npj Precision Oncology

Progress in precision oncology, including biomarker discovery and individualized treatment selection, is limited by the complexity of clinico-genomic data and the scarcity of large multimodal patient cohorts. Here, we introduce PanoraOnc, a pan-cancer artificial intelligence (AI) model pretrained on real-world clinical, genomic, and imaging data from 84,131 patients spanning 66 cancer types. PanoraOnc enables transferable treatment outcome prediction through pan-cancer pretraining and generalizes to unseen cohorts across cancer types, institutions, and therapeutic settings. Evaluation and fine-tuning were performed on cohorts comprising diverse modalities, including clinical features, targeted gene panels, immunofluorescence imaging, whole-exome sequencing, and transcriptomic profiles. Across these settings, PanoraOnc consistently outperforms statistical, machine-learning, survival, and AI baselines, with the largest improvements observed in zero- and few-shot scenarios, demonstrating that large-scale clinico-genomic pretraining enables robust and generalizable outcome predictions across previously unseen conditions. In addition, PanoraOnc supports biomarker discovery through explainable AI, revealing both established and underappreciated features, including tumor-infiltrating clonal hematopoiesis, oncogenic signaling pathways, and DNA damage response mechanisms in immunotherapy-treated melanoma and non-small cell lung cancer. Furthermore, PanoraOnc enables the identification of patient subgroups potentially benefitting from alternative treatments by estimating personalized treatment outcomes across therapeutic scenarios. These findings establish pan-cancer multimodal pretraining as a scalable paradigm for AI-assisted discovery in precision oncology.

1Accurate subtype diagnosis is essential for guiding therapy and predicting patient outcome in malignant mesothelioma. Most computational pathology models are trained on large tissue images from resection specimens, which maximises information for training but limits model relevance in real-world diagnostic settings where small biopsies are the most usual tissue source. In this work, we assembled a large multicentre cohort of HES- and HPS-stained mesothelioma biopsy slides. We used a self-supervised learning model to evaluate the associations of biopsy-driven morphology patterns with histological subtype, molecular markers, and survival. The discovered histomorphology patterns captured a continuum of tissue phenotypes spanning epithelioid, sarcomatoid, and non-tumour morphologies. Also, patient-level HPC representations achieved excellent performance for distinguishing epithelioid from non-epithelioid mesothelioma (AUC = 0.94) and demonstrated predictive value for immunohistochemistry (IHC) markers. Additionally, HPC-derived features alone achieved performance comparable to established clinical and molecular variables (C-index = 0.65), while integration of HPCs with clinical and marker information improved performance to a C-index of 0.69. Several HPCs were significantly associated with favourable or adverse prognosis and reflected known subtype-specific biological patterns. In conclusion, self-supervised learning can discover interpretable histomorphological phenotypes directly from routine mesothelioma biopsies without further training. These AI-derived phenotypes capture clinically and biologically relevant information, linking tissue architecture to molecular characteristics, histological subtypes, and patient outcomes. The proposed framework provides a thorough evaluation of real-world biopsy data using a pre-trained model, without the need for computationally intensive retraining, and addresses the question of whether SSL-based AI can be deployed out of the box in clinical settings.

