Ung, Choong Yong

Background/Objectives: Glioblastoma surveillance after surgery and chemoradiation remains challenging because MRI findings of tumor recurrence can overlap with pseudo-progression, treatment-related effects, and postoperative tissue changes. Methods: We developed a seven-channel postoperative MRI framework using a deep learning transfer method to support non-invasive glioblastoma treatment-effect assessment. The model used T1 contrast-enhanced, FLAIR, T1, RSI-Cell, ADC, T2, and cerebral blood flow volumes from 124 postoperative glioblastoma patients (164 MRI timepoints) as input. Images were processed using a 3D ResNet18 encoder pretrained on 588 postoperative glioma samples and fine-tuned using task-specific classification heads. The cohort comprised 124 patients contributing 164 postoperative MRI timepoints, all acquired at 3T on scanners from a single vendor. Performance was evaluated with nested five-fold cross-validation stratified and assigned at the patient level, so that all timepoints from a given patient fell in one-fold and the training epoch was selected on an inner split rather than on the fold being reported. Because some of the clinical labels were incomplete, the number of evaluable timepoints differed by task (recurrence versus pseudo-progression, 164; MGMT, 99; short-term survival, 139). The whole procedure was repeated under three independent random seeds and results are reported as the mean and standard deviation across seeds. Results: The strongest clinical endpoint was recurrence versus pseudo-progression, where nested cross-validation across three random seeds gave a pooled out-of-fold AUC of 0.935 (SD = 0.014), area under the precision-recall curve of 0.973, balanced accuracy of 0.880, sensitivity of 0.917, and specificity of 0.843. No other endpoint reached reliable discrimination. Radiation decision reached an AUC of 0.658 (SD = 0.039), while MGMT promoter methylation (AUC = 0.532, SD = 0.082) and short-term survival (AUC = 0.514, SD = 0.027) were indistinguishable from chance. Conclusions: In this single-center proof-of-concept study, postoperative MRI successfully distinguished tumor recurrence from pseudo-progression. However, the models did not reliably predict the other three outcomes related to molecular status, treatment planning, and prognosis. Overall, the model learned imaging features specifically associated with recurrence, rather than a more general representation of the tumor that can predict many different clinical outcomes. These findings support technical feasibility for a single endpoint rather than clinical readiness, for which external multi-center validation is required.

Accurate estimation of Alzheimer’s disease (AD) severity from structural magnetic resonance imaging (MRI) remains difficult, as disease-associated anatomical alterations are often subtle and publicly available datasets are typically too small to support robust deep learning model training. This feasibility study sought to determine how much Alzheimer’s disease spectrum-related information could be extracted from a small structural MRI cohort using a deliberately lightweight two-dimensional convolutional neural network (2D CNN), and whether transfer learning improves model performance. This study was intended as a methodological proof of concept rather than the development of a clinically deployable diagnostic tool. Structural scans and Clinical Dementia Rating (CDR) labels from the OASIS-1 dataset were filtered to 214 subjects: 124 cognitively normal (CN), 65 with mild cognitive impairment (MCI; CDR = 0.5), and 25 with AD-level impairment (CDR ≥ 1). A compact 2D CNN trained from scratch and a transfer learning model (frozen ImageNet MobileNetV2 features) were evaluated on four binary tasks (CN vs. AD, MCI vs. AD, CN vs. MCI, and CN vs. any impairment) under identical pre-processing and subject-level repeated 5-fold cross-validation (10 repeats), with the decision threshold tuned only on an inner split. Discrimination was summarized by ROC-AUC with 95% confidence intervals (CIs), permutation tests against chance, and per-task sensitivity and specificity. The from-scratch CNN recovered only a broad normal-versus-impaired signal (CN vs. any impairment AUC 0.59) and was at chance on adjacent-stage tasks (MCI vs. AD 0.41; CN vs. MCI 0.51). Transfer learning improved every task: CN vs. AD AUC 0.745 (95% CI 0.730–0.763), CN vs. any impairment 0.642, CN vs. MCI 0.601, and MCI vs. AD 0.599. On an independent OASIS-2 cohort, the transfer learning CN vs. AD model retained AUC 0.748. In this small-data regime, transfer learning recovers substantially more Alzheimer-spectrum signals than a from-scratch CNN, but performance remains modest because it is bounded by CDR-based, non-biomarker-confirmed labels, suggesting the model separates CDR-defined cognitive-status groups rather than detecting AD pathology.

