Science Bulletin

Cardiovascular disease is one of the leading causes of death worldwide, with myocardial infarction (MI) being a major cause of both morbidity and mortality among cardiovascular patients. MI Patients face a higher risk of cardiovascular disease recurrence afterwards. Therefore, accurately predicting the risk of recurrence and identifying key risk factors are crucial for clinical decision-making. In this paper, we consider the interrelationships among cardiovascular factors from a systemic perspective. We first construct a differential network for each patient to capture individual-specific deviations in factor relationships and propose a novel method, termed Causal Factor-aware Graph Neural Network (CFGNN), which integrates factor interactions to predict the recurrence risk of MI patients while uncovering key risk factors from a causal perspective. Experimental results demonstrate that CFGNN performs well on hospital-derived datasets in real world, effectively identifying several key risk factors. This method not only deepens our understanding of cardiovascular disease, but also paves the way for more targeted and effective interventions.

Structural-functional coupling (SFC) provides critical insights into how the brains structural architecture constrains functional organization for supporting higher-order cognition. Investigating evolutionary differences in SFC between humans and macaques may provide new insights into the neural basis of unique human cognitive abilities. In this study, we analyzed multimodal magnetic resonance imaging data from anesthetized and awake adult rhesus macaques and adult humans to examine cross-species divergences in SFC. Moreover, we integrated transcriptomic data to elucidate the molecular mechanisms underlying evolutionary differences in SFC patterns. We find that humans and macaques exhibit distinct SFC patterns: the human brain shows high SFC in the lateral and medial prefrontal cortex, whereas macaques show high SFC in the sensorimotor cortex. Notably, language-related regions in the human lateral temporal cortex exhibit relatively low SFC. Furthermore, the human whole-brain SFC pattern and the evolutionary differences in SFC between humans and macaques are negatively correlated with cortical evolutionary expansion. By integrating human and macaque transcriptomes, we reveal that the macaque SFC specifically associated genes are primarily involved in basic physiological functions, whereas the human SFC specifically associated genes exhibit evolutionary adaptations in synaptic function, neurotransmitter secretion, and other molecular processes. Moreover, the human-specific genes showing significant overlap with Human Accelerated Regions genes were mainly enriched in cell types of astrocyte and oligodendrocyte and in diseases of schizophrenia and Alzheimers diseases. Overall, these findings advance our understanding of the intricate relationships of SFC in human and macaque brains and provide novel insights in understanding evolutionary conservation and species specificity in cognitive function and gene regulation.

Network analysis of human brain connectivity provides a fundamental framework for identifying the neurobiological mechanisms that cause cognitive variations and neurological disorders. However, existing diagnostic models often treat structural connectivity (SC) as a fixed or optimal topological scaffold for functional connectivity (FC). This consequently overlooks the higher-order dependencies between brain regions that are critical for characterizing pathological alterations. Moreover, the distinct spatial organizations of SC and FC complicate their direct integration, as naive alignment methods may distort the inherent nonlinear patterns of brain connectivity. To address these limitations, we propose the Graph Diffusion Optimal Transport Network (GDOT-Net), which models disease-related topological evolution and achieves precise alignment between SC and FC. Unlike existing diffusion studies, the proposed model introduces an evolvable brain connectome modeling approach to infer the complex topological structure of brain networks, unveiling higher-order connectivity patterns linked to specific neuropsychiatric disorders. Furthermore, GDOT-Net incorporates a Pattern-Specific Alignment mechanism, leveraging optimal transport to align structural and functional topological representations in a geometry-aware manner. To capture nonlinear topological relationships between brain regions, a Neural Graph Aggregator Module was developed, which adaptively learns complex node interaction patterns in brain networks. By leveraging this module, GDOT-Net generates highly discriminative representations that form a robust basis for the precision diagnosis of brain disorders. Experiments on REST-meta-MDD and ADNI demonstrate that GDOT-Net surpasses SOTA methods in uncovering structural-functional misalignments and disorder-specific subnetworks. The source code is publicly available at this Link.

