Energy-ensemble concept bottleneck models for enhancing interpretability and accuracy in concept-based learning

Abstract Tumor heterogeneity fueled by plasticity and genetic diversification of cancer cells is key to therapy failure of malignant glioma. We implemented spatial and genetic platforms at single cell resolution to explore the trajectory of evolution of glioblastoma. Using spatial analysis of whole glioblastoma sections, we established a homotypic clustered cell identity paradigm whereby tumor cell state coherence was maximal in cells organized in homotypic clusters, whereas dispersed cells downregulated the original state, acquired alternative phenotypes and exhibited changes in the microenvironment, thus linking the process of plasticity to loss of the cell adhesion mechanisms that preserve the clustered spatial pattern of glioblastoma cells. We also used single cell DNA-sequencing methods integrated with the single cell transcriptome of patient-matched primary-recurrent glioblastoma pairs to resolve the clonal substructure of untreated glioblastoma and determine the clonal evolution at recurrence driven by therapeutic resistance. The evolutionary trajectory of glioblastoma identified a bottleneck model as the predominant pattern of evolution and converged on the identification of a rare persister subclonal state in primary glioblastoma exhibiting distinct phenotypic hallmarks that evolves and diversifies to populate the recurrent tumor mass. We used preclinical tumor models to trace the individual lineages associated with the persister subclones and experimentally illuminated the biological and metabolic activities of the persister cellular state in brain tumors. Persister glioblastoma cells in untreated tumors lacked spatial segregation and were independent predictors of timing to recurrence for glioblastoma patients. Thus, genetic and non-genetic co-evolution mechanisms forge the acquisition of plasticity and therapy resistance in glioblastoma. Citation Format: Antonio Iavarone. Shaping tumor cell plasticity and therapy resistance in glioblastoma [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Brain Cancer; 2026 Mar 23-25; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(6_Suppl):Abstract nr IA005.

Concept Bottleneck Models facilitate interpretable image classification by predicting human-understandable concepts prior to class labels. However, when constructed upon CLIP, they exhibit unreliable concept scores stemming from CLIP's global representation bias and insufficient region-level sensitivity, which severely constrain their effectiveness in sensor-driven applications like remote sensing and medical imaging where localized visual evidence is critical. To mitigate this, we propose the Local-Global Aware Concept Bottleneck Model (LGA-CBM), which improves concept prediction through a training-free refinement pipeline. Building on initial CLIP-derived concept scores, LGA-CBM incorporates three key components: a Dual Masking Guided Concept Score Refinement (DMCSR) module that exploits attention weights to strengthen region-concept alignment; a Local-to-Global Concept Reidentification (L2GCR) strategy to harmonize local and global activations; and a Similar Concepts Correction Mechanism (SCCM) integrating Grounding DINO for fine-grained disambiguation. A sparse linear layer then maps the refined concepts to class labels, enabling highly interpretable classification with minimal concept usage. Experiments across six benchmark datasets demonstrate that LGA-CBM consistently achieves state-of-the-art performance in both accuracy and interpretability, producing explanations that align closely with human cognition.

Distilling knowledge from large language models: A concept bottleneck model for hate and counter speech recognition

With the increase in cross-border mobility, passports, as critical identity documents, require robust anti-counterfeiting security. While existing deep learning-based automatic detection methods achieve high accuracy, they lack interpretability. This paper introduces the Concept Bottleneck Model (CBM) to construct a transparent passport printing method detection framework. By defining interpretable intermediate concepts and integrating linear reasoning, the model significantly enhances reliability and debugging efficiency. The article systematically analyzes the advantages, challenges, and future directions of this approach.

Underwater optical images are the primary carriers of underwater scene information, playing a crucial role in marine resource exploration, underwater environmental monitoring, and engineering inspection. However, wavelength-dependent absorption and scattering severely deteriorate underwater images, leading to reduced contrast, chromatic distortions, and loss of structural details. To address these issues, we propose a U-shaped underwater image enhancement framework that integrates Swin-Transformer blocks with lightweight attention and residual modules. A Dual-Window Multi-Head Self-Attention (DWMSA) in the bottleneck models long-range context while preserving fine local structure. A Global-Aware Attention Map (GAMP) adaptively re-weights channels and spatial locations to focus on severely degraded regions. A Feature-Augmentation Residual Network (FARN) stabilizes deep training and emphasizes texture and color fidelity. Trained with a combination of Charbonnier, perceptual, and edge losses, our method achieves state-of-the-art results in PSNR and SSIM, the lowest LPIPS, and improvements in UIQM and UCIQE on the UFO-120 and EUVP datasets, with average metrics of PSNR 29.5 dB, SSIM 0.94, LPIPS 0.17, UIQM 3.62, and UCIQE 0.59. Qualitative results show reduced color cast, restored contrast, and sharper details. Code, weights, and evaluation scripts will be released to support reproducibility.

