JCO Clinical Cancer Informatics

The increasing use of tumor sequencing has intensified the need for fast, traceable interpretation of genomic variants. General-purpose large language models can produce fluent answers, but unsupported statements, weak provenance, and stale knowledge limit their suitability for clinical genomics. We developed OncoGenRAG, a research framework that combines a parameter-efficiently fine-tuned BioBERT classifier with an entity-aware retrieval system over a curated, multi-source oncology knowledge base. The reported knowledge base contains 933 harmonized records derived from CIViC, ClinVar/dbSNP, Open Targets, UniProtKB/Swiss-Prot, Ensembl Variation, and linked PubMed literature. The classifier assigns one of five labels: Pathogenic, Likely Pathogenic, Variant of Uncertain Significance, Benign, or Oncogenic; the retrieval component ranks evidence records using subword TF-IDF similarity and explicit gene, variant, and cancer-type matches. A rejection rule suppresses answers when retrieval support is below a prespecified threshold. In the authors held-out evaluation, the classifier achieved 92.40% accuracy, 93.15% weighted precision, 92.40% weighted recall, and 92.65% weighted F1 score. In a separate benchmark of 100 clinical-style queries, OncoGenRAG achieved reported Precision@1 of 94.5%, Precision@3 of 96.8%, and 100% database grounding. No hallucinated answer was observed under the study operational definition, compared with a 41.0% no-hallucination rate for the ungrounded baseline. These results should be interpreted as internal validation rather than proof of universal safety because query construction, annotator agreement, class-specific performance, calibration, and external validation data were not available for independent analysis. OncoGenRAG provides a transparent design for evidence retrieval and abstention, but it is a research prototype and must not be used to select treatment without expert review.

Clinical records preserve patient findings and final recommendations but rarely capture how specialists select evidence, develop alternative interpretations, disagree, and converge on a management strategy. We describe a dataset derived from ten retrospectively selected, completed cases of histologically confirmed colorectal cancer. For each case, an administrator uploaded de-identified textual clinical records to an authenticated web-based annotation platform. A medical oncologist, surgical oncologist, and radiation oncologist, each with more than ten years of clinical experience, independently linked relevant source fragments to patient-specific interpretations and documented clinical problems, differential hypotheses, uncertainty, and diagnostic and treatment strategies. After the three records were locked, a large language model generated candidate questions for review by the tumour-board chair. The four physicians then discussed each case online; recordings were automatically transcribed and verified by two medical residents. The dataset comprises 30 source-linked specialist reasoning records, 10 verified case-level discussion transcripts, 10 final tumour-board decisions, and associated provenance and quality-control metadata. It supports research on evidence-grounded reasoning, cross-specialty disagreement, multidisciplinary decision-making, and multi-agent clinical decision support.

Background: Large language models (LLMs) may answer cancer questions plausibly while accepting a false presupposition (FP). Prompts that increase FP correction may also challenge questions in which no FP is present (NFP). We compared SkillOpt, an automated skill-optimisation method developed by Microsoft Research, with conventional prompt engineering across models. Methods: Using the physician-verified Cancer-Myth benchmark, GPT-5.4 optimised a written skill for GPT-4o in three independent 12-step executions. Training used 117 FP questions. Rewrites were selected by combined accuracy on a validation set of 58 FP and 30 NFP questions, then tested on 410 unseen FP questions. A separate GPT-4o judged responses. Across seven LLMs, we compared no correction instruction, a deliberately strong seven-example few-shot prompt, three GPT-5.4-generated instruction-only prompts, and full and compact SkillOpt skills. Further tests examined portability, new questions and valid clinical assumptions. Results: SkillOpt increased GPT-4o FP correction from 15.4-17.6% to 82.2-90.5%, gains of 66.3-72.9 percentage points (95% bootstrap confidence intervals 61.7-77.3; all Holm-adjusted p<0.001). Only two or three rewrites were accepted per execution. The first full skill achieved 92.7-96.6% correction on three recipient models but 6.3-16.3% on two weaker models, where few-shot prompting achieved 44.6-54.6%. On Gemini 3.5 Flash and GPT-5.6 Luna, compact SkillOpt retained 90.0-92.9% correction while increasing NFP accuracy from 59.2-79.2% with the full skill to 85.8-93.3%. Both approaches sometimes challenged valid clinical assumptions. Conclusions: SkillOpt produced large, repeatable gains, but it was neither universally portable nor free of over-correction. Automated and conventional prompts should be compared for each intended model using tests of both FP correction and NFP accuracy. Compact SkillOpt may improve this balance on capable models.

