Accurate prediction of patient-specific therapeutic response from pre-treatment transcriptomes is hindered by the scarcity of matched clinical response labels and post-treatment molecular profiles. Preclinical transfer-learning models can simulate drug-induced expression changes but are often hard to interpret and unstable, whereas knowledge-graph methods provide mechanistic context yet remain static and fail to capture drug-induced transcriptomic perturbation dynamics. We propose PREDIKTOR, a patient-centered multi-view framework that aligns a personalized network view with a transferable transcriptomic perturbation view to predict clinical drug response. For each patient, we construct an individualized gene regulatory network from tumor expression using DysRegNet and augment it with drug-target links from DrugBank; a graph neural encoder yields a drug-centric, mechanistically grounded embedding. In parallel, a frozen condition-specific gene-gene attention model pretrained on LINCS L1000 generates a simulated post-perturbation transcriptomic profile for the same patient-drug pair. We align the two views in a shared latent space via a CLIP-style contrastive objective with drug-context hard negatives, then concatenate the representations for end-to-end response classification. On TCGA, PREDIKTOR consistently outperforms state-of-the-art baselines under patient-, drug-, and tissue-split evaluations, and transfers zero-shot to the I-SPY2 trial, improving AUROC by 5.6

Abstract Introduction: The dysregulation of kinases is a major mechanism for cancer development and progression. Rapid advances in generative deep learning technologies now allow models to generate candidate kinase inhibitors. However, with over 500 human kinases, the major challenge is to design inhibitors that are highly selective for the oncogenic kinase without blocking other essential ones. Thus, all generated molecules need to be tested rigorously for interactions with specific residues and overall affinity to pockets of both target and off-target kinases. This task requires orchestrating multiple AI tools in a well-defined workflow, which is not trivial to researchers. Methods: To address the challenges, we developed DrugVLAB for Kinase Inhibitor Generation, a comprehensive Amazon cloud-based workflow, built through our collaboration, that enables human-in-the-loop search. Workflow: The workflow consists of more than 20 cutting edge AI tools. It begins with in-house fragment-based molecule generations (ICLR 2025). Users can specify requirements such as residue-atom interactions and drug-likeness filters using in-house tools (JCIM 2025, ICML 2025, ISMB 2025). Candidate molecules then undergo docking simulations with Autodock Vina, followed by additional filtering based on residue-atom interactions while considering docking pose. Finally, drug target affinity (DTA) can be rigorously checked using in-house tools (ICLR 2025, ISMB 2025). At this stage, molecules are ranked by DTA values for synthesis and evaluation. As molecules are generated with fragments, most of them are synthesizable. This will conclude the execution of one round of DrugVLABTM for Kinase. A unique and notable feature of our cloud system is to incorporate assay results of newly synthesized and evaluated molecules. Our system identifies fragments or substructures enriched in active and inactive molecules. With these new fragment sets, itinitiates the next round of assay-guided molecule generation. Our experience is that better, more active molecules are generated as the round goes on. Results: DrugVLAB can generate 3000 molecules on Amazon cloud in 2.5 hours and a complete round of evaluation can be done in 2.5 hours for a target kinase and five off-targets. Conclusion: Our system is implemented on Amazon cloud, enabling researchers around the world to generate and evaluate molecules as kinase inhibitors. DrugVLAB for kinase is designed in a modular way so that any newly developed AI tools can be incorporated easily and timely. Citation Format: Sun Kim, Bokyung hyerin kim, Park, Joonho Seong, Seokchol kim youngkuk, Hong, Changyun Cho, Heejoon Chae, Kyoung Jae Won. DrugVLAB for oncogenic kinase inhibitor generation [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 979.

Accurately identifying drug-target interactions (DTIs) is a critical step in drug discovery. While structure-based drug design methods demonstrate impressive docking prediction accuracy, their heavy computational demands and resource-intensive nature make them impractical for directly processing vast chemical spaces containing a large number of compounds. This limitation highlights the need for a computational tool that balance speed and accuracy to rank and filter potential drug candidates efficiently. In contrast, existing ligand-based drug design methods, which learn representations from diverse protein and molecule features, often fail to make consistent predictions on unseen data or external databases, limiting their applicability for ranking and filtering potential drug candidates accurately. To address these challenges, we propose EnsDTI, a novel framework that bridges the gap between structure-based and ligand-based drug design approaches. EnsDTI utilizes a mixture-of-experts architecture to enhance DTI predictions using existing deep learning models and incorporates an inductive conformal predictor to assess prediction quality with confidence scores, ensuring reliability. Experimental results on four widely used benchmark datasets show that EnsDTI consistently achieves high performance in both prediction accuracy and confidence estimation. In addition, its candidate rankings correlate well with actual docking affinities, suggesting its practical utility in drug discovery.