Next-generation sequencing (NGS) for the detection of somatic variants has become the method of choice in a variety of molecular oncology fields and in the clinic. Its use ranges from sequencing entire tumor genomes and transcriptomes to targeted clinical diagnostic gene panels. The NYU Langone Genome PACT (Profiling of Actionable Cancer Targets, LG-PACT) assay is a qualitative in vitro diagnostic test that uses targeted next generation sequencing (NGS) of formalin-fixed paraffin-embedded (FFPE) tumor tissue matched with normal specimens from patients to detect gene alterations in a targeted panel covering 606 genes and the TERT promoter. Indications for testing are cancer (solid tumors and hematological malignancies) where a mutational profile from multiple genes would be informative for disease stratification, prognosis, or treatment options including targeted therapies and eligibility for clinical trials. The test is intended to provide information on somatic mutations including point mutations, small insertions/deletions (indels), and copy number aberrations for diagnostic and treatment decisions. LG-PACT is a United States Food and Drug Administration (FDA) cleared diagnostic test (510K: K202304). The clinical interpretation of sequencing data of molecular tumor markers from NGS encompasses automated variant calling tools with human interpretation. This final mostly manual review of data step is intensive, involving highly trained scientists, encompassing literature review, interpretation and clinical tier classification by pathologists, who then provide a complete molecular diagnostic report to the treating oncologists. We provide analysis of 1339 clinical genomic profiles from 31 different cancers and their subtypes, comprising of central nervous system (CNS) 792 (59%) cases (incl. meningioma, glioma and glioblastoma), with 267 (20%) cases predominantly of lung, pancreatic and colorectal and 280 of others (21%). Here, we present the technical challenges of validating an NGS oncological diagnostic targeted assay for clinical grade accuracy and sensitivity for patient care. We show how copy number alterations provide a more comprehensive description of the tumors genomic profile. We then outline the utility of targeted panel sequencing based on certified pathologist selection of reportable variants for our current patient cohort. Where analysis of variant detection has led to 49.4% (661/1339) of our clinical tumor samples containing mutations in known therapy targeted genes, 35.6% (477/1339) with mutation detected in other genes, and 15% (201/1339) cases being negative.

Tumor cellular composition--including malignant cell states, immune populations, and stromal populations--is increasingly recognized as a determinant of therapeutic response and resistance to anti-cancer agents, yet comprehensive cellular profiling remains largely confined to research settings. Here, we present a clinically compatible sample-to-report workflow for tumor composition profiling from routine formalin-fixed paraffin-embedded (FFPE) clinical specimens. By combining low-input single-nucleus RNA sequencing with foundation model- based automated cell annotation, this workflow enables prospective sample-by-sample analysis without dedicated research material or cohort-based processing. Across 116 clinical specimens representing six cancer types, we generated reproducible measurements of cellular composition and cell-type-specific gene expression, demonstrated high technical reproducibility, and showed concordance with pathological assessment of immune infiltration. The workflow was similarly applicable to archival FFPE material and ultra-low-input biopsy specimens. Together, these findings establish a practical framework for routine single-cell profiling from standard pathology specimens and open the perspective of prospective evaluation of cellular composition as a clinical biomarker in precision oncology.

Patient-derived tumor organoids provide a physiologically relevant 3D disease model for preclinical drug discovery, surpassing the limitations of conventional 2D cell lines. To better capture the dynamic nature of organoid drug responses, we developed a new systematic evaluation method called SCOPE (Systematic Classification of Organoids for Phenotypic Evaluation), harnessing phenotypic assessments from multi-timepoint 3D imaging data. By integrating artificial intelligence (AI)-based image analysis of organoid viability with tracking and mathematical modeling of organoid growth over time, we captured temporal-and dose-dependent dynamics of phenotypic changes, culminating in two novel metrics: a combined growth and viability (GV) score as well as a cytostatic-cytotoxic transition range (CCTR) that separates drug effects on organoid growth and viability. Our approach supports classification of specific drug responses into four distinct phenotypic groups: (1) cytotoxic, (2) cytostatic plus cytotoxic, (3) late cytotoxic, and (4) cytostatic. This novel drug evaluation system can identify previously unknown drug effects or new therapeutic use cases for existing drugs, facilitating the design of alternative therapeutic options to overcome efficacy or drug resistance challenges and improving the clinical applicability of organoid-based drug discovery results.

Purpose: To compare the cross-site generalization of radiomic features and deep learning embeddings for MRI prediction of substantial lymphovascular space invasion (LVSI) in endometrial cancer. Materials and Methods: This retrospective two-center study included 206 women (mean age, 59.8 years) with endometrial cancer who underwent preoperative 3-T MRI from March 2016 to March 2023. Hospital A (n = 130) was used for development and Hospital B (n = 76) for strict external testing. T2-weighted, reduced field-of-view diffusion-weighted, and apparent diffusion coefficient images were manually segmented. Radiomic features and seed-pooled embeddings from 3D ResNet18, DenseNet121, and U-NEXtractor were modeled with elastic-net logistic regression or XGBoost. Out-of-fold Platt calibration and sensitivity-targeted thresholds were estimated using development data only. AUCs were summarized with 95% bootstrap confidence intervals. Results: External radiomics with elastic-net achieved an AUC of 0.609 (95% CI: 0.464, 0.740) and sensitivity of 0 of 12 (0%). DenseNet121 with elastic-net had the highest external AUC (0.685; 95% CI: 0.538, 0.822) but sensitivity of 3 of 12 (25%). U-NEXtractor with elastic-net detected 10 of 12 positive cases (83.3%) with specificity of 32 of 64 (50.0%) and balanced accuracy of 0.667. XGBoost showed higher apparent development performance but weaker external operating behavior. Conclusion: Under real-world cross-site MRI acquisition shift, DenseNet121 and U-NEXtractor embeddings showed better external generalization than handcrafted radiomic features for substantial LVSI prediction.