Purpose: Digital-twin frameworks can potentially offer a pathway toward individualized treatment but remain largely untested in radiation oncology. We present a causal digital-twin model for head-and-neck (H&N) cancer that estimates patient-specific survival benefit from chemoradiotherapy (CRT) versus radiotherapy (RT) alone, stratified by HPV status. Methods: Using the public RADCURE cohort (n = 3,346), we developed a causal digital-twin framework in which patients were divided into HPV-positive and HPV-negative groups and assigned to training (RT start year [&le;] 2007) and test sets (RT start year > 2007). Bayesian-network structure learning identified each cohort's Markov Blanket around the 2-year survival endpoint (SVy2), yielding the minimal causally relevant feature set. A causal-forest model (double machine learning, DML framework) estimated the individualized treatment effect {tau}=P(SVy_2|CRT,X)-P(SVy_2|RT,X), with {tau} denoting the model-estimated effect. Overlap restrictions were applied to mitigate extrapolation. Predictive discrimination (AUC, C-index), average treatment effects (ATE), individualized effects (ITE), and decision-curve analyses were computed. Results: For HPV-negative disease, internal evaluation demonstrated strong discriminative performance (test AUC = 0.72; C-index = 0.81, p < 0.01). The estimated test-set ATE favored CRT with a modest but positive improvement in SVy2 (mean {tau} = 0.13). Patterns in {tau} were clinically coherent, including increased benefit in node-positive and good-performance patients and reduced benefit in high-burden or poor-performance presentations. Decision-curve analysis showed that a {tau} -guided selective-CRT strategy provided higher net benefit than uniform treatment policies. In HPV-positive disease, treatment-effect heterogeneity was attenuated, consistent with its more uniformly favorable prognosis. Conclusion: This study demonstrates a proof-of-concept causal digital-twin framework for interpretable, patient-specific survival simulation in H&N cancer. The model captures clinically plausible treatment-effect heterogeneity and establishes a foundation for multi-cohort validation and future integration into individualized treatment decision support.

Background: Much of the symptom burden in colorectal cancer (CRC) patients is documented in unstructured discharge-note narrative, and manual extraction is not scalable. Whether large language models (LLMs) outperform rule-based and named entity recognition (NER) methods has not been rigorously benchmarked. Objective: To benchmark rule-based, NER, and zero-shot LLM methods for extracting 46 cancer-related symptoms from CRC discharge notes against an adjudicated ground truth. Methods: We analyzed 2,704 discharge notes from CRC patients in MIMIC-IV. A 46-symptom target list was built from the Memorial Symptom Assessment Scale and the EORTC QLQ-CR29. Four approaches -- dictionary-based rule matching, pretrained clinical NER, and zero-shot Claude Haiku and Gemini 3.5 Flash -- plus two hybrid variants (LLM output with post-hoc rule-based negation filtering) were evaluated against a 200-note gold standard adjudicated by two raters (pooled kappa=0.71, macro kappa=0.49), using Macro/Micro F1, precision, and recall. Results: Gemini 3.5 Flash performed best (Macro F1=0.70, Micro F1=0.86, Macro Precision=0.74), followed by Claude Haiku (Macro F1=0.63, Macro Recall=0.71); both substantially outperformed rule-based (Macro F1=0.44) and NER (Macro F1=0.38) methods. Post-hoc negation filtering paradoxically degraded LLM performance (Gemini+Hybrid Macro F1=0.58; Claude+Hybrid Macro F1=0.54) by overriding correct predictions through rigid, fixed-window matching. Conclusions: Zero-shot LLMs substantially outperform rule-based and NER approaches for CRC symptom extraction; post-hoc negation correction should not be applied to LLM outputs without syntactic scope validation. Implications for Practice: Zero-shot LLM extraction offers a scalable, accurate alternative to manual chart review and traditional NLP pipelines for oncology symptom surveillance, without institution-specific rule development or model training.