The simultaneous use of multiple medications is a common practice in disease treatment, yet the same drug combination can lead to different effects under varying physiological, pharmacological, or genomic conditions-collectively referred to as the 'context'. Accurately predicting the outcomes of drug combinations across diverse contexts, also known as drug relational learning (DRL), is essential for improving therapeutic efficacy and safety. Despite its importance, existing methods face two major challenges: they are often tailored to specific DRL tasks, lacking generalizability, and they fail to explicitly model the influence of context on drug interactions. This limitation arises because most methods focus primarily on whole-drug compound structures, overlooking the fine-grained atomic-level interactions critical for context-aware predictions. To address these challenges, we propose a novel context-aware hierarchical fusion architecture for DRL. By formulating the problem as the label prediction of drug-drug-context triplets, our approach explicitly models the interaction between drugs by first learning their intrinsic atomic-level interactions and then incorporating context into their embeddings at the atomic level through information fusion. Experiments across diverse tasks-such as synergy prediction, polypharmacy side effect detection, and drug-drug interaction prediction-demonstrate our model's capability to effectively capture context-aware information. Importantly, our method consistently achieves robust performance in highly complex scenarios, highlighting its adaptability and utility in advancing context-aware drug relational learning.

Deep learning models for drug–target interaction (DTI) prediction often achieve strong benchmark performance while relying on input patterns that are not captured by standard accuracy-based evaluation. We introduce ISAAC (Intervention-based Structural Auditing ApproaCh for Model Evaluation), a post-hoc framework that evaluates prior-relative structural sensitivity by probing frozen models through matched prior-aligned and prior-misaligned input-level interventions, independently of predictive accuracy. Applied to three sequence-based DTI architectures on the Davis kinase benchmark, ISAAC reveals statistically significant differences in prior-alignment ratios across models operating within a comparable predictive regime, differences that are not explained by predictive performance alone and that are robust to the choice of perturbation operator. These discrepancies are not reflected by the predictive ranking and motivate the use of post-hoc structural auditing as a complement to standard performance evaluation in DTI modeling and, more broadly, in scientific machine learning settings where structured priors are available.

Virtual screening (VS) aims to identify bioactive compounds from vast chemical libraries, but remains difficult in low-label regimes where only a few actives are known. Existing methods largely rely on general-purpose molecular fingerprints and overlook class-discriminative substructures critical to bioactivity. Moreover, they consider molecules independently, limiting effectiveness in low-label regimes. We introduce SubDyve, a network-based VS framework that constructs a subgraph-aware similarity network and propagates activity signals from a small known actives. When few active compounds are available, SubDyve performs iterative seed refinement, incrementally promoting new candidates based on local false discovery rate. This strategy expands the seed set with promising candidates while controlling false positives from topological bias and overexpansion. We evaluate SubDyve on ten DUD-E targets under zero-shot conditions and on the CDK7 target with a 10-million-compound ZINC dataset. SubDyve consistently outperforms existing fingerprint or embedding-based approaches, achieving margins of up to +34.0 on the BEDROC and +24.6 on the $EF_{1\\%}$ metric.