Abstract Neuroblastoma (NB) accounts for 10% of pediatric cancer-related deaths. High-risk NB is frequently metastatic at diagnosis and has a poor 5-year survival rate (<50%) despite aggressive treatment. GD2, a disialoganglioside glycolipid that is expressed in nearly all NBs, has been validated as a promising immunotherapy target. Although GD2-CAR-T cells have been clinically evaluated, their efficacy has been limited by poor tumor infiltration and the immunosuppressive tumor microenvironment (TME), highlighting the need for alternative immune-based strategies. With inherent capacity of macrophages to infiltrate tumors, phagocytose cancer cells, and modulate TME, they represent a promising alternative strategy for immunotherapy in NB. In this study, we developed and evaluated GD2-CAR-macrophages (GD2-CAR-Ms) as a novel therapeutic approach for NB. GD2-specific CAR constructs were designed, validated, and introduced into multiple human and murine monocyte/macrophage cell lines via lentiviral transduction. Fluorescent microscopy and flow cytometry confirmed a significant enhancement in the phagocytosis of GD2-high NB cells by GD2-CAR-Ms containing Hu3F8 single-chain variable fragment (scFv). A strong positive correlation was observed between GD2 expression levels on the target cells and phagocytosis efficiency of GD2-CAR-Ms, indicating that the phagocytic activity of GD2-CAR-Ms is antigen-specific. Anti-tumor efficacy was further evaluated in vivo using an immunocompetent Mycn-expressing NB non-genetically engineered mouse model (MYCN-NGEMM). Mice treated with GD2-CAR-Ms exhibited significantly reduced tumor burden and prolonged survival compared to control groups. Notably, treatment of mice bearing GD2-high tumors resulted in markedly improved survival, including one complete remission among ten treated animals, underscoring the therapeutic potential of GD2-targeted macrophage therapy. Histological analysis showed increased deposition of connective tissues and decreased tumor cells density, suggesting TME remodeling. Furthermore, immunohistochemistry reveals increased immune cell infiltration, particularly of macrophages, neutrophils and T cells, indicating a favorable shift in the immune TME from “cold” to “hot” following GD2-CAR-M treatment. Additionally, inhibition of tumor growth was also observed in our zebrafish NB model after GD2-CAR-Ms treatment, further supporting their therapeutic potential. Together, our findings demonstrate that GD2-CAR-Ms exhibit potent antitumor activity and reshape the tumor immune microenvironment, supporting their translational potential as a therapeutic strategy for high-risk NB. Citation Format: Ke-En Tan, Kok Siong Yeo, Cheng Zhang, Caleb Baker, Shyang Hong Tan, Alexis M. Sosic, Yat-Yuen Lim, Choong-Yong Ung, Cristina Correia, Stephanie F. Polites, Yi Lin, William A. Weiss, Hu Li, Shizhen Zhu. Engineering macrophages as a novel therapeutic approach for high-risk neuroblastoma [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 7810.

Abstract Background: Symmetry principles, a long foundational framework in physics and chemistry, have rarely been applied to understand biological phenotypes especially in cancers. Here we examine whether symmetric relationships in gene expression can characterize and differentiate healthy from disease conditions. Method: To test this concept, we built a hybrid machine-learning framework, Learning-Based Invariant Feature Engineering (LIFE), that applies two symmetric invariant feature functions, IFF1 and IFF2, to all possible gene pairs in bulk transcriptomic datasets to identify invariant feature genes (IFGs) - gene pairs whose transformed expression values by either IFF1 or IFF2 produce quasi-constant single-value outputs within a phenotype despite inter-individual variability. Results: Using bulk transcriptomes from 25 normal organs (GTEx) and 25 cancer types (TCGA), we computed IFF values for all normally distributed gene pairs, selected the 1000 most stable pairs per phenotype, and evaluated their performance in multiclass classification with five-fold cross-validation and independent hold-out testing. IFGs generated >70% accuracy across organs and cancers, establishing the existence of robust phenotype-specific symmetry “fingerprints.” Mapping approved and experimental drug targets onto networks construction from IFGs (IF-Nets) showed strong enrichment of hubs in cancer networks and highlighted the use of IF-Nets as drug discovery platforms in cancer treatment. Conclusion: Our findings demonstrate that gene-expression symmetry as a unifying organizing principle for phenotype definition and illustrate how IFGs and IF-Nets can guide biomarker design and symmetry-aware pharmacological intervention via “symmetry breaking.” Citation Format: Cristina Correia, Choong-Yong Ung, Cheng Zhang, Shizhen Zhu, Hu Li, . Learning-based invariant feature engineering reveals symmetry-encoded fingerprints of cancers that facilitate drug discovery [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 5516.