Background: Adverse event (AE) coding is essential for safety monitoring in oncology clinical trials, particularly in acute myeloid leukemia (AML), where intensive therapies are associated with frequent and heterogeneous toxicities requiring standardized MedDRA (Medical Dictionary for Regulatory Activities) coding. However, manual Low-Level Term (LLT) assignment remains labor-intensive, subjective, and difficult to scale. Although large language models (LLMs) have emerged as promising decision-support tools for automated coding, unguided zero-shot generation remains insufficient for reliable fine-grained MedDRA coding. Objective: To develop and evaluate a retrieval-augmented reasoning pipeline for clinically aligned LLT-level MedDRA coding of free-text adverse events from prospective AML clinical trials. Methods: We implemented a retrieval-augmented reasoning pipeline inspired by the retrieval-augmented generation (RAG) paradigm using LLaMA-3.3-70B-Instruct as the primary backbone and benchmarked the framework across multiple open instruction-tuned LLMs. Dense semantic retrieval first generated a constrained top-100 LLT candidate set for each AE, followed by structured LLM reasoning to select a single best-matching LLT and deterministic mapping to Preferred Term (PT) and System Organ Class (SOC) levels. The pipeline was evaluated retrospectively on AE datasets from three prospective AML clinical trials (MOSAIC, DELTA, and DaunoDouble) with automated LLT/PT/SOC metrics and expert-assessed Clinical Correctness Rate (CCR). Results: Clinical expert review showed high clinical acceptability of the RAG pipeline across datasets (91-97%). Under automated evaluation, the pipeline achieved LLT exact accuracy of 50-58%, PT accuracy of 78-85%, and SOC accuracy of 90-93%. Zero-shot generation and random candidate selection performed substantially worse. Semantic retrieval more often included the coder-assigned LLT among the candidate terms available to the model than retrieval based on lexical similarity. Multi-model benchmarking showed that backbone choice mainly affected LLT exact agreement, whereas PT and SOC performance remained comparatively stable. Conclusions: Retrieval-augmented reasoning supports clinically aligned MedDRA coding of free-text adverse events under realistic candidate constraints in AML clinical trials. Evaluation across three AML clinical trials showed that strict LLT-level string agreement underestimated clinical ap-propriateness, highlighting the importance of combining hierarchical evaluation metrics with clini-cal expert validation for AI-assisted MedDRA coding in hematology trials.

Cancer registries are essential for population-level surveillance, epidemiologic research, and public health planning, but their accuracy depends on human review of key variables documented in unstructured clinical text. This is time-consuming, and current quality assurance (QA) typically relies on reviewing fewer than 10% of records at random, which can miss facility- or time-specific errors. We developed a large language model (LLM)-based framework for targeted cancer registry QA. The framework uses LLMs to extract biomarker values from clinical text, compares them with registry entries, and flags disagreements for expert review. Six open-source LLMs were benchmarked locally in a privacy-preserving environment, and the best model per site and report type was applied to breast and prostate cancer, two of the most common cancers in the United States. For the four breast cancer biomarkers ER, PR, HER2, and Ki-67, the disagreement rate against quality-assured registry values in 1,852 breast cancer cases was lowest for ER at 1.67% and highest for Ki-67 at 9.18%. In 920 prostate cancer cases, the disagreement rate with the deployed models stayed below 8% for every categorical Gleason variable, while the exact-value PSA metric reached 17.28%. PSA therefore generated the largest volume of candidate discrepancies, though individual PSA flags require more corroboration than categorical flags. Applied to a full year of records (11,740 breast and 6,792 prostate cases), the framework flagged disagreements and assigned each a probable cause, grounded in NAACCR coding rules, to prioritize expert review. Extracted values are linked to their supporting text, and model, prompt, and pipeline versions are logged for reproducibility. By directing review toward records most likely to contain discrepancies, the framework offers an efficient, auditable alternative to random QA sampling for cancer registries.