BackgroundOvarian cancer (OvCa) ranks as the most lethal gynecological malignancy in women worldwide. This complex disease, which can develop independently of a womans age, is characterized by late diagnosis, pronounced tumor heterogeneity, and an immunosuppressive tumor microenvironment (TME). Incremental diagnostic tools that could better inform clinicians on potential therapy resistance or subsets of patients that could benefit from new drug modalities represent a critical unmet need to improve patient care and potentially the identification of new biomarkers. ObjectiveThis study aimed to establish a reconfigurable patient-derived OvCa-on-chip platform for longitudinal functional profiling of tumor cell death, immune activation, and patient-specific responses to TIL-mediated killing, PD-L1 blockade, and sequential chemo-immunotherapy. MethodsPatient-derived OvCa microtumors (PDM) were integrated with sequential integration of autologous tumor-infiltrating lymphocytes (TILs) into a perfusable microfluidic chip in the presence of different single and combination treatment regimens of chemotherapy and immune checkpoint inhibitors (ICIs). Treatment responses were assessed by longitudinal quantification of caspase-cleaved cytokeratin 18 (ccCK18) as marker of apoptotic epithelial tumor cell death, as well as cytokine/chemokine release in chip effluents, and multiplex flow cytometry-based characterization of autologous TIL subsets. ResultsThe perfusable OvCa-on-chip platform supported long-term culture of PDM while maintaining key structural and microenvironmental features of the primary tumor. Multidimensional analyses, including time-resolved assessment of tumor cell death, secretome profiling and correlative analysis of autologous TIL subsets revealed patient-specific tumor-immune response patterns and heterogenous sensitivity to TIL-mediated killing, PD-L1 blockade, and sequential chemo-immunotherapy. Correlation analyses identified treatment-dependent associations between specific TIL phenotypes and functional tumor cell killing. PD-1-expressing CD4 TIL subsets correlated with enhanced tumor cell killing, whereas terminally exhausted CD8PD-1Tcf1- TILs negatively correlated with durvalumab responses. In contrast, tumor-reactive CD8CD39 TILs were associated with improved responses under sequential chemo-immunotherapy conditions. ConclusionCollectively, this OvCa-on-chip system represents a complex in vitro model (CIVM) that combines 3D tumor tissue with autologous immune cells in a microfluidic platform. Resembling a physiologically relevant human preclinical platform, it allows for the time-resolved functional assessment of patient-specific responsiveness to OvCa therapies, with direct implications for personalized treatment stratification.

Triple-Negative Breast Cancer (TNBC) is characterized by high heterogeneity, poor prognosis, and limited targeted treatment options. Bridging the gap between molecular alterations and histopathological morphology remains a major challenge in precision oncology. We propose an interpretable, multi-modal framework that integrates histopathological image analysis with multi-omics profiling (somatic mutations, DNA methylation, copy number alterations), leveraging U-Net-based nuclei segmentation, vision-language models (BLIP), biomedical language models (BioGPT), and explainable AI (SHAP, LIME). Our framework achieves strong predictive performance (AUC = 0.989) and provides transparent, biologically grounded interpretations by integrating morphological features with genomically prioritized biomarkers. Cross-modal analysis confirms established TNBC drivers and generates novel, testable hypotheses associating specific epigenetic alterations with distinct morphological phenotypes. While causal validation requires future wet-lab experiments, our framework accelerates hypothesis-driven biomarker discovery by integrating complementary data modalities with language-based reasoning, providing a transparent foundation for hypothesis generation and clinical translation.