Human diseases are complex and arise from the coordinated action of multiple genes and their protein products. Genes’ behaviors extend beyond genetic variants, mutations, and differential expressions. Their coordinated activity across biological scales (molecules, cells, tissues, organs) produces emergent behaviors that shape health and disease. These emergent behaviors span time and space and are often hard to measure directly from observation when using standard experimental measurements. Yet these “hidden” or latent gene characteristics can be powerful drivers of disease. We propose a Mini-Galaxy Model (MGM), a systems-level AI-driven network framework that models cells as “mini-galaxies” composed of multilayered biological information, with each layer encoding a different dimension of genes’ behavior. Here, we delineate a strategy on how to construct and compare MGMs across health and disease and map their etiological relatedness. We also operationalize the MGM as a discovery platform for translational medicine, offering modules to allow target prioritization and editing. By reframing human diseases as the result of emergent behavior of multilayered multimode biological networks and their perturbations, the MGM yields actionable rules to streamline biomarker discovery, guide target selection and enable rational design of combinatorial interventions, and accelerate drug repurposing.

Abstract Neuroblastoma (NB) accounts for approximately 10% of pediatric cancer-related deaths. High-risk NB is often metastatic at diagnosis with a poor 5-year survival rate (<50%) despite intensive multimodal therapy. GD2, a disialoganglioside glycolipid expressed in nearly all NBs, is a well-validated immunotherapy target. Although GD2-CAR-T cells have been clinically evaluated, their efficacy remains limited due to poor tumor infiltration and the immunosuppressive tumor microenvironment (TME), underscoring the need for alternative immune-based approaches. Given the inherent capacity of macrophages to infiltrate solid tumors, phagocytose cancer cells, and modulate the TME, we engineered the GD2-CAR macrophage (GD2-CAR-M) as a novel therapeutic platform for NB. GD2-specific CAR constructs containing the Hu3F8 single-chain variable fragment (scFv) were designed, validated, and introduced into multiple human and murine monocyte/macrophage cell lines via lentiviral transduction. Fluorescence microscopy and flow cytometry demonstrated a significant enhancement in the phagocytosis of GD2-high NB cells and the phagocytic efficiency correlating positively with GD2 expression levels, suggesting antigen-specific phagocytosis by GD2-CAR-M. Our further in vivo studies using an immunocompetent MYCN-driven NB non-genetically engineered mouse model (MYCN-NGEMM) showed that treatment with GD2-CAR-M markedly reduced tumor burden and prolonged survival compared with controls. Notably, mice bearing GD2-high-tumors exhibited significantly improved survival, including one complete remission among ten treated animals. Histological analyses revealed increased connective tissue deposition and decreased tumor cell density, suggesting TME remodeling. Single-cell RNA sequencing and immunohistochemistry further demonstrated enhanced infiltration of macrophages, neutrophils, and T cells, indicating a shift from a “cold” to a “hot” immune TME following GD2-CAR-M treatment. Consistent antitumor effects were also observed in our dbh-MYCN transgenic zebrafish with spontaneous development of NB following GD2-CAR-M administration. Taken together, GD2-CAR-M exhibits potent antigen-specific antitumor activity and promotes favorable TME remodeling in NB models. These findings highlight GD2-CAR-M as a promising translational immunotherapy strategy for high-risk NB. Citation Format: Shizhen (Jane) Zhu, Ke-En Tan, Kok Siong Yeo, Cheng Zhang, Caleb T. Baker, Alexis M. Sosic, Eddie Lim . Yat Yuen, Choong Yong Ung, Stephanie F. Polites, Yi Lin, William A. Weiss, Hu Li. Harnessing GD2-CAR macrophages to target high-risk neuroblastoma [abstract]. In: Proceedings of the AACR Immuno-Oncology Conference (AACR IO): Discovery and Innovation in Cancer Immunology: Revolutionizing Treatment through Immunotherapy; 2026 Feb 18-21; Los Angeles, CA. Philadelphia (PA): AACR; Cancer Immunol Res 2026;14(2 Suppl):Abstract nr B005.

Cancer is a complex disease involving the deregulation of intricate cellular systems beyond genetic aberrations and, as such, requires sophisticated computational approaches and high-dimensional data for optimal interpretation. While conventional artificial intelligence (AI) models excel in many prediction tasks, they often lack interpretability and are blind to the scientific hypotheses generated by researchers to enable cancer discoveries. Here we propose that hypothesis-driven AI, a new emerging class of AI algorithm, is an innovative approach to uncovering the complex etiology of cancer from big omics data. This review exemplifies how hypothesis-driven AI is different from conventional AI by citing its application in various areas of oncology including tumor classification, patient stratification, cancer gene discovery, drug response prediction, and tumor spatial organization. Our aim is to stress the feasibility of incorporating domain knowledge and scientific hypotheses to craft the design of new AI algorithms. We showcase the power of hypothesis-driven AI in making novel cancer discoveries that can be overlooked by conventional AI methods. Since hypothesis-driven AI is still in its infancy, open questions such as how to better incorporate new knowledge and biological perspectives to ameliorate bias and improve interpretability in the design of AI algorithms still need to be addressed. In conclusion, hypothesis-driven AI holds great promise in the discovery of new mechanistic and functional insights that explain the complexity of cancer etiology and potentially chart a new roadmap to improve treatment regimens for individual patients.