PURPOSE: Radiology and treating-oncologist documentation are recorded on different schedules in routine oncology care. We characterized the opposite clinical source at its first evaluable opportunity after radiology- or medical-oncology-documented progression in advanced pancreatic cancer using an exhaustive six-state framework. METHODS: We performed a retrospective secondary analysis of AACR Project GENIE BPC PANC v1.0-public. Patients entered the treatment-index cohort at the first systemic regimen initiated on or after advanced-disease onset. Process T0 was the first genuine radiology- or medical-oncology-documented progression after cohort entry and not after subsequent same-cancer treatment or death. The opposite source's first evaluable opportunity was classified as same-day concordance, confirmation, explicit nonprogression, mixed, indeterminate, or no evaluable opportunity. The secondary estimand was explicit nonprogression among first opportunities classifiable specifically as progression or explicit nonprogression. RESULTS: Of 1,109 patients, 882 entered the treatment-index cohort and 645 qualified for the primary estimand. States were same-day concordance in 25 (3.88%), confirmation in 219 (33.95%), explicit nonprogression in 108 (16.74%), mixed in 19 (2.95%), indeterminate in 166 (25.74%), and no evaluable opportunity in 108 (16.74%). Among 327 progression/nonprogression-classifiable first opportunities, 108 (33.03%; 95% CI, 28.15%-38.30%) documented explicit nonprogression. Opposite-source opportunity occurred after 88.47% of imaging-first events at a median of 5 days versus 62.94% of medical-oncology-first events at a median of 38 days. A known-regimen-completion sensitivity yielded 84/144 (58.33%). CONCLUSION: Cross-source progression assessment frequently did not yield immediate confirmation and often remained indeterminate or unevaluable. An opportunity-adjusted six-state framework makes observation-process and clinical-boundary dependencies explicit rather than compressing them into a binary discordance estimate.

Background: Adjuvant therapy decision in early-stage node-negative breast cancer remains one of the most challenging problems in oncology. Although most patients in this population achieve favorable outcomes with endocrine therapy alone, a subset remains at risk of relapse and may benefit from chemotherapy. Accurate identification of high-risk and low-risk patients is therefore essential to prevent both undertreatment and overtreatment. Methods: In this study, machine learning models were assessed using MSK-IMPACT (n=645) and METABRIC (n=503) cohorts comprising node-negative stage I-II breast cancer patients. Three experimental settings were evaluated: training on MSK-IMPACT and testing on METABRIC, the reverse direction, and five-fold cross-validation on the combined cohort. Sensitivity analysis and data harmonization were applied to identify and mitigate cohort-specific bias. Survival analysis, SHapley Additive exPlanations (SHAP) interpretability, and ablation analyses were performed to validate clinical relevance and characterize feature contributions. Results: The combined cohort model achieved an area under the receiver operating characteristic curve (AUC) of 0.788{+/-}0.028 across five-fold cross-validation, while bidirectional cross-cohort validation yielded AUC of 0.741 when testing on MSK-IMPACT and AUC of 0.714 when testing on METABRIC. Survival validation on METABRIC out-of-fold predictions from the combined cohort model showed hazard ratios of 3.01 (95\% CI: 2.28-3.98) for disease-free survival and 2.35 (95\% CI: 1.79-3.09) for overall survival, consistent across Luminal A, Luminal B, and Triple-Negative subtypes. Conclusion: Validated across two geographically independent cohorts, the proposed model enables cost-effective, gene expression-free identification of patients at high-risk of distant relapse and may facilitate adjuvant chemotherapy decision-making in early-stage node-negative breast cancer.

Oncology knowledge spans structured and unstructured datasets of drugs, regimens, conditions, and identifiers, as well as clinician-authored narratives describing treatment sequencing across disease settings. Answering clinical questions may require accessing both databases and free-text narratives, but existing intelligent retrieval systems favor either relational facts or narrative context. We developed HemOncAgent, an artificial intelligence agent that selects among retrieval tools for the HemOnc knowledge ecosystem: HemOncKB, a curated knowledge graph for pharmacologic relationships, and HemOnc.org, a narrative resource for care pathways. Across structured and narrative benchmarks, HemOncAgent maintained high-fidelity retrieval across question types for which single-source systems showed domain-specific limitations, supporting hybrid tool-based retrieval for more reliable oncology knowledge access.