CD19-targeted chimeric antigen receptor (CAR) T cell therapy achieves high initial response rates in B-cell acute lymphoblastic leukemia (B-ALL), yet half of patients relapse within one year. Pre-infusion product composition decoded by single-cell RNA sequencing (scRNA-seq) carries information predictive of long-term CAR T persistence, but extracting this information from individual patients typically requires highly sophisticated bioinformatics expert annotation, limiting clinical translation. Here, we evaluate whether single-cell foundation models (scFMs) can extract clinically actionable information from engineered CAR T products. We applied four scFMs (scGPT, scFoundation, CellPLM and UCE), including fine-tuned versions of scGPT and scFoundation, to paired basal and CD19-stimulated pre-infusion CAR T products from 33 pediatric patients with B-ALL. Although annotation accuracy declined relative to healthy peripheral blood references, scFM-derived cell composition stratified patients with long-duration B-cell aplasia with a leave-one-out cross-validated area under the receiver operating characteristic curve of 0.879 (95% confidence interval, 0.742-0.986). Notably, foundation-model-identified cell proportion analysis matched or exceeded expert annotations for several predictive features, demonstrating that accurate clinical prediction may not require perfect per-cell annotation to begin with. CD8+XCL1/2+ cells were further identified as the biomarker consistently associated with durable CAR T persistence across models under CD19 stimulation, whereas other candidate populations showed limited reproducibility. Finally, we translate these findings into a locally deployable decision-support AI agent that predicts the probability of sustained CAR T persistence from pre-infusion CAR T scRNA-seq data.

Longitudinal circulating tumor DNA (ctDNA) measurements offer a noninvasive means to monitor treatment response, but clinical trial data present substantial methodological challenges due to high-dimensional short longitudinal ctDNA sequences and limited sample sizes. We introduce SHERLOC, a deep learning framework specifically designed for survival analysis using longitudinal on-treatment ctDNA data, which integrates shared temporal representations of gene-level variant allele frequencies, feature-specific temporal trajectories of panel-level ctDNA biomarkers, and survival-aware genomic representations pre-trained on a large pan-cancer tissue-biopsy dataset (MSK-CHORD), within an interpretable Cox proportional hazards framework. Benchmarked against diverse statistical, ensemble, and deep learning approaches in a non-small-cell lung cancer cohort from the phase III IMpower150 trial, SHERLOC consistently achieved superior survival discrimination and calibration, while remaining interpretable and robust to reductions in the number of available longitudinal liquid biopsy time points per patient. The resulting ctDNA-based risk score provided prognostic information both independent of and complementary to standard radiographic response assessments, and enabled patient stratification within homogeneous RECIST response groups--highlighting its potential as an early, non-invasive decision-support tool to guide treatment adaptation and patient management.

Patient-derived organoids (PDOs) have transformed translational cancer research, allowing tractable models that better represent clinical features than traditional immortalized cell lines. Here we describe two PDOs with differential responses to carboplatin derived from sequential ascites fluid collections from a patient with high-grade mullerian carcinoma, that could not be further subclassified on the omental biopsy. Uterine origin was clinically excluded by pelvic imaging/CT scan of the uterus and absence of vaginal bleeding. Successful derivation from independent collections enabled comparison of intra-patient heterogeneity across sequential ascites samples and demonstrates that PDO efficiency rate is at least partly patient-specific or tumor-dependent. We performed long-read whole genome sequencing on the two PDOs, OC104 and OC109, to better characterize the structural variant landscape while also obtaining information on single nucleotide variants and DNA methylation. In addition to confirming single nucleotide variants noted in clinical sequencing (TP53, KRAS, SPOP, PPP2R1A, KMT2D), we identified additional variants in TSC2, NCOR2, and CTNNA2 that are predicted to be likely pathogenic. The spectrum of mutations, particularly the coincident KRAS and TP53, highlighted unexpected overlap with ovarian mucinous carcinoma. We also identified larger insertions and deletions that result in non-synonymous variants in MUC5AC, TPRX1, and BMX, as well as four translocation events, including two that could not have been resolved with short-read sequencing. Differentially methylated promoters between the two PDOs include 201 oncogenes and tumor suppressor genes, with HNF1A, MSI2, and SETBP1 having methylation directions consistent with these genes' roles in platinum response differences observed between the PDOs. Notably, the clonal nature of PDOs produced from two samples taken one week apart is important for the field to appreciate, particularly since they have clonal differences in platinum response. The temporal differences in clonality may indicate a limitation of low volume sampling, however may provide opportunity to longitudinally predict clinical outcomes. We also demonstrate the ability of long-read sequencing to add detail into the genomics and epigenetics of ovarian cancer.

