İnci Meliha Baytaş

Motivation: Accurate prediction of protein function is essential for elucidating molecular mechanisms and advancing biological and therapeutic discovery. Yet experimental annotation lags far behind the rapid growth of protein sequence data. Computational approaches address this gap by associating proteins with Gene Ontology (GO) terms, which encode functional knowledge through hierarchical relations and textual definitions. However, existing models often emphasize one modality over the other, limiting their ability to generalize, particularly to unseen or newly introduced GO terms that frequently arise as the ontology evolves, and making the previously trained models outdated. Results: We present STAR-GO, a Transformer-based framework that jointly models the semantic and structural characteristics of GO terms to enhance zero-shot protein function prediction. STAR-GO integrates textual definitions with ontology graph structure to learn unified GO representations, which are processed in hierarchical order to propagate information from general to specific terms. These representations are then aligned with protein sequence embeddings to capture sequence–function relationships. STAR-GO achieves state-of-the-art performance and superior zero-shot generalization, demonstrating the utility of integrating semantics and structure for robust and adaptable protein function prediction. Availability: Code and pre-trained models are available at https://github.com/boun-tabi-lifelu/stargo, https://doi.org/10.5281/zenodo.18643082

Background and Objectives: Mild cognitive impairment (MCI) is heterogeneous within the Alzheimer’s disease (AD) continuum, and categorical labels may not reflect biological variability. We evaluated whether multimodal machine learning using routine clinical data and neuroimaging could support biologically informed enrichment across MCI and AD in a real-world memory clinic cohort. Methods: We analyzed 474 patients (1547 visits) with clinical and cognitive measures, laboratory parameters, MRI regional volumes, and FDG-PET regional uptake. Elastic Net and gradient boosting models were trained using nested cross-validation with strict patient-level separation. Results: Model discrimination improved as additional data modalities were added, and FDG-PET contributed the largest performance improvement. Hypometabolism in posterior default mode network regions consistently emerged as the most influential predictor. In the MCI subgroup, AD-like scores showed a continuous distribution consistent with biological enrichment. Conclusions: Multimodal models may provide an interpretable enrichment framework in heterogeneous memory clinic populations.

Over-squashing is a challenge in training graph neural networks for tasks involving long-range dependencies. In such tasks, a GNN's receptive field should be large enough to enable communication between distant nodes. However, gathering information from a wide range of neighborhoods and squashing its content into fixed-size node representations makes message-passing vulnerable to bottlenecks. Graph rewiring and adding virtual nodes are commonly studied remedies that create additional pathways around bottlenecks to mitigate over-squashing. However, these techniques alter the input graph's global topology and disrupt the domain knowledge encoded in the original graph structure, both of which could be essential to specific tasks and domains. This study presents Local Virtual Nodes (LVN) with trainable embeddings to alleviate the effects of over-squashing without significantly corrupting the global structure of the input graph. The position of the LVNs is determined by the node centrality, which indicates the existence of potential bottlenecks. Thus, the proposed approach aims to improve the connectivity in the regions with likely bottlenecks. Furthermore, trainable LVN embeddings shared across selected central regions facilitate communication between distant nodes without adding more layers. Extensive experiments on benchmark datasets demonstrate that LVNs can enhance structural connectivity and significantly improve performance on graph and node classification tasks. The code can be found at https://github.com/ALLab-Boun/LVN/}{https://github.com/ALLab-Boun/LVN/.

Sign language, primary communication medium of the Deaf, uses visual cues from upper body, hands and face. Sign Language Recognition (SLR) aims to learn salient representations from these cues to bridge the communication gap between Deaf and hearing communities. The state-of-