Background. Accurate delineation of the gross tumour volume (GTV) - primary tumour (GTVp) and nodal disease (GTVn) - on FDG-PET/CT is a critical step of head and neck radiotherapy planning. Comparisons between lightweight custom networks and the auto-configured nnU-Net v2 are usually reported as end-to-end pipelines, conflating the contribution of the network with that of the inference-time post-processing applied on top of it. We separated the two. Methods. MiniUNet3D (custom 3D U-Net, 18.3 M parameters) and nnU-Net v2 (3d_fullres, 88.2 M parameters) were trained on the same 578 FDG-PET/CT cases (85/15 author-defined split of the HECKTOR 2025 Task 1 set, 8 centres) and evaluated on the same internal cohort. Three arms were compared pairwise: MiniUNet3D raw output at a fixed 0.5 threshold, MiniUNet3D with a locked adaptive post-processing pipeline, and nnU-Net v2. Comparisons used paired Wilcoxon tests with bootstrap confidence intervals, Bonferroni and Benjamini-Hochberg correction, and Cohen's d; catastrophic failure (Dice < 0.01) was compared with an exact McNemar test. Cases with an empty reference for a given target were excluded from that target's analysis (n = 98 GTVp, n = 93 GTVn). Results. With post-processing matched off, nnU-Net v2 was superior: median GTVp Dice 0.799 versus 0.592 (mean difference -0.244, 95 % CI -0.300 to -0.191; d = -0.88) and GTVn 0.774 versus 0.598 (d = -0.82). Post-processing raised MiniUNet3D to 0.800 (GTVp) and 0.738 (GTVn), recovering 79 % of that difference. Post-processed, MiniUNet3D matched nnU-Net v2 on GTVp Dice (p = 0.113) but remained inferior on nodal disease after Bonferroni correction (Dice p = 0.041; surface Dice p = 0.049). Catastrophic GTVp failures were 25/98 raw, 8/98 post-processed and 1/98 for nnU-Net v2 (McNemar p = 0.016). Inference took 34 s versus 78 s per case on the same GPU. Conclusions. Post-processing recovered most, but not all, of the difference between the two models, and it did not confer robustness: an eight-fold higher rate of empty contours on small primaries persisted, which is the more consequential difference for planning safety. Pipeline comparisons reported without a post-processing ablation risk attributing to a network what post-processing supplied.

Background: Reported performance for multi-modal cancer prognostic models is often obtained from a single train-test split and without comparison against a simple clinical baseline. We evaluated an integrated framework combining language-model representations of clinical records with multi-omics data under a leakage-controlled protocol. Methods: Clinical variables from 589 TCGA-BRCA patients were converted to natural language sentences and encoded with BioBERT; four molecular modalities were reduced by Sparse principal component analysis. Clinical-only, multi-omics-only and integrated feature sets were evaluated with Cox regression, XGBoost and random survival forest (RSF) using repeated stratified cross-validation with pooled out-of-fold predictions, against a marginal Kaplan-Meier null and a conventional age-and-stage Cox model. Results: Among 589 patients with 88 deaths (14.9%; median follow-up 33.7 months), the integrated RSF model performed best (Harrell's C 0.685, 95% CI 0.620-0.754; Uno's C 0.748), exceeding the conventional clinical Cox model (0.637, 0.559-0.707). The ordering integrated > clinical-only > multi-omics-only held in all nine feature-set by algorithm comparisons and at all three component settings tested. The integrated model significantly outperformed every multi-omics-only model (P = 0.008-0.044) but not clinical-only models (dC = 0.037, P = 0.128). Brier scores approached the null reference at 2 and 3 years and improved modestly at 5 years. Conclusion: Integration gave the highest discrimination and exceeded a conventional clinical model, but the gain over clinical features within the same pipeline was not statistically resolvable at this event count. Reporting simple baselines, null-model calibration and cross-validated estimates should be standard in multi-modal prognostic modelling.

CXR report generation may require a vision-language model (VLM) to produce both textual findings and spatial bounding boxes. Generative 4B-7B VLMs can emit non-empty outputs on normal images and empty outputs on abnormal images, motivating explicit structural routing. To evaluate whether a hard inference-time gate before a probabilistic VLM changes output-presence performance and to identify the mechanisms underlying paired STRUCT outcomes. We evaluated CXRxVLM v2, combining a frozen microsoft/rad-dino ViT-B/14 encoder with a 768[-&gt;]1 logistic probe (threshold 0.0557) and google/medgemma-4b-it with the pamessina/medgemma-4b-it-cure LoRA adapter. A seed=42 stratified cohort of 500 VinDr-CXR train-pool images (250 NORMAL, 250 ABNORMAL) was compared with Lingshu-7B A_baseline and D_fewshot configurations. Exact paired McNemar tests and stratum-level output-presence analyses were prespecified for the primary configurations; MedGemma 1.5 SigLIP was exploratory. CURE achieved STRUCT = 78.0% (390/500; Wilson 95% CI 74.2-81.4), versus 73.8% for Lingshu A_baseline and 74.2% for D_fewshot. Pairwise p-values were 0.0778, 0.1042, and 0.8642. The paired decomposition showed CURE ABNORMAL non-empty-output advantage of +13.6 percentage points versus Lingshu A (p = 0.0012; +14.0 points versus D, p = 0.0007), while Lingshu had higher NORMAL empty-output rates (+5.2 to +6.4 points; p = 0.0106 and p = 0.0004). The full pipeline used 8.87 GB VRAM and 4.92 s/image mean latency; 53% of records used a 25.7 ms warm gate-negative path after model loading. Equivalent overall STRUCT scores concealed two mechanistically different output regimes: CURE favored ABNORMAL non-empty outputs, whereas Lingshu favored NORMAL empty outputs. This paired decomposition, rather than the aggregate score alone, characterizes how hard-gated and probabilistic systems route output presence.