Most patients with cancer of unknown primary (CUP) still receive platinum-based chemotherapy and have a poor prognosis, with overall survival of less than one year. Recent studies suggest improved outcomes with molecularly guided or site-specific therapies informed by molecular tissue profiling. Here, we analyzed ctDNA from 190 CUP patients using an integrated genomic and epigenomic assay to identify actionable alterations and predict tissue-of-origin (ToO). Integration of actionable biomarkers, ToO prediction and clinical data yielded diagnostic, prognostic or therapeutic information in 90% of unfavorable CUP cases and 88% of patients analyzed at first diagnosis. High ctDNA tumor fraction was associated with poorer prognosis in both favorable and unfavorable CUP. These findings highlight the clinical utility of ctDNA analysis for therapeutic decision-making in CUP and support its incorporation into the diagnostic work-up, particularly when tissue samples are unavailable or insufficient for molecular testing.

Background: Liquid biopsy-based monitoring of circulating tumor DNA (ctDNA) holds promises for real-time assessment of tumor burden and treatment response in precision oncology. However, mutation-based approaches alone show limited sensitivity, particularly in low-shedding tumors. We evaluated whether integrating epigenomic biomarkers (specifically LINE-1 retrotransposon (L1PA) hypomethylation and copy number variation (CNV)) with standard mutation-based ctDNA analysis could improve cancer detection and longitudinal monitoring in patients enrolled in the SHIVA02 precision oncology trial. Methods: We performed a retrospective analysis of 32 patients with advanced or metastatic solid tumors who received molecularly matched targeted therapies within the SHIVA02 trial (NCT03084757). Plasma samples were collected at baseline and longitudinally every two months and at progression. Multimodal ctDNA profiling was performed using the DRAGON targeted NGS panel covering SNVs, indels and focal CNVs in 571 genes and the DIAMOND assay profiling L1PA methylation and genome-wide CNV. A three-step classification algorithm integrating maximum variant allele frequency (MaxVAF), L1PA methylation-based cancer probability (MethPCancer), and genome-wide CNV scores was developed to maximize ctDNA detectability. Results: At baseline, mutation-based profiling detected ctDNA in 62.5% of patients. L1PA hypomethylation alone identified ctDNA in 78.1% of patients, including cases with undetectable mutations. The three-step integrative model increased overall detectability to 93.8%. Concordance analyses between tumor tissue and plasma revealed that 80.6% of mutations and 60% of CNVs identified in tumor biopsies were detectable in matched ctDNA. The three modalities showed limited pairwise correlation at baseline, supporting their complementarity. Longitudinal analysis demonstrated that changes in MaxVAF, MethPCancer, and L1PA CNV scores over time were informative on treatment response across tumor types, with ctDNA detected in 94.5% of samples collected at disease progression. In selected patients, ctDNA alterations preceded radiological progression by four months. Conclusions: Multimodal ctDNA profiling integrating mutation analysis, CNV profiling, and LINE-1 hypomethylation substantially improves ctDNA detection at baseline and during treatment in a pan-cancer precision oncology setting. These complementary genomic and epigenomic biomarkers provide a more comprehensive and dynamic assessment of tumor burden than single-modality approaches, supporting prospective validation in larger cohorts for integration into precision oncology workflows.