Abstract Introduction: Measuring perturbations of the transcriptome upon drug treatment can be highly informative, yet perturbed transcriptome profiles are often unavailable at both the cell-level and the patient-level. A key challenge, therefore, is how to infer such unobserved drug responses. Because drug treatment perturbs not only individual gene expression levels but also the underlying gene-gene interactions, the major challenge is how to model perturbations of these gene-gene interactions at the gene-level in LINCS so that unobserved drug responses of cancer cells and patients can be inferred. Methods: To predict unobserved perturbation of transcriptome after drug treatment, we have developed a deep learning model that captures perturbations of gene-gene interactions in LINCS, Condition-Specific Gene-Gene Attention (CSG2A, Bioinformatics/ISMB 2024). CSG2A models dosage- and time-dependent transcriptome perturbation and serves as as perturbation-based pretrained model. As change in transcriptome can be also seen as transcriptomic cellular state change, we used an existing pre-trained rank-based model (Geneformer, Nature 2024). Given these models of transcriptomic changes, we model the cancer patient in two different approaches in terms of transcriptome profiles. In the first approach, a deep learning model, called THERAPI, embeds a tumor by combining arbitrary sets of cancer cells in GDSC, after which perturbations of gene-gene interactions are predicted with the perturbation-based and rank-based models. In the second approach, we proposed PREDIKTOR, which uses a network-based representation of the patient transcriptome with DysRegNet (British Journal of Pharmacology 2024), and then predicts drug-induced perturbations of gene-gene interactions using the perturbation-based pre-trained model. Results: We developed two computational frameworks for predicting patient-level drug response from cell-line response, THERAPI and PREDIKTOR. Both THERAPI and PREDIKTOR outperformed existing deep learning models with significant margin up to 8.5% improvement in experiment with TCGA data and the I-SPY2 patient cohort. Conclusion: These two frameworks are being deployed on the Amazon cloud as “DrugVLABTM response” so that any researchers around the world can utilize for drug response prediction. Our systems can evolve with upcoming deep learning models, propriety or public, for more accurate drug response prediction. Citation Format: Sun Kim, Bonil Koo, Dongmin Bang, Inyoung Sung, Changyun Cho, Sangseon Lee, Kyoung Jae Won. Deep learning frameworks for translating cancer drug response from cell-level to patient-level by modeling transcriptome [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 978.

Prediction of patient-level drug response is critical for precision oncology but remains limited by the scarcity of clinical data. While machine learning models trained on cell lines offer a scalable alternative, biological differences introduce domain shifts that hinder direct translation to patient tumors. Here, we present THERAPI (Tumor Heterogeneity-aware Embedding for Response Adaptation and Patient Inference), a deep learning framework designed to bridge this gap. First, THERAPI aligns patient tumors to cell lines through attention-based aggregation guided by tissue context, modeling each tumor as a linear combination of cell lines. Second, THERAPI transfers gene- and cell-level knowledge from pre-trained perturbation and rank embeddings to train drug response predictors. THERAPI outperforms 11 baselines on TCGA dataset, generalizes to external breast and colorectal cancer cohorts, and supports interpretable gene/pathway-level analysis. These results highlight the value of integrating tumor-biology context and perturbation-aware modeling for generalizable and interpretable drug response prediction towards precision oncology.

Deep learning models for drug–target interaction (DTI) prediction often achieve strong benchmark performance without necessarily relying on mechanistically meaningful molecular features, a limitation that standard accuracy-based evaluation cannot detect. We introduce ISAAC (Intervention-based Structural Auditing Approach for Causal Reasoning), a post-hoc framework that evaluates prior-relative structural sensitivity by probing frozen models through matched mechanistic and spurious input-level interventions, independently of predictive accuracy. Applied to three sequence-based DTI architectures on the Davis benchmark, ISAAC reveals approximately 25% relative differences in reasoning scores across models with comparable AUROC (within around 3%), stable across training and intervention seeds and two distinct perturbation operators. These discrepancies, undetectable under conventional accuracy metrics, motivate the use of post-hoc structural auditing as a complement to standard performance evaluation in scientific machine learning for molecular modeling.

Biomedical knowledge graphs that include drugs, genes, and diseases support drug repurposing by connecting drugs to diseases through gene-mediated multi-hop paths, thereby enabling mechanism-of-action reasoning. However, deeper traversal does not necessarily improve mechanistic reasoning: long paths grow combinatorially and frequently pass through hub genes, producing irrelevant gene regulatory signals, whereas overly constrained or sparse paths may miss broader biological context. We propose Context-Aware REasoning Path (CAREPath), a knowledge graph (KG)-large language model framework inspired by depth-search and breadth-search reasoning to balance mechanistic specificity, scalability, and context recovery. The depth-search strategy constrains traversal to short disease-gene-drug paths, converts each path into a structured prompt, and encodes it with a biomedical language model to generate semantic path embeddings. Complementarily, the breadth-search strategy constructs entity-level mechanism-context embeddings from one-hop gene neighborhoods and enriches them through similarity-guided augmentation using pharmacologically related drugs and gene-signature-similar diseases. Across five biomedical KGs, CAREPath achieves the best area under the precision-recall curve (AUPRC) in the disease cold-start setting among 18 baselines, improving performance by up to 3.6%. Additional analyses show that semantic short-path encoding contributes most to performance, while mechanism-context augmentation improves robustness under sparse path signals and strengthens gene ontology functional agreement. Case studies and recently U.S. Food and Drug Administration (FDA)-approved indications further demonstrate its practical relevance, positioning CAREPath as a framework that supplies interpretable mechanistic rationales where constrained path is available, while remaining robust when it is not. Source code is available at https://github.com/hamppy-song/CAREPath.