Human coronavirus OC43 (HCoV-OC43) constitutes one of the most common causes of the seasonal cold but can also cause severe disease among elderly and immuno-compromised. Currently, there are no approved antiviral drugs to combat HCoV-OC43 infection. Coronaviruses are positive single-stranded RNA (+ssRNA) viruses and utilize -1 programmed ribosomal frameshifting (-1 PRF) to obtain the correct stoichiometry of viral protein components. As such, the ribosomal frameshifting stimulation element (FSE) is a promising target for antiviral drug discovery, due to its high conservation. By repurposing available drugs, we identified a group of anthracycline compounds that can reduce -1 PRF of HCoV-OC43 and reduce viral infection of cells. Furthermore, we show that anthracyclines that reduce infection also bind the FSE and reduce frameshift frequency. We also show that the selected anthracyclines reduce SARS-CoV-2 infection, but without affecting -1 PRF frequency. All together, we demonstrate that a subset of anthracyclines selectively binds and inhibit the HCoV-OC43 FSE and could thus serve as a robust framework when developing new antivirals targeting coronaviruses.

The recurrent outbreaks and geographical expansion of mosquito-borne arboviruses pose a significant challenge to public health worldwide. The disease outcome for arboviral infections ranges from acute febrile illness to severe conditions such as encephalitis, hemorrhagic shock, and mortality. Current treatment options for these viruses are limited to supportive care, necessitating an urgent need for a safe and effective broad-spectrum antiviral. In this study, we have identified Trifluoperazine (TFP), an FDA-approved antipsychotic, as a potent broad-spectrum antiviral against Japanese encephalitis Virus (JEV), Dengue virus (DENV) and Chikungunya virus (CHIKV) infections. The antiviral effect of TFP was also seen in the animal models of JEV and CHIKV with significantly reduced disease severity. Mechanistically, TFP treatment increased the phosphorylation of eIF2a and induced an adaptive ER stress response in diverse cell types. Alleviation of TFP-induced ER stress by chemical chaperone 4PBA abolished the antiviral activity of the drug and rescued virus replication in cells. The robust in vitro and in vivo efficacy of the drug against arboviruses highlights the potential for repurposing TFP as a broad-spectrum antiviral candidate.

Despite of the highly potent antiretroviral therapies, HIV-1 establishes persistent infection and causes chronic inflammation in AIDS patients. Beyond CD4+ T cells, HIV-1 infects myeloid cells, including circulating monocytes and tissue-resident macrophages, and integrates with host genomes to form stable viral reservoirs. To achieve a functional HIV cure, latency-promoting agents (LPAs) have been developed for the "block-and-lock" strategy to reinforce deep HIV-1 latency and permanently silence proviruses. However, most LPAs have been tested mainly in CD4+ T cells, and their efficacy in myeloid cells remains unclear. In this study, we reported that levosimendan (LSM), a drug approved for clinic use to treat heart failures, is able to inhibit HIV lytic infection and reactivation in myeloid cells. LSM blocked viral lytic reactivation in HIV-1 latently infected monocytic cells (TH89GFP, U1) and microglial cells (HC69). LSM also inhibited HIV infection in human induced pluripotent stem cell (iPSC) derived microglia (iMG), primary human resident liver macrophages (Kupffer cells) as well as human monocyte-derived macrophages (MDMs). Furthermore, we demonstrated that overexpression of a predicted drug target of LSM, the conserved serine/threonine kinase RIOK1 (RIO kinase 1), overcomes LSMs anti-HIV effect. Overall, our studies concluded that LSM is a promising LPA to inhibit HIV-1 infection in myeloid cells in the RIOK1-dependent manner.

Coherent Ising machines (CIMs) excel at solving large-scale combinational optimization problems (COPs), but their insufficient long-term stability has hindered their applications in compute-intensive tasks like computer-aided drug discovery (CADD). By improving fiber vibration isolation and temperature control system, we have implemented a 2000-node CIM named QBoson-CPQC-3Gen achieving stable solutions over one hour on large-scale COPs. Graph-based encoding schemes were further introduced to realize a CIM-based CADD workflow including allosteric site detection, protein-peptide docking and intermolecular similarity calculation. CIM-based methods demonstrated superior speed and accuracy than heuristic algorithms. Especially, QBoson-CPQC-3Gen identified 2 novel druggable sites and bioactive compounds for 6 targets, which were further validated in vitro, in-cell and by crystal structures. Our contributions established a quantum-computing framework for multi-stage drug discovery, representing a significant advancement in both quantum computing applications and pharmaceutical research.