Diagnostic assessment of breast cancer biopsies remains reliant on resource-intensive histopathology and molecular profiling, which often lack real-time physiological readouts. Magnetic levitation (MagLev) enables label-free density profiling of single cells, yet its application to intact tissue biopsies has been precluded by size-dependent geometric artifacts and the absence of analytical frameworks for biopsy-scale samples. Here, we report the first application of MagLev to intact invasive breast carcinoma biopsies (200-600 m) for biophysical profiling, generating multivariate biophysical signatures from 203 samples across 17 patients. We developed a physics-based size-correction algorithm (xmc) that isolates biological density from geometric artifact, and demonstrate that tissue viability is predicted not by average levitation height, but by spatial heterogeneity across replicate samples, reflecting the microenvironmental complexity of metabolically active tumors. Multivariate integration using Partial Least Squares (PLS) regression and Factor Analysis of Mixed Data (FAMD) identified nodal status (N) as the strongest biophysical predictor, suggesting that lymphatic dissemination capacity leaves a measurable signature in the primary tumor density profile. Unsupervised patient clustering in PLS-derived latent space recovered three clinically coherent subgroups aligned with molecular subtypes. This 30-minute, low-cost assay provides exploratory biophysical stratification complementary to existing diagnostics, particularly in resource-limited settings.

Copy number alterations (CNAs) are major contributors to genomic instability in cancer, and copy number signatures (CNS) provide a compact representation of the processes shaping CNA landscapes. However, the relationships among existing CNS frameworks and their predictability from molecular data other than whole-genome sequencing remain unclear. Here, we compare three major CNS compendia across more than 5,800 TCGA cancer samples, evaluating their overlap, complementarity, biological relevance, and prognostic associations. Individual signatures showed limited cross-study concordance, whereas signature-derived clusters identified biologically distinct patient groups, including favorable-outcome clusters observed across all frameworks. Using gene expression, DNA methylation, somatic mutation features, age, and tumor purity, XGBoost models predicted cluster membership with framework-dependent performance, achieving high F1-scores for the Drews and Steele compendia but limited performance for Tao. Feature importance analysis highlighted expression-driven predictors and pathways linked to genomic instability. These findings show that current CNS frameworks capture complementary rather than interchangeable dimensions of tumor genome instability and suggest that multi-omic profiles can extend signature-based stratification to cohorts without whole-genome sequencing.

Prognostic models in oncology are developed one cancer at a time, from that cancer's own labelled outcomes, and fail where prognostic information is scarcest. Rare cancers account for roughly a fifth of diagnoses and most paediatric malignancies, yet seldom supply enough events for a reliable time-to-event model. We therefore asked whether a representation learned without outcome labels can supply what those cohorts cannot. A Transformer encoder was pretrained by masked field-value modelling on 9425135 tumour records from the SEER 17 registries, diagnosed in 2000 to 2023. Only diagnosis-time fields passing a fail-closed coding-verification gate were admitted, and each record was emitted as an era-specific and a harmonised view, keeping two decades of recoding auditable. The encoder was then frozen and read by a linear Cox head for overall survival. Nine rare cancers were removed from the pretraining corpus entirely, each requiring an independent pretraining run. On a sealed test partition, all nine exceeded an architecture-identical random frozen encoder in Harrell concordance by +0.0034 to +0.0368, every lower confidence limit above zero. At 256 labelled patients, all 67 cancers favoured the pretrained representation over budget-matched Cox regression, median difference +0.0283. The advantage was bounded: given the entire training set, Cox regression was favoured in seven of nine rare cancers. The encoder did not outperform a field-frequency baseline on its own objective, so upstream reconstruction did not predict downstream transfer. Outcome-agnostic registry pretraining carries prognostic signal into cancers it has never seen, and is most useful where labels are fewest, without establishing clinical utility.

Hematoxylin and eosin (H&E) slides are routinely available in cancer care, but molecular profiling often requires additional tissue processing and turnaround time. We introduce HistoMethyl, a DNA methylation-aware pathology foundation model that aligns whole-slide histopathology with matched genome-scale methylation beta-value profiles during pre-training while requiring only an H&E slide at inference. We evaluated HistoMethyl across four task groups: gene mutation prediction, morphology-associated classification, overall survival prediction, and direct DNA methylation beta-value recovery. Evaluation spanned TCGA cross-validation, a disease-held-out lower-grade glioma cohort, and external cohort validation in CPTAC glioblastoma and SurGen rectal adenocarcinoma, totaling 14 cancer cohorts and 81 gene mutation tasks. The best-performing configuration improved mean mutation AUROC by 5.03 percentage points. By converting DNA methylation supervision into an H&E-only representation, His-toMethyl could support an early molecular triage layer that helps prioritize cases for confirmatory sequencing, methylation profiling, immunohistochemistry, and molecular tumor board review. These image-only predictions are intended to accelerate downstream molecular testing and tissue allocation while leaving final diagnosis and treatment selection anchored in validated molecular assays.