Molecular language modeling tasks such as molecule captioning have been recognized for their potential to further understand molecular properties that can aid drug discovery or material synthesis based on chemical reactions. Unlike the common use of molecule graphs in predicting molecular properties, most methods in molecular language modeling rely heavily on SMILES sequences. This preference is because the task involves generating a sequence of multiple tokens using transformer-based models. Therefore, a main challenge is determining how to integrate graph data, which contains structural and spatial information about molecules, with text data. In addition, simply using both 1D SMILES text and 2D graph as inputs without addressing how they align and represent the molecule structure in different modalities makes it challenging to fully utilize structural knowledge about molecules. To this end, we propose GraphT5, a multi-modal framework that integrates 1D SMILES text and 2D graph representations of molecules for molecular language modeling. Specifically, we introduce a novel cross-token attention module in GraphT5 to bridge the gap arising from the fundamental differences between the two modalities of molecule representations. Cross-token attention exploits implicit information between SMILES and graphs of molecules, resulting from their interactions at a fine-grained token level that benefits molecular language modeling. Extensive experiments including molecule captioning, IUPAC name prediction tasks, and case studies show that our GraphT5 outperforms the latest baseline approaches, which validates the effectiveness of our GraphT5 in sufficiently utilizing 1D SMILES text and 2D graph representations.

Accurately predicting protein-ligand binding affinity is a critical challenge in drug discovery, crucial for understanding drug efficacy. While existing models typically rely on atom-level interactions, they often fail to capture the complex, higher-order interactions, resulting in noise and computational inefficiency. Transitioning to modeling these interactions at the cluster level is challenging because it is difficult to determine which atoms form meaningful clusters that drive the protein-ligand interactions. To address this, we propose CheapNet, a novel interaction-based model that integrates atom-level representations with hierarchical cluster-level interactions through a cross-attention mechanism. By employing differentiable pooling of atom-level embeddings, CheapNet efficiently captures essential higher-order molecular representations crucial for accurate binding predictions. Extensive evaluations demonstrate that CheapNet not only achieves state-of-the-art performance across multiple binding affinity prediction tasks but also maintains prediction accuracy with reasonable computational efficiency. The code of CheapNet is available at https://github.com/hyukjunlim/CheapNet.

IntroductionDrug-induced liver injury (DILI) has been investigated at the patient level. Analysis of gene perturbation at the cellular level can help better characterize biological mechanisms of hepatotoxicity. Despite accumulating drug-induced transcriptome data such as LINCS, analyzing such transcriptome data upon drug treatment is a challenging task because the perturbation of expression is dose and time dependent. In addition, the mechanisms of drug toxicity are known only as literature information, not in a computable form.MethodsTo address these challenges, we propose a Multi-Dimensional Transcriptomic Ruler (MDTR) that quantifies the degree of DILI at the transcriptome level. To translate transcriptome data to toxicity-related mechanisms, MDTR incorporates KEGG pathways as representatives of mechanisms, mapping transcriptome data to biological pathways and subsequently aggregating them for each of the five hepatotoxicity mechanisms. Given that a single mechanism involves multiple pathways, MDTR measures pathway-level perturbation by constructing a radial basis kernel-based toxicity space and measuring the Mahalanobis distance in the transcriptomic kernel space. Representing each mechanism as a dimension, MDTR is visualized in a radar chart, enabling an effective visual presentation of hepatotoxicity at transcriptomic level.Results and DiscussionIn experiments with the LINCS dataset, we show that MDTR outperforms existing methods for measuring the distance of transcriptome data when describing for dose-dependent drug perturbations. In addition, MDTR shows interpretability at the level of DILI mechanisms in terms of the distance, i.e., in a metric space. Furthermore, we provided a user-friendly and freely accessible website (http://biohealth.snu.ac.kr/software/MDTR), enabling users to easily measure DILI in drug-induced transcriptome data.