Voxel-level prediction of treatment response from longitudinal fluorodeoxyglucose (FDG) positron emission tomography/computed tomography (PET/CT) can enable spatially adaptive dose planning in advanced non-small cell lung cancer (NSCLC), but point predictions alone lack the uncertainty quantification needed for safe clinical decision-making. Standard conformal prediction (CP) provides distribution-free coverage but yields spatially uniform intervals and assumes exchangeability, an assumption violated by intra-tumoral spatial correlation. We propose a multiscale CP framework built on a re-optimized VoxelForecast generalized least squares model that explicitly captures voxel-level covariance. Variogram sensitivity analysis identified the Stable model as significantly superior to the previously used Matern family. The framework combines hierarchical nested leave-one-patient-out calibration, which restores exchangeability for valid voxel-level coverage, with residual-variance CP (VarCP), which learns a heteroscedastic, location-dependent scale to produce adaptive intervals while preserving finite-sample marginal coverage guarantees. On two prospective NSCLC trial cohorts, FLARE-RT (N=25 locally advanced, chemoradiotherapy; 11,100 voxels) and BRIGHT (N=19 metastatic, chemoimmunotherapy; 26,980 voxels), multiscale feature integration reduced voxel-level prediction error by 35% and 31%, respectively. VarCP maintained nominal coverage while producing 7-18% (FLARE-RT) and 4-8% (BRIGHT) narrower voxel-level intervals than a hierarchically calibrated standard CP baseline (paired Wilcoxon signed-rank, significant at most miscoverage levels), with comparable lesion-level gains (7-18% and 6-10%) via median-based aggregation of voxel residuals. VarCP delivers tighter yet mathematically guaranteed spatial uncertainty maps, supporting both voxel-level dose adjustment and whole-lesion treatment planning in advanced NSCLC.

Living clinical guidelines revise recommendations as evidence emerges, so an AI answer can be faithful to a real guideline and still be out of date. Here we examine this failure using ASCOBench, 288 unique questions in 96 three-turn conversations grounded in versioned American Society of Clinical Oncology (ASCO) breast and prostate cancer guidelines, with oncologist-reviewed reference answers. Oncologists corrected one in four version-sensitive answers drafted by a frontier model, against fewer than one in ten factual ones. Across the guideline corpus, 19 recommendations changed between versions, seven of them reversals, yet only 1 of 14 superseded documents states that it has been replaced. With a strong model, retrieval over several guideline versions made stale answers four times more frequent than answering without retrieval. SentryLine, a verification-first system, lowered incorrect answers to 4.2-5.2% from 9.7-22.9% for baselines across three models. Recognizing change before a guideline declares it remains an open problem.