Hematologic diagnostics and especially cytomorphologic assessment are time-intensive and require high levels of expertise. Vision Language Models (VLM) show promise in medical image analysis in radiology and histopathology, while an evaluation on detecting acute myeloid leukemia (AML) is lacking. Our goal was to evaluate three Vision Language Models regarding their diagnostic accuracy and safety in clinical decision support in detecting AML from digitized bone marrow smears (BMS). Whole slide images were obtained from bone marrow smears of 50 AML patients and 50 bone marrow donors. Ten representative fields of view per sample were extracted manually. Three VLMs were used, two of which are considered generalist models (Qwen3.5-397B-A17B-FP8, GLM-4.6V-FP8), while the other one is a medically adapted model (Medgemma-27b-it). All models performed zero-shot analysis using two prompting strategies: First, a context-rich prompt requesting reporting of WHO/FAB diagnostic criteria in a structured manner, and secondly a minimal prompt without specific hematologic context. Overall diagnostic accuracy was poor for all models as they exhibited the overwhelming tendency to classify most samples as leukemic: With context-rich prompts, GLM4.6 identified 90% of leukemic samples while also labeling 92% of bone marrow donors as AML. The medical specialist model MedGemma-27b showed similar failure, misclassifying 86% of healthy donors and correctly detecting AML in only 66% of cases. Qwen3.5 performed best under detailed prompting, achieving a specificity of 0.26 and accuracy of 0.51. Accuracy of all models improved with context-free prompts (accuracies range 0.47-0.79), yet they still lacked the ability to correctly distinguish between leukemia and healthy bone marrow. Qwen3.5 was the only model to maintain meaningful specificity (0.64) and correctly identified 94% of AML, yielding an overall accuracy of 0.79. Morphologic feature-level agreement with human expert reports was poor across all models, indicating poor recognition of cell-level morphologies. This failure is likely driven by the fact that pathology imaging archives are vastly scraped during model training while hematological samples are not as widely available and therefore, hematology is an out-of-bounds use-case for these models, rendering them currently unsuitable for clinical decision support in hematology.

Precision oncology seeks to match each tumor with the most effective anti-cancer therapy. Advances in pharmacogenomics and machine learning enabled drug response prediction models with strong performance in cancer cell lines. Nonetheless, patient-centric evaluation of drug prioritization and systematic assessment of model generalization in patient-derived systems across cancer types remain largely absent. Here we introduce a translational framework combining patient-centric benchmarking with a pan-cancer pharmacogenomic atlas of patient-derived organoids, together with NELLY, a deep learning model integrating transcriptomic and chemical information to predict drug response and prioritize therapies. NELLY outperformed existing methods for patient-specific drug prioritization across cancer cell lines and patient-derived organoids, including under out-of-distribution evaluation. Its dynamic weighting mechanism provided patient-specific gene attributions, offering a route to connect predicted drug response to molecular programs associated with drug resistance. Our results support NELLY as a promising framework for translationally relevant and interpretable drug response prediction in precision oncology.

Functional ex vivo assays using live tumor tissues have demonstrated strong predictive accuracy for response to immune checkpoint inhibitors (ICIs) but are not scalable, requiring manual processing of large resections collected at academic centers. Here, an ex vivo live tumor fragment (LTF) platform was developed using standard-of-care biopsies from 228 patients with suspected malignancy collected across prospective, multicenter observational trials and biobanks. Hierarchical clustering of ICI-mediated changes in cytokine production identified two groups: responders and nonresponders. A binary classifier (elive index) using 8 cytokines achieved an AUC of 0.99 for cluster prediction. elive index correctly predicted clinical benefit in 93% (26/28) of patients (P = 3.2x10-5) and accurately identified 83% (10/12) of objective responders. Critically, elive responders were identified among biomarker-negative patients, highlighting the platform as a scalable approach that complements existing companion diagnostics and expands the population of patients identified to benefit from ICI therapy.