Accurate survival prediction is essential in healthcare as it guides treatment strategies and improves patient outcomes. While clinical features provide valuable prognostic information, they often fail to represent the molecular complexity of diseases. Transcriptomic data, which reflects gene expression patterns of tumors, present a complementary perspective to address this limitation. We introduce Transcriptome Transformer (TxT), a multitask learning framework that uses a transcriptome-centric approach to improve patient survival prediction. TxT employs a Transformer-based architecture with multihead attention mechanisms to effectively capture complex dependencies among genes, enabling dynamic modeling of gene-gene interactions while using shared information across multiple clinical prediction tasks. By jointly analyzing transcriptomic data and incorporating clinical features, TxT offers a more complete representation of patient biology. In experiments across both single-task and multitask datasets, TxT outperformed existing methods in survival prediction and related clinical tasks. Additionally, TxT offers biological insights through attention-derived gene interaction networks, identifying immune-related pathways in longer-surviving Luminal A patients and coagulation and epithelial-mesenchymal transition pathways in shorter-surviving counterparts. Differential attention analysis further revealed that integrating clinical features enhances the model's ability to prioritize genes involved in biologically meaningful pathways that are known to influence tumor progression and distant recurrence. The source code of TxT is available at https://github.com/BonilKoo/TxT.

Knowledge of unintended effects of drugs is critical in assessing the risk of treatment and in drug repurposing. Although numerous existing studies predict drug-side effect presence, only four of them predict the frequency of the side effects. Unfortunately, current prediction methods 1) do not utilize drug targets, 2) do not predict well for unseen drugs, and 3) do not use multiple heterogeneous drug features. We propose a novel deep learning-based drug-side effect frequency prediction model. Our model utilized heterogeneous features such as target protein information as well as molecular graph, fingerprints, and chemical similarity to create drug embeddings simultaneously. Furthermore, the model represents drugs and side effects into a common vector space, learning the dual representation vectors of drugs and side effects, respectively. We also extended the predictive power of our model to compensate for the drugs without clear target proteins using the Adaboost method. We achieved state-of-the-art performance over the existing methods in predicting side effect frequencies, especially for unseen drugs. Ablation studies show that our model effectively combines and utilizes heterogeneous features of drugs. Moreover, we observed that, when the target information given, drugs with explicit targets resulted in better prediction than the drugs without explicit targets.

BACKGROUND:The immune response to SARS-CoV-2 varies greatly among individuals yielding highly varying severity levels among the patients. While there are various methods to spot severity associated biomarkers in COVID-19 patients, we investigated highly mutated regions, or mutation hotspots, within the SARS-CoV-2 genome that correlate with patient severity levels. SARS-CoV-2 mutation hotspots were searched in the GISAID database using a density based clustering algorithm, Mutclust, that searches for loci with high mutation density and diversity. RESULTS:Using Mutclust, 477 mutation hotspots were searched in the SARS-CoV-2 genome, of which 28 showed significant association with severity levels in a multi-omics COVID-19 cohort comprised of 387 infected patients. The patients were further stratified into moderate and severe patient groups based on the 28 severity related mutation hotspots that showed distinctive cytokine and gene expression levels in both cytokine profile and single-cell RNA-seq samples. The effect of the SARS-CoV-2 mutation hotspots on human genes was further investigated by network propagation analysis, where two mutation hotspots specific to the severe group showed association with NK cell activity. One of them showed to decrease the affinity between the viral epitope of the hotspot region and its binding HLA when compared to the non-mutated epitope. CONCLUSION:Genes related to the immunological function of NK cells, especially the NK cell receptor and co-activating receptor genes, were significantly dysregulated in the severe patient group in both cytokine and single-cell levels. Collectively, mutation hotspots associated with severity and their related NK cell associated gene expression regulation were identified.