Managing commuting flow from residential communities to subway stations is particularly challenging due to travel time uncertainty, which disrupts both commuter travel patterns and station inbound behavior. This study proposes a reservation service for the community-station commuting scenario, enabling the coexistence of reserved and unreserved channels. The service is expected to provide more stable travel and efficient entry via dedicated reservation channels. A bottleneck model is developed to incorporate travel time uncertainty into the community-station process, analyzing its influence on demand split, departure time, and queuing time. From both commuter and operator perspectives, the implementation boundary of reservation services is explored, and an upper bound for the unit reservation cost is derived. Given specific parameter settings, the feasible travel time range, demand split, and demand-to-capacity ratio for each channel can be analytically determined, providing decision support for operators considering the deployment of reservation services.

BackgroundHistopathological grading is a key prognostic marker for hepatocellular carcinoma (HCC). However, the clinical application of deep learning models (DLMs) for predicting HCC grading from medical imaging is limited by their black-box nature. We aimed to develop an interpretable DLM, interpretable HCC grading network (iHCG-Net), to predict HCC grading preoperatively using multi-phase contrast-enhanced magnetic resonance imaging (CEMRI).MethodsThis study retrospectively enrolled 370 HCC patients who underwent preoperative CEMRI before curative resection. Based on postoperative pathology, the patients were categorized into high-grade (n=136) and low-grade (n=234) HCC groups. They were then stratified into a training cohort (n=259) and a time-independent validation cohort (n=111). Twenty-three clinical-radiological features were collected for all patients. The iHCG-Net, based on the Concept Bottleneck Model (CBM) framework, first encodes CEMRI images using a DenseNet-121 backbone and then leverages a concept regressor to predict the twenty-three clinical-radiological features for final prediction of HCC histological grade. A feature importance score plot was generated to assess the contribution of each feature to the differential diagnosis. Nine baseline predictive models were developed for comparison. The models were evaluated using receiver operating characteristic (ROC) curve analysis and DeLong’s test.ResultsiHCG-Net demonstrated strong predictive performance for HCC grading, achieving areas under the receiver operating characteristic curve (AUCs) of 0.893 in the training cohort and 0.802 in the validation cohort. The model significantly outperformed conventional models, including the clinical-radiological model (CM), radiomics models (RMs), and a clinical-radiomic combined model (CRM) (AUCs: 0.675–0.778, 0.617–0.723; P<0.05). Furthermore, iHCG-Net exhibited performance comparable to that of the DLM (AUCs: 0.920, 0.774; P>0.05), while providing inherent interpretability and mitigating the risk of overfitting. Feature importance analysis identified intratumoral arteries as the most influential feature for predicting HCC grading, with an importance score of 0.213.ConclusionsThe iHCG-Net can be a promising interpretable artificial intelligence tool for the preoperative prediction of HCC grading.