AbstractBackgroundContext‐specific networks can capture functionally important genes that may not be differentially expressed. Leveraging seven different gene expression datasets of Alzheimer’s disease (AD) and healthy control brains, we constructed a novel pipeline that includes such network analysis, augmented by in vivo validations and in silico drug repurposing to facilitate prioritization of AD therapeutic targets.MethodWe utilized publicly available gene expression data from 733 AD and 454 control brains across 7 datasets and the NetDecoder (PMID: 26975659) network biology platform to build context‐specific networks. Based on the known protein‐protein interaction and gene expression correlations, information flow from the top differentially expressed genes (DEG) to transcriptional regulator genes was calculated using a process‐guided flow algorithm. Genes with the greatest informational flow changes across AD cohorts were systematically prioritized. We next validated these genes in a medium‐throughput screen of a tau‐transgenic Drosophila model (PMID: 11408621). Further, we identified existing small‐molecule drugs to target the candidate genes utilizing curated databases.ResultsA total of 14 context‐specific networks were constructed across 7 datasets. We found AD‐related genes, such as APP, to have high flow differences between AD and control networks and high connectivity. Notably, many such genes were not DEGs between AD and control, suggesting that the information flow captures novel aspects of AD pathophysiology. We nominated 224 unique genes with the greatest flow differences for functional validation. We screened 111 of the 224 candidate genes with readily available, high‐confidence orthologs in Drosophila. We identified genes that suppress (N = 68) or enhance (N = 27) tau‐related neurodegeneration in this model. We confirmed the effects of previously nominated genes, such as HDAC1, which enhances tau‐neurodegeneration if overexpressed and suppresses it if knocked down. We also identified and validated additional therapeutic targets. Lastly, we applied a network‐proximity‐based algorithm and identified small molecule drugs for the candidate genes.ConclusionWe present a novel pipeline that integrates analytic approaches, experimental validation, and drug repurposing to facilitate therapeutic discovery in AD. We applied this pipeline and nominated previously known and novel genes implicated in AD pathogenesis. Our pipeline demonstrates a cross‐species, systems‐level approach to therapeutic nomination in AD and related neurodegenerative diseases.

Abstract Despite intensive characterization of immune responses after COVID-19 infection and vaccination, research examining protective correlates of vertical transmission in pregnancy are limited. Herein, we profiled humoral and cellular characteristics in pregnant women infected or vaccinated at different trimesters and in their corresponding newborns. We noted a significant correlation between spike S1–specific IgG antibody and its RBD-ACE2 blocking activity (receptor-binding domain–human angiotensin-converting enzyme 2) in maternal and cord plasma (P < .001, R > 0.90). Blocking activity of spike S1–specific IgG was significantly higher in pregnant women infected during the third trimester than the first and second trimesters. Elevated levels of 28 cytokines/chemokines, mainly proinflammatory, were noted in maternal plasma with infection at delivery, while cord plasma with maternal infection 2 weeks before delivery exhibited the emergence of anti-inflammatory cytokines. Our data support vertical transmission of protective SARS-CoV-2–specific antibodies. This vertical antibody transmission and the presence of anti-inflammatory cytokines in cord blood may offset adverse outcomes of inflammation in exposed newborns.

Anticipating and understanding cancers’ need for specific gene activities is key for novel therapeutic development. Here we utilized DepMap, a cancer gene dependency screen, to demonstrate that machine learning combined with network biology can produce robust algorithms that both predict what genes a cancer is dependent on and what network features coordinate such gene dependencies. Using network topology and biological annotations, we constructed four groups of novel engineered machine learning features that produced high accuracies when predicting binary gene dependencies. We found that in all examined cancer types, F1 scores were greater than 0.90, and model accuracy remained robust under multiple hyperparameter tests. We then deconstructed these models to identify tumor type-specific coordinators of gene dependency and identified that in certain cancers, such as thyroid and kidney, tumors’ dependencies are highly predicted by gene connectivity. In contrast, other histologies relied on pathway-based features such as lung, where gene dependencies were highly predictive by associations with cell death pathway genes. In sum, we show that biologically informed network features can be a valuable and robust addition to predictive pharmacology models while simultaneously providing mechanistic insights.