Identification of virulence factors (VFs) is critical for expanding our knowledge on bacterial pathogenesis and also for developing targeted strategies for the prevention and treatment of related infectious diseases. Understanding virulence factors requires to consider coevolutionary information, as it reveals the evolutionary interdependencies between amino acid residues, which can provide some biological insights into their functional and structural roles in bacterial pathogenicity. Previous studies have conducted VF predictions without considering coevolutionary information of proteins. In this paper, we introduce MSA-VF Predictor (MVP), a novel deep learning-based method that effectively captures coevolutionary features inherent in protein sequences for VF prediction. The first step of our method is to generate multiple sequence alignment (MSA) that can represent evolutionary information of VF related protein sequences. Then, we utilize the MSA Transformer to extract features from the MSA data that capture coevolutionary information and homologous protein information. Using these coevolutionary features along with the residue level information, we propose MSA-composition, which consists of latent vectors for amino acids in matrix form. Our approach achieved a prediction accuracy of 0.869, outperforming existing state-of-the-arts (SOTA) models. We conducted experiments to interpret the relationship between MVP’s performance and coevolutionary information, and presented the interpretation results. To further investigate the MSA transformer model, we performed experiments of pruning attention blocks, which shows attention blocks that play a crucial role in VF prediction are also significant to VF proteins with high coevolutionary information. In summary, MVP ( http://bhi4.snu.ac.kr:7978 ) successfully incorporates coevolutionary information for predicting VF proteins using MSA transformer.

Toxicity is a critical hurdle in drug development, often causing the late-stage failure of promising compounds. Existing computational prediction models often focus on single-organ toxicity. However, avoiding toxicity of an organ, such as reducing gastrointestinal side effects, may inadvertently lead to toxicity in another organ, as seen in the real case of rofecoxib, which was withdrawn due to increased cardiovascular risks. Thus, simultaneous prediction of multi-organ toxicity is a desirable but challenging task. The main challenges are (1) the variability of substructures that contribute to toxicity of different organs, (2) insufficient power of molecular representations in diverse perspectives, and (3) explainability of prediction results especially in terms of substructures or potential toxicophores. To address these challenges with multiple strategies, we developed FATE-Tox, a novel multi-view deep learning framework for multi-organ toxicity prediction. For variability of substructures, we used three fragmentation methods such as BRICS, Bemis-Murcko scaffolds, and RDKit Functional Groups to formulate fragment-level graphs so that diverse substructures can be used to identify toxicity for different organs. For insufficient power of molecular representations, we used molecular representations in both 2D and 3D perspectives. For explainability, our fragment attention transformer identifies potential 3D toxicophores using attention coefficients. Scientific contribution: Our framework achieved significant improvements in prediction performance, with up to 3.01% gains over prior baseline methods on toxicity benchmark datasets from MoleculeNet (BBBP, SIDER, ClinTox) and TDC (DILI, Skin Reaction, Carcinogens, and hERG), while the multi-task learning approach further enhanced performance by up to 1.44% compared to the single-task learning framework that had already surpassed these baselines. Additionally, attention visualization aligning with literature contributes to greater transparency in predictive modeling. Our approach has the potential to provide scientists and clinicians with a more interpretable and clinically meaningful tool to assess systemic toxicity, ultimately supporting safer and more informed drug development processes.

Phenotype-driven drug discovery leverages cellular responses to guide the design of therapeutic molecules. Recent advancements in transcriptomics have provided extensive datasets describing how gene expression changes in response to various chemical stimuli, presenting an opportunity to directly link molecular generation to specific cellular phenotypes. However, effectively linking transcriptomic perturbations to chemical structure generation remains challenging due to the complexity of gene interactions and chemical feasibility constraints. We developed GGIFragGPT, a novel generative model that integrates transcriptomic perturbation profiles with biologically informed gene-gene interaction embeddings to guide fragment-based molecular generation. The model employs an autoregressive transformer architecture to sequentially assemble chemically valid fragments, with cross-attention mechanisms highlighting biologically relevant genes guiding the molecular generation process. Comparative analysis confirmed that the proposed approach yields chemically feasible, novel, and diverse molecules. By leveraging transcriptomic profiles, GGIFragGPT successfully generated compounds aligned with the biological context suggested by transcriptomic data, validated through gene-level interpretability analysis that identified key target genes. Case studies demonstrated the model’s capability to produce structurally plausible inhibitors, exemplified by targeted molecule generation against CDK7. This work demonstrates the potential of integrating biological insights into chemical generation processes, offering a promising approach for phenotype-driven therapeutic discovery.