Interpretable prognostics with concept bottleneck models

Background. Takayasu arteritis (TAK) is a large-vessel vasculitis with fluctuating disease activity and a lack of validated biomarkers for relapse prediction. While [18F]FDG PET provides diagnostic sensitivity, its prognostic role remains debated. This study aimed to assess whether convolutional neural networks (CNNs) applied to baseline [18F]FDG PET scans can predict clinical relapses within 12 months in patients with TAK. Methods. We retrospectively analysed [18F]FDG PET scans acquired from patients with TAK between 2013 and 2023 at a tertiary centre. All scans were harmonised using a liver-based standardisation method to correct inter-scanner variability. Clinical relapse was defined as recurrence of TAK-related signs or symptoms, based on blinded chart review. CNNs were trained using four architectures: Baseline model (4 convolutional layers), Residual bottleneck model (3 blocks), and two models retrieved from the MONAI library – ResNet18 and DenseNet121. The dataset was split into training, validation, and test sets. To address class imbalance, we applied class weighting, early stopping, and 10-fold subsampling of the majority class. Data augmentation techniques included RandAffine and RandAxisFlip. Performance was evaluated using accuracy, class-specific precision and recall, specificity, and normalised Matthews correlation coefficient (nMCC). Results. A total of 306 scans from 74 patients with TAK (median age: 47 years [range: 14–75]; 62 females) were included. Sixty-four scans (21%) were followed by clinical relapse within 12 months. Of these, 25 were correctly identified by blinded nuclear medicine readers. The Baseline model-1 achieved the best overall performance with 82% accuracy, 85% precision for class 0, 62% precision for class 1, 38% recall, 94% specificity, and 70% nMCC. DenseNet121-2 yielded 81% accuracy, 100% precision for class 1 (relapse), 8% recall, 100% specificity, and 62% nMCC, maintaining 100% precision and specificity in the subsampled test set. Performance metrics for all models are summarised in Figure 1. Compared to standard visual PET interpretation, which correctly identified fewer than 40% of relapses, CNN-based models significantly improved predictive performance (Figure 2). Conclusions. This is the first study applying CNN-based [18F]FDG PET analysis to predict 12-month clinical relapse in patients with TAK, involving one of the largest PET datasets in this disease. Despite the retrospective design and class imbalance, CNNs demonstrated strong performance in identifying those likely to remain in remission. These findings support the use of imaging-based computational tools to inform treatment tapering and follow-up strategies in patients with TAK.

Practicing two simultaneous tasks in an extensive manner reduces the performance impairments (i.e., dual-task costs) that occur in dual-task situations compared to single-task situations. The present study provides empirical tests of the latent bottleneck model to explain this reduction and thus the practice-related improvement in dual-task performance. To do so, in three experiments, participants practiced a visual-manual and an auditory-verbal task in single-task and dual-task trials for several sessions. In these experiments, we changed the duration of the response selection stages of the two tasks after practice and analyzed the resulting effects on the reaction times (RTs) during subsequent transfer. The results showed a pattern of selective prolongations of the RTs in the two tasks, which depends on the location of the manipulated process relative to a presumed latent processing bottleneck. The manipulation of the time at bottleneck stages in the longer (auditory-verbal) task did not propagate into the RTs of the shorter task, while prolongations of bottleneck stages of a shorter (visual-manual) task propagated into longer task RTs after practice. These results are consistent with a latent bottleneck model of dual-task practice.

The main challenges limiting the adoption of deep learning-based solutions in medical workflows are the availability of annotated data and the lack of interpretability of such systems. Concept Bottleneck Models (CBMs) tackle the latter by constraining the model output on a set of predefined and human-interpretable concepts. However, the increased interpretability achieved through these concept-based explanations implies a higher annotation burden. Moreover, if a new concept needs to be added, the whole system needs to be retrained. Inspired by the remarkable performance shown by Large Vision-Language Models (LVLMs) in few-shot settings, we propose a simple, yet effective, methodology, CBVLM, which tackles both of the aforementioned challenges. First, for each concept, we prompt the LVLM to answer if the concept is present in the input image. Then, we ask the LVLM to classify the image based on the previous concept predictions. Moreover, in both stages, we incorporate a retrieval module responsible for selecting the best examples for in-context learning. By grounding the final diagnosis on the predicted concepts, we ensure explainability, and by leveraging the few-shot capabilities of LVLMs, we drastically lower the annotation cost. We validate our approach with extensive experiments across four medical datasets and twelve LVLMs (both generic and medical) and show that CBVLM consistently outperforms CBMs and task-specific supervised methods without requiring any training and using just a few annotated examples. More information on our project page: https://cristianopatricio.github.io/CBVLM/.

An inland waterway navigation congestion mitigation strategy based on a multi-objective moment bottleneck model

Overtreatment of patients with ductal carcinoma in situ (DCIS) is driven by a lack of a reliable prognostic biomarker. Widespread mammographic screening has resulted in a substantial increase in women diagnosed with DCIS. To improve patient risk stratification, we investigate circulating tumor cells (CTCs) as a biomarker for DCIS patients’ biological aggressiveness and as an indicator of early dissemination. We apply microfluidics to enrich CTCs from 34 patients with DCIS and find a significantly higher concentration of CTCs compared to in healthy controls. We profile CTCs and matched DCIS tissues using single-cell RNA sequencing. We find that CTCs express higher clonal aberrations when compared to white blood cells from the same samples, and clonal comparisons between matched tissue and CTC samples provide evidence for an evolutionary bottleneck model. mRNA expression in CTC reveals EMT/MET and immunoregulatory pathway regulation with a suggestion of racial differences. Last, we provide support for early dissemination in DCIS using a Mouse IntraDuctal model.

Evaluating population genetic inference methods is challenging due to the complexity of evolutionary histories, potential model misspecification, and unconscious biases in self-assessment. The Genomic History Inference Strategies Tournament (GHIST) is a community-driven competition designed to evaluate methods for inferring evolutionary history from population genomic data. The inaugural Genomic History Inference Strategies Tournament competition ran from July to November 2024 and featured four demographic history inference challenges of varying complexity: a bottleneck model, a split with isolation model, a secondary contact model with demographic complexity, and an archaic admixture model. Data were provided as error-free VCF files, and participants submitted numerical parameter estimates that were scored by relative root-mean-squared error. Approximately 60 participants competed, using diverse approaches. Results revealed the current dominance of methods based on site frequency spectra, while highlighting the advantages of flexible model-building approaches for complex demographic histories. We discuss insights regarding the competition and outline the next iteration, which is ongoing with expanded challenge diversity. By providing standardized benchmarks and highlighting areas for improvement, Genomic History Inference Strategies Tournament represents a substantial step toward more reliable inference of evolutionary history from genomic data.

This study examines departure time decisions in Vickrey's bottleneck model by incorporating bounded rationality through a quantal response equilibrium (QRE) framework. Traditional bottleneck models assume perfect rationality and may fail to capture real-world commuter behaviour. To address this limitation, a discrete version of the bottleneck model was developed and experimentally tested to compare the predictive performance of QRE and symmetric mixed-strategy equilibrium (SMSE). In a controlled laboratory setting with two conditions differing in the shadow price of travel time, departure time choices initially deviated from SMSE predictions, particularly under a low shadow price, but gradually converged toward them over repeated rounds. QRE provided a closer fit to observed departure time distributions than SMSE, especially in the low shadow price condition. Based on two out-of-sample validation measures, QRE demonstrated superior predictive performance, highligting the value of incorporating bounded rationality into bottleneck congestion models.

Are closed-form solutions attainable after incorporating Logistic scheduling preference into the bottleneck model?

Departure time choices and comparisons of household commuting based on the stochastic bottleneck model

Evaluating population genetic inference methods is challenging due to the complexity of evolutionary histories, potential model misspecification, and unconscious biases in self-assessment. The Genomic History Inference Strategies Tournament (GHIST) is a community-driven competition designed to evaluate methods for inferring evolutionary history from population genomic data. The inaugural GHIST competition ran from July to November 2024 and featured four demographic history inference challenges of varying complexity: a bottleneck model, a split with isolation model, a secondary contact model with demographic complexity, and an archaic admixture model. Data were provided as error-free VCF files, and participants submitted numerical parameter estimates that were scored by relative root mean squared error. Approximately 60 participants competed, using diverse approaches. Results revealed the current dominance of methods based on site frequency spectra, while highlighting the advantages of flexible model-building approaches for complex demographic histories. We discuss insights regarding the competition and outline the next iteration, which is ongoing with expanded challenge diversity. By providing standardized benchmarks and highlighting areas for improvement, GHIST represents a substantial step toward more reliable inference of evolutionary history from genomic data.

The rapid expansion of single-cell transcriptomics necessitates scalable and efficient deep-learning frameworks. These frameworks must process massive datasets and reduce dependency on export-constrained hardware. In this study, we present the porting, optimization, and benchmarking of the Geneformer model on the Cambricon MLU290 intelligent accelerator. This device serves as a high-performance, domestically available alternative to mainstream NVIDIA GPUs. We systematically adapted the PyTorch and transformers codebases to leverage Cambricon’s Torch-MLU and CNCL libraries. This enabled seamless distributed training with enhanced communication and memory management. Experimental results show that the MLU290 delivers a remarkable 1,027,734 training iterations per card in 8-card clusters. This performance outpaces the NVIDIA V100 and RTX 4090 by nearly eightfold in comparable configurations. Communication efficiency consistently exceeded 91% in multimachine settings. Optimization of data loading reduced average iteration time by 7.8%. Detailed profiling revealed that synchronization operations accounted for over 33% of total runtime. This identifies clear targets for further performance improvement. Our findings validate the MLU290 as a cost-effective and scalable accelerator for biomedical AI. This study provides a transferable methodology for hardware-agnostic, large-scale deployment of deep learning models in bioinformatics. The work lays the groundwork for a broader adoption of accessible AI hardware in genomics. Finally, this study also suggests future directions in bottleneck optimization, model generalization, and secure federated learning.