Background. Residual cancer burden (RCB) after neoadjuvant chemotherapy (NAC) offers finer prognostic stratification than binary pathologic complete response, and increasingly guides adjuvant treatment intensity. Predicting four-tier RCB class from preoperative data could inform adjuvant planning before surgery, yet this remains an unmet need; and when two models reach equal discrimination, the key question is which generalizes most reliably. We compared a radiology-focused model with a fully integrated multimodal model for preoperative four-class RCB prediction. Methods. In a single-center, retrospective cohort of 328 patients treated with NAC followed by surgery, 64 clinicopathologic and radiologic variables were organized into thematic blocks. Two configurations were compared: a 17-variable radiology model (Model R) and a 62-variable multimodal model (Model ALL). Three algorithms (Random Forest, XGBoost, LightGBM) were evaluated with and without SMOTE using an 80/20 stratified split and 5-fold cross-validation. Model selection combined test AUC, macro-F1, cross-validation-to-test gap, nested cross-validation, bootstrap confidence intervals, and SHAP explainability, following the TRIPOD+AI guidance. Results. RCB classes were distributed as RCB-0 27.4% (n=90), RCB-I 10.4% (n=34), RCB-II 43.6% (n=143), and RCB-III 18.6% (n=61). Model R and Model ALL reached identical test AUC (0.838). Model ALL, however, achieved higher accuracy (0.636 vs 0.530) and macro-F1 (0.602 vs 0.598), together with a substantially smaller cross-validation-to-test gap (0.015 vs 0.099), pointing to more stable generalization; this gap difference persisted across all three algorithms. SHAP analysis showed that the multimodal model drew jointly on imaging phenotype, tumor biology, and disease extent. Both models remained weakest in the RCB-III class. Conclusions. At equivalent discrimination, the multimodal model was methodologically preferable for preoperative RCB prediction, owing to its stability and interpretability - qualities relevant to trustworthy clinical decision support. It remains investigational; a model flagging likely RCB-0 or RCB-III before surgery could prioritize adjuvant-therapy discussions earlier in the care pathway, pending prospective external validation.

The rapid expansion of the oncology literature has outpaced manual curation of clinically relevant gene-cancer-drug associations and oncogenic driver evidence. Existing automated approaches often lack transparency or are difficult to scale across heterogeneous data sources. To address this gap, we developed megaMine, a transparent, rule-based, and context-aware literature-mining framework that integrates therapeutic and driver evidence from PubMed, PubTator, and Europe PMC by combining entity recognition, hierarchical heuristics, and contextual labeling. In therapy mode, megaMine was applied to approximately 100,000 oncology articles published between 2015 and 2025, yielding more than 23,000 structured sentence-level evidence records, with standardized annotations for drug response, resistance, and study context. Internal evaluation of context labels showed the strong separability between efficacy and non-efficacy evidence using ridge logistic regression (AUROC = 0.915; AUPRC = 0.941). Benchmarking against NCI/OncoKB-supported drug-cancer associations showed that curated clinical associations had higher megaMine composite evidence scores than unlabeled comparison pairs [median (IQR): 25.6 (9.07-72.5) vs. 3.61 (1.69-8.69); Wilcoxon rank-sum test, P < 2.2 x 10-16]. In driver mode, megaMine retrieved mutation- and biomarker-related evidence from an ERBB-focused gastric cancer query, generating 750 evidence rows from 200 PMIDs. These results demonstrate that deterministic and interpretable approaches can support scalable evidence extraction for downstream applications such as knowledge graph construction and literature-based evidence synthesis.

Background: Clinicopathologic (SARCULATOR) and transcriptomic (CINSARC) models are established prognostication tools in soft tissue sarcoma, but neither directly represent the sequence context of tumor genomic alterations. Evo2 is a genomic foundation model trained to learn the grammar of DNA/RNA sequences. We tested whether Evo2-derived features from tumor short variants, fusions, and rearrangements improve survival prognostication beyond clinicopathologic and transcriptomic risk models. Methods: We analysed formalin-fixed, paraffin-embedded tumor material and clinical data from 102 patients with dedifferentiated liposarcoma (n = 25), high-grade leiomyosarcoma (n = 25), and undifferentiated pleomorphic sarcoma (n = 52). A separate TCGA-SARC cohort of comparable histologies (n = 129) served as external validation. Reconstructable alterations (short variants, fusions and rearrangements) were converted into sequence inputs, embedded with Evo2 and aggregated into patient-level genomic representations. Cox proportional hazards models incorporated genomic features alone (Evo2) or together with clinicopathologic (SARCULATOR) or transcriptomic (CINSARC) predictors across overall survival (OS), progression-free survival (PFS), disease-free survival (DFS), relapse-free survival (RFS), and metastasis-free survival (MFS). Discrimination was assessed using Harrell's concordance index (C-index). Results: Evo2 features were generated from 174 study-cohort sequences and 7298 TCGA-SARC sequences. In the study cohort, adding Evo2 to SARCULATOR increased the reported C-index for OS from 0.620 to 0.770, PFS from 0.625 to 0.755, DFS from 0.612 to 0.714, and RFS from 0.633 to 0.771. In TCGA-SARC, SARCULATOR plus Evo2 increased the OS C-index from 0.595 to 0.716 and the MFS C-index from 0.548 to 0.667. Also in TCGA-SARC, adding Evo2 to CINSARC increased the OS C-index from 0.476 to 0.653 and the MFS C-index from 0.553 to 0.650. These are descriptive point-estimate comparisons. Moreover, in TCGA-SARC Evo2 alone outperformed SARCULATOR alone and CINSARC alone, supporting independent prognostic information within altered sequence-context representations. Conclusions: Evo2-derived genomic representations improved clinicopathologic and transcriptomic prognostication across sarcoma cohorts and survival endpoints. These findings support genomic foundation models as an independent and complementary layer for sarcoma risk stratification through direct modeling of altered tumor sequence biology. Confirmation of model transportability, calibration and clinical utility is required before clinical application.

