities for certain candidates warrant further experimental scrutiny. Our findings provide structurally diverse scaffolds and a computational prioritization framework for early-stage discovery and optimization of SERT inhibitor candidates that merit subsequent experimental validation.
The HIF-1α/VHL protein–protein interaction regulates cellular hypoxic adaptation. Targeting this PPI offers therapeutic potential for ischemia, yet the structural diversity and direct cytoprotective applications of reported VHL ligands remain limited. Here, we developed a computational virtual-screening workflow informed by deep learning-based interface analysis. DDMut-PPI was used to nominate putative auxiliary interface residues for construction of an alternative pharmacophore model. Screening of 24,893 molecules followed by fluorescence-polarization testing identified four candidates with IC50 values below 10 μM. Cmpd16 (CAS 1072833–77-2; ixazomib) showed the highest measured affinity in this panel (IC50 = 0.41 μM). Cmpd16 showed no detectable loss of viability under the reported assay conditions and produced a VHL-dependent pattern of HIF-1α and hydroxylated HIF-1α stabilization. In an oxygen–glucose deprivation/reoxygenation model, 10 μM Cmpd16 improved endothelial-cell migration and tube formation and was associated with increased VEGF and GLUT1, reduced ROS accumulation, and reduced cleaved caspase−3. Three independently initialized 200-ns Desmond production simulations showed recurring Pro99 and His110 contacts but replica-dependent protein and ligand dynamics, supporting a computationally plausible rather than unique Cmpd16 orientation. The comparison between the two selected ten-compound panels remained exploratory (two-sided Fisher’s exact test, p = 0.0867), and residue causality requires experimental mutagenesis.
Peroxisome proliferator-activated receptor gamma (PPAR-γ) is a ligand-activated nuclear receptor involved in adipogenesis, glucose homeostasis, lipid metabolism, and inflammation, making it an important therapeutic target for metabolic disorders. However, the complex pharmacology of PPAR-γ presents significant challenges for rational drug discovery. In this study, we developed Meta-iPPAR, an integrative in silico framework combining stacked machine learning, molecular docking, and molecular dynamics (MD) simulations for the identification of PPAR-γ agonists. Meta-iPPAR was constructed using diverse SMILES-based molecular descriptors and multiple machine learning algorithms integrated through a stacking strategy. The proposed model achieved strong predictive performance on the independent test set, with an ACC of 0.926, AUC of 0.965, and MCC of 0.848. Scaffold analysis and SHAP interpretation further identified important chemotypes and molecular features associated with PPAR-γ activation. Large-scale virtual screening of more than 36,000 compounds from the natural product atlas identified three promising fungal-derived candidates. Subsequent docking and 300 ns MD simulations demonstrated stable binding conformations and favorable interactions with key residues in the PPAR-γ ligand-binding domain, comparable to known agonists and co-crystal ligands. Collectively, these findings suggest that Meta-iPPAR provides a reliable computational framework for screening and prioritizing potential PPAR-γ agonists in early-stage drug discovery. Future experimental validation and biological evaluation of the identified compounds are warranted. We anticipate that Meta-iPPAR will be an effective computational tool for screening and prioritizing potential compounds targeting PPAR-γ in the early stage of drug development pipelines.
Osteosarcoma is a primary bone tumor in adolescents and young adults, characterized by high chemotherapy resistance and poor prognosis. Ferroptosis, an iron‑dependent and lipid-peroxidation‑driven cell death, has become a key target to overcome chemoresistance and inhibit tumor progression. Natural products, with structural diversity, multi‑target regulation, and low toxicity, show great potential in ferroptosis‑based osteosarcoma therapy. This review summarizes the core molecular mechanisms of ferroptosis, focusing on the regulatory networks of Xc⁻-GSH-GPX4, Nrf2/HMOX1, p53, MAPK, and STAT3 pathways in osteosarcoma. It further categorizes natural products (flavonoids, terpenoids, alkaloids, naphthoquinones, and isothiocyanates) and discusses their targets and mechanisms in inducing ferroptosis. Current bottlenecks, including insufficient mechanistic validation, poor target specificity, limited clinical translation, and a lack of combination therapy strategies, are critically assessed. Future research directions are also proposed. This review aims to provide a theoretical basis and new insights for developing natural-product-based ferroptosis‑targeting drugs to address clinical treatment dilemmas in osteosarcoma.