Astrocytes play a key role in neuronal homeostasis and in various neural disorders. The generation of astrocytes from neural progenitor cells (NPCs) and its functions are under a complex control of several signaling networks and transcription factors. In this study, we demonstrate that the transcription factor, GLIS similar 3 (GLIS3), which has been implicated in several neurodegenerative diseases, is highly expressed in astrocytes, and is required for the efficient differentiation of human NPCs into astrocytes. Loss of GLIS3 function greatly impairs astrocytes differentiation, resulting in reduced expression of astrocyte markers, whereas expression of exogenous GLIS3 restores the induction of astrocyte specific genes indicating a critical role for GLIS3 in astrocyte differentiation. Integrated transcriptomic and cistromic analyses revealed that GLIS3 directly regulates the transcription of several astrocyte-associated genes, including GFAP, SLC1A2, NFIA, and ATF3, in coordination with lineage-determining factors, such as STAT3, NFIA, and SOX9. We hypothesize that GLIS3 dysfunction disrupts this transcriptional network thereby contributing to astrocyte-associated neurological disorders. Identification of GLIS3 as a key regulator of astrocyte differentiation and gene expression will advance our understanding of its role in neurodegenerative diseases and may provide a new therapeutic target.

Chronic kidney disease (CKD) affects nearly 850 million individuals globally; the prevalence of undiagnosed CKD is 60%. Taking advantage of the relationship between CKD and cardiovascular disease, we developed a deep learning (DL) model to detect CKD from parasternal long-axis (PLAX) videos using 325,377 PLAX videos from 62,818 patients at Cedars-Sinai Medical Center (CSMC). We externally validated our model in two independent cohorts of 2,224 patients at Stanford Healthcare (SHC) and 41,611 patients at Kaiser-Permanente Northern California (KPNC). In a held-out test cohort at CSMC, our model detected any stage of CKD with an area under the curve (AUC) of 0.756 [95% confidence interval 0.749 - 0.763], with consistently strong performance in KPNC (AUC 0.718 [0.714 - 0.723]) and SHC (AUC 0.719 [0.704 - 0.735]). Our DL echo model detected CKD with robust performance at two external clinical sites, offering an avenue for noninvasive screening and improved detection rates.

Emotion and mood regulation critically depends on interactions between the anterior cingulate cortex (ACC) and the amygdala. However, the detailed architecture of ACC projections to their major targets, the basal (BA) and accessory (AcBA) basal nuclei of the amygdala, remains unclear. To address this issue, a combined retrograde and anterograde tracing with viral vectors were performed in macaques to map the projection patterns from pregenual (pgACC), subgenual (sgACC), and dorsal (dACC) subareas. Data revealed that ACC neurons projecting to the BA arose predominantly from the superficial layers (II/III) of all subareas and the deep layers (V/VI) of the sgACC, whereas ACC neurons projecting to the AcBA originated mainly in the deep layers of the sgACC and dACC. The present study defines the topographic and layer-specific organization of ACC-amygdala connectivity in primates and subserves to provide an anatomical basis for future causal and translational approaches, such as targeted interventions against ACC-related mood disorders. TeaserPrimate anterior cingulate cortex has topographic and layer-specific projections to amygdala that are involved in emotion and mood regulation.

7q11.23 Duplication Syndrome (7Dup) is a type of syndromic autism spectrum disorder caused by duplication of a typically 1.5-1.8 Mb segment in section q11.23 of chromosome 7, including 25-27 genes. Previous work has highlighted the GTF2I gene as playing a major role in the phenotype of 7Dup patients. Accordingly, mice with Gtf2i duplication (Gtf2i+/dup) are commonly used as an animal model of 7Dup. We previously reported deficits in several behavioral and physiological modalities, which were associated with Gtf2i dosage in mice conducting a battery of social discrimination tests. Here, we report the effect of treating Gtf2i+/dup mice with Baicalin, a naturally occurring flavonoid added to the mices drinking water (0.15 mg/ml), on these deficits. We found that Baicalin treatment ameliorated the higher surface temperature observed in Gtf2i+/dup males and the lower tail temperature observed in Gtf2i+/dup females during the social behavior tests. It also prevented the increased defecation rate exhibited by Gtf2i+/dup mice during the social preference test. We further analyzed the effect of Baicalin treatment on cortical neurons differentiated from 7Dup patient-derived IPSCs. Using whole cell patch clamp and calcium imaging, we found an increased rate of excitatory postsynaptic currents in Baicalin-treated cells, without a change in their firing rate, indicating a stronger synaptic activity in the Baicalin-treated cells. Altogether, our results reveal that Baicalin administration alleviates some of the behavioral and physiological effects of Gtf2i duplication in mice, and affects neuronal activity in cultured 7Dup human neurons. Thus, Baicalin administration has the potential to serve as a treatment for 7Dup patients.

Neuromodulators influence critical functions of the developing human brain and regulate behavioral states. Dysfunction of neuromodulatory systems is often involved in neuropsychiatric disease and many drugs for these conditions act on these signaling pathways. Recent advances in stem cell biology have made it possible to derive a wide range of cells across the developing human nervous system in regionalized organoids and to functionally integrate them into assembloids, however they currently do not systematically incorporate neuromodulation. Here, we generated human midbrain-hindbrain organoids (hMHO) from human induced pluripotent stem (hiPS) cells and fused them with human cortical organoids (hCO) to form neuromodulatory assembloids (hNMA). We focus on serotonin (5-hydroxytryptamine, 5-HT) as one key neuromodulator and found characteristic gene expression patterns and electrophysiological properties of serotonergic neurons (5-HT neurons) in the hMHO. In hNMA, 5-HT neurons projected into hCO, released 5-HT and modulated cortical network activity. To explore the applicability of this system in human disease, we studied 22q11.2 deletion syndrome (22q11.2DS), a common microdeletion associated with high risk for neuropsychiatric disease and defects in 5-HT signaling. We found aberrant 5-HT dynamics in hNMA from patient hiPS cell lines that were rescued by administration of a selective serotonin reuptake inhibitor (SSRI). Taken together, hNMA can be used to study human 5-HT dynamics and uncover disease phenotypes which could facilitate therapeutic development.

African swine fever (ASF) is a highly pathogenic disease caused by the African swine fever virus (ASFV) infection, which can affect pigs of all ages and breeds, posing significant threat to the global pig farming industry. The ASFV p30 protein is an early-expressed viral structural protein; however, its function is not fully understood. In this study, the interaction of viral p30 with host TRIM21 was identified. The ectopic TRIM21 inhibited ASFV replication, while knockdown or knockout of TRIM21 promoted ASFV replication. Further, p30 was found to interact with RIG-I-like receptor (RLR) signaling adaptor MAVS, and during ASFV infection, p30-TRIM21-MAVS interacted with each other. Mechanistically, TRIM21 activated the K27 polyubiquitination of MAVS to induce IRF3 mediated type I interferon (IFN) production, whereas p30 counteracted TRIM21 activated MAVS K27 polyubiquitination to evade RLR signaling mediated antiviral IFN induction. In summary, our study revealed a novel function of ASFV p30, and provided new insights into the immune evasion of ASFV.

Missense variants are a common type of genetic mutation that can alter the structure and function of proteins, thereby affecting the normal physiological processes of organisms. Accurately distinguishing damaging missense variants from benign ones is of great significance for clinical genetic diagnosis, treatment strategy development, and protein engineering. Here, we propose the VarDCL method, which ingeniously integrates multimodal protein language model embeddings and self-distilled contrastive learning to identify subtle sequence and structural differences before and after protein mutations, thereby accurately predicting pathogenic missense variants. First, leveraging sequence and structural information before and after mutations, VarDCL generates sequence-structural multimodal features via different language models. It incorporates both global and local perspectives of feature embeddings to provide the model with dynamic, multimodal, and multi-view input data. Additionally, a Self-distilled Contrastive Learning (SDCL) module was proposed to enable more effective information integration and feature learning, enhancing the models ability to detect sequence and structural changes induced by mutations. Within this module, the multi-level contrastive learning framework excels at capturing information differences before and after mutations within the same modality; meanwhile, the feature self-distillation mechanism effectively utilizes high-level fused features to guide the learning of low-level differential features, facilitating information interaction across different modalities. The VarDCL framework not only ensures the models capacity to learn dynamic changes pre- and post-mutation but also significantly improves cross-modal information interaction between sequence and structure, thereby remarkably boosting the models performance in distinguishing pathogenic mutations from benign ones. To validate the effectiveness of VarDCL, extensive experiments were conducted. The ablation study demonstrates that all key components of VarDCL contribute significantly. On an independent test set containing 18,731 clinical variants, VarDCL achieved an AUC of 0.917, an AUPR of 0.876, an MCC of 0.690, and an F1-score of 0.789, outperforming 21 state-of-the-art existing methods. Benchmark analysis shows that VarDCL can be utilized as an accurate and potent tool for predicting missense variant effects.

The impact of cellular activities on large-scale brain dynamics is thought to determine brain functioning and disease, yet the causal relationships of neural mechanisms across scales remain unclear. Recently, the cerebellum has been reported to affect whole-brain dynamics during sensorimotor integration. To disclose the underlying mechanisms, we have developed a multiscale digital brain co-simulator, in which a spiking neural network of the olivo-cerebellar microcircuit is embedded in a mouse virtual brain and wired with other nodes using an atlas-based long-range connectome. Parameters and bi-directional interfaces between the spiking olivo-cerebellar network and other rate-coded modules were tuned to match experimental data of primary sensory and motor cortex (M1 and S1) power spectral densities and neuronal spiking rates. Then, the role of the cerebellar circuitry on sensorimotor integration was analyzed by lesioning critical circuit connections in silico. Simulations showed that spike processing within the cerebellar circuit is key to explaining the gamma-band coherence between M1 and S1 during sensorimotor integration. These results provide a mechanistic explanation of how the cerebellum promotes the formation of sensorimotor contingencies in relevant cortical modules as the basis of its critical role in sensorimotor prediction. On a broader perspective, this modelling approach opens new perspectives for the multiscale investigation of brain physiological and pathological states in relation to specific cellular and microcircuit properties.

Targeted drug therapies offer a promising approach for treating complex diseases, with combinational drug therapies often employed to enhance therapeutic efficacy. However, unintended drug-drug interactions may undermine treatment outcomes or cause adverse side effects. In this work, we propose a novel joint learning framework for the simultaneous prediction of effective drug combinations and drug-drug interactions, based on coupled tensor-tensor factorization. Specifically, we model drug combination therapies and DDI by representing drug-drug-disease associations and drug-drug interaction profiles as coupled three-way tensors. To address the challenges of data incompleteness and sparsity, the proposed model integrates auxiliary drug similarity information, such as chemical structure similarities, drug-specific side effects, drug target profiles, and drug inhibition data on cancer cell lines, within a multi-view learning frame-work. For optimization, we adopt a modified Alternating Direction Method of Multipliers (ADMM) algorithm that ensures convergence while enforcing non-negativity constraints. In addition to standard tensor completion tasks, we further evaluate the proposed method under a more realistic new-drug prediction setting, where all interactions involving a previously unseen drug are withheld. This scenario closely aligns with real-world applications, in which reliable predictions for emerging or under-studied compounds are essential. We evaluate the proposed method on a comprehensive dataset compiled from multiple sources, including DrugBank, CDCDB, SIDER, and PubChem. Our experiments show that SI-ADMM maintains robust performance and achieves the best results comparing to other tensor factorization approaches, with or without auxiliary information, particularly in the new-drug prediction setting. The implementation of our method is publicly available at: https://github.com/Xiaoge-Zhang/SI-ADMM.

Hepatitis B and D virus (HBV, HDV) enter hepatocytes through a coordinated process mediated by a receptor complex consisting of sodium taurocholate co transporting polypeptide (NTCP) and its entry cofactors, including epidermal growth factor receptor (EGFR). Here, we established an in vitro assay to evaluate the NTCP-EGFR interaction and identified Oxy229, an oxysterol-based compound that disrupted this molecular interaction. Oxy229 selectively inhibited HBV and HDV infection to HepG2-NTCP cells and primary human hepatocytes. Mechanistic analysis revealed that Oxy229 impaired the relocalization of the HBV-NTCP complex from plasma membrane to intracellular vesicles. Notably, Oxy229 did not compromise the physiological functions of NTCP and EGFR, i.e., bile acid transport and activation of downstream EGFR signaling pathways including Ras-MAPK and PI3K-Akt pathways, indicating selective inhibition of viral entry. Compound derivative analysis identified Oxy283, which acquired dual inhibitory activity against both NTCP-EGFR interaction and NTCP multimerization, resulting in enhanced anti-HBV potency. These findings establish the functional significance of the NTCP-receptor complex formation in HBV/HDV entry and highlight this machinery as a potential target for antiviral intervention.

The escalating crisis of antimicrobial resistance poses a devastating and immediate threat to human life. Antimicrobial peptides (AMPs) are a promising antibiotic substitute to combat antimicrobial resistance. Compared with the traditional wet-lab screening approaches, computational models have largely improved the efficiency of predicting antimicrobial peptide. However, most computational models overlook or underutilize the evolution and structural information of peptides, which is crucial for understanding the peptide functions. Here, we proposed a sophisticated deep learning model to predict AMPs, Antimicrobial Peptide Bilinear Attention Network (AMPBAN), which incorporates peptide evolution features from ESM3 protein language model, structure features from ESMFold predicted with equivariant graph neural network (EGNN), and the joint information from sequence and structure learned via Bilinear Attention Network. AMPBAN consistently demonstrated superior accuracy and generalization compared to nine state-of-the-art AMP prediction models across multiple independent benchmarks. Furthermore, an ablation study confirms that our multimodal fusion strategy significantly refines the integration of sequence and structural signals, yielding superior predictive balance over single-modality models. This framework provides a robust tool for the accelerated discovery of novel AMPs and the advancement of next-generation antimicrobial drug development. The datasets, source code and models are available at https://github.com/baiwenhuim/ampban.

The escalating crisis of antimicrobial resistance necessitates novel therapeutic strategies, among which drug combination therapy shows great promise by enhancing efficacy and reducing toxicity. However, identifying effective synergistic pairs from the vast combinatorial space remains experimentally challenging and resource-intensive. To address this, we introduce GCN-Mamba, a deep learning framework that integrates Graph Convolutional Networks (GCN) with the Mamba State Space Model. This architecture captures both local molecular topological structures and global implicit interactions by leveraging Extended 3-Dimensional Fingerprints (E3FP) and bacterial gene expression profiles. Evaluation on a comprehensive dataset demonstrated that GCN-Mamba significantly outperforms classical machine learning models in predictive accuracy. In a targeted case study against Methicillin-resistant Staphylococcus aureus (MRSA), the model successfully rediscovered known synergistic pairs, such as Quercetin and Curcumin, consistent with recent literature. Furthermore, prospective in vitro validation confirmed a novel synergistic combination of Shikimic acid and Oxacillin, validating the models practical utility. By efficiently prioritizing potential candidates, GCN-Mamba serves as a powerful and reliable tool for accelerating the discovery of synergistic antimicrobial combinations, effectively bridging the gap between computational prediction and experimental validation.

Neonates requiring intensive care are at increased risk for long-term neuropsychiatric disorders. However, clinical adoption of risk prediction models remains limited when their performance lacks adequate interpretability for informed clinical decision-making. Here, we investigated whether longitudinal neonatal electronic health record (EHR) data from the first 90 days of life can support clinically meaningful interpretation of long-term risk signals for major neuropsychiatric diagnoses by age seven. In a retrospective register-based cohort of 17,655 at-risk children from an academic medical center, of whom 8.0\% (1,420) received a major neuropsychiatric diagnosis during follow-up, we applied a time-aware transformer model (Self-supervised Transformer for Time-Series; STraTS) and thoroughly evaluated its predictions using three complementary interpretability approaches: perturbation-based variable importance, value-dependent effect analysis, and leave-one-out (LOO) feature attribution. STraTS achieved the highest area under the precision--recall curve (AUPRC 0.171 {+/-} 0.022), compared with Random Forest (0.166 {+/-} 0.008), logistic regression (0.151 {+/-} 0.007), and XGBoost (0.128 {+/-} 0.010). Across interpretability methods, five predictors were consistently identified: birth weight, gender, Apgar score at 1 minute, umbilical serum thyroid stimulating hormone (uS-TSH), and treatment time in hospital. Indicators of early clinical severity, including chromosomal abnormalities and neonatal cerebral-status disturbances, showed the largest risk-increasing effects. Furthermore, the model's learned vector representations of subject-specific EHR sequences formed clinically coherent latent embeddings that reflect population heterogeneity along established perinatal risk dimensions. These findings demonstrate that combining multiple complementary interpretability methods yields stable, clinically plausible risk signals while revealing limitations that would remain undetected by any single approach, highlighting the importance of careful interpretability analysis of deep learning-based risk predictions.