Background: Precision oncology relies on accurate interpretation of tumour-detected gene variants, to guide personalized treatment decisions. However, accurate interpretation of variants in context requires extensive information that is often buried within unstructured biomedical literature and obscured by inconsistent nomenclature, making manual retrieval labour-intensive and prone to omissions. Methods: To address this challenge, we developed Variantscape, a large-scale, automated pipeline and open-access web tool. It integrates traditional natural language processing methods with state-of-the-art large language models to extract, standardize, and analyze co-associations between genetic variants, cancer types, and therapeutic interventions from published biomedical abstracts. Findings: From over 3 million abstracts screened, 335,817 gene name-containing articles were eligible for downstream extraction. Among these, 7,423 (2.2%) simultaneously mentioned a variant, cancer type, and therapeutic agent, encompassing 3,902 unique variants across 98 cancer types and 388 therapeutic agents. This highlights the inefficiency of manual literature retrieval in molecular tumour board (MTB) workflows. Network analysis revealed 14,831 statistically significant co-associations, represented in a literature-derived graph with 4,388 nodes and 46,943 edges. Canonical alterations in well-studied cancers (e.g., BRAF V600E in melanoma) were strongly linked to established treatments, while several rare variants also emerged with high-confidence literature support. Interpretation: By applying large language models to biomedical literature, Variantscape enables scalable, context-aware extraction of trilateral variant-treatment-cancer relationships. This approach supports early evidence synthesis/hypothesis generation, highlights underrecognized or rare associations, and offers a practical resource for accelerating discovery and supporting precision oncology research and translation. Unlike static databases, Variantscape is continuously updatable and leverages large language model-based inference to uncover putative associations without manual curation. Variantscape has the potential to support MTB workflows and translational research by rapidly revealing signals from underlying abstracts.

Objective: To determine how cohort versus nested case-control (NCC) design affects EHR-based prediction of a rare outcome. Materials and Methods: We compared two cohort and five NCC designs for predicting early-onset gastrointestinal cancer using L1-regularized logistic regression (LASSO), XGBoost, and multilayer perceptron (MLP). Models underwent nested five-fold cross-validation with model-specific hyperparameter tuning and temporal evaluation in a held-out prospective cohort representing intended deployment. Model discrimination was assessed using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). Model calibration was assessed using Brier scores and calibration plots before and after recalibration with isotonic regression. Feature-ranking similarity was quantified using rank-biased overlap. Results: The cohort development datasets included 256,854-415,529 patients, with 2-yr cumulative incidence of 0.12%-0.17%, compared with approximately 9% in the 1:10 matched NCC datasets. NCC designs generally produced higher internal discrimination, but temporal performance varied by algorithm and index-date definition. Cohort designs achieved the highest temporal AUPRC for LASSO (0.068), whereas diagnosis-anchored NCC achieved the highest temporal AUPRC for XGBoost (0.086) and MLP (0.070). Temporal AUPRC generally decreased as the required lead time before diagnosis increased. Raw predicted probabilities substantially overestimated observed risks in the temporal cohort. Isotonic recalibration improved probability scaling while largely preserving discrimination, although residual miscalibration remained. Cohort-based feature rankings were more consistent across algorithms, whereas NCC-based rankings depended more strongly on algorithm and index-date anchoring. Conclusion: Study design should be selected and evaluated according to the intended clinical prediction task and deployment setting.