Breast cancer (BC) represents a major public-health burden, and epidemiological evidence suggests a potential association with exposure to methyl 4-hydroxybenzoate (MEP), a widely-used cosmetic preservative and estrogen-mimicking endocrine-disrupting chemical. Nevertheless, the potential mechanisms underlying MEP-associated BC oncogenesis and progression remain poorly understood. BC-related targets were curated from CTD, GeneCards, and OMIM, whereas MEP-related targets were interrogated from ChEMBL, PharmMapper, and SEA using stringent filters. The intersecting targets informed subsequent protein-protein interaction network construction and molecular docking studies. Subsequently, consensus molecular subtypes of BC were derived by applying ten clustering algorithms to multi-omics data, which were subsequently employed in three machine learning algorithms to develop a consensus MEP-related signature (CMEPRS) for BC patients. Five core putative toxicological targets (HSP90AA1, CTNNB1, TP53, MYC, and EGFR) with critical regulatory roles in MEP-associated molecular alterations were identified. Based on these findings, we generated MEP-toxicity-related classifiers and the CMEPRS prognostic model, which may facilitate patient stratification and support personalized clinical management for BC patients. The high-CMEPRS patients displayed prominent infiltration of macrophages, myeloid-derived suppressor cells, and cancer-associated fibroblasts. Apart from lapatinib, the high-CMEPRS patients showed higher predicted sensitivity to most conventional chemotherapeutic drugs. This computational study provides preliminary insights into molecular alterations linked to MEP exposure and offers a feasible analytical framework for patient stratification and therapeutic-target exploration in breast cancer.
Accurate drug–target affinity (DTA) prediction is essential for virtual screening and computational drug discovery. However, existing benchmarks such as Davis and KIBA often suffer from imbalanced label distributions and potential sampling bias, which can lead to overestimated performance and limited generalization in realistic settings. In addition, integrating heterogeneous modalities such as molecular graphs, protein sequences, and chemical fingerprints remains challenging due to unstable cross-modal interactions and inconsistent representation scales. To address these issues, we first construct three curated datasets from the Therapeutic Target Database (TTD), namely TTD_IC50, TTD_EC50, and TTD_KI, using a stratified sampling and normalization strategy to improve distribution balance while maintaining biochemical diversity. These datasets are designed to better reflect real-world distribution shifts in DTA prediction. Based on these datasets, we propose SCAD-DTA, a geometry-aware multimodal learning framework that explicitly addresses instability in cross-modal fusion under distribution shift. The key idea is to stabilize multimodal interaction by jointly modeling adaptive fusion dynamics and representation geometry constraints. SCAD-DTA introduces a Dynamic Cross-Modal Attention mechanism that adaptively reweights modality contributions conditioned on sample-specific context, mitigating modality dominance. To further improve representation stability, a Spherical Constrained Projection module enforces unit-norm geometry in the latent space, reducing scale inconsistency across modalities. In addition, a Concept Alignment module maps fused representations into a learnable prototype space, enabling structured and interpretable modeling of drug–target interactions. Extensive experiments on six benchmark datasets show that SCAD-DTA achieves competitive or superior performance in most evaluation settings, with particular strength under cold-start and cross-distribution scenarios; however, its gains are less pronounced in some cold-start splits (e.g., cold-drug on KIBA, cold-target on Metz), which we discuss explicitly. The source code and datasets are publicly available at: https://github.com/xwtxbzz/SCAD-DTA.
Retrosynthetic planning is a core task in computer-aided synthesis design. Existing multi-step retrosynthesis methods mainly focus on route accessibility and search success rates. However, in practical synthesis, route quality depends partly on yield, whose importance may vary across applications. Existing methods often struggle to adjust their search strategies according to changing yield preferences. To address this problem, we propose Hyp-Retro, a hypernetwork-guided and preference-conditioned retrosynthetic planning model. Hyp-Retro first collects candidate-pool decision data through a yield-guided search process. It then constructs a hypernetwork-driven tree policy network conditioned on yield preference, allowing the model to adjust candidate-node scores under different preference settings. On this basis, online reinforcement learning is employed to fine-tune the policy network, further enhancing the model's long-horizon decision-making capability and preference alignment in realistic search environments. Experimental results show that Hyp-Retro outperforms comparative methods in terms of search success rate and route yield. Moreover, it can adaptively adjust its planning strategy under different yield preferences, thereby generating high-quality retrosynthetic routes that better satisfy target-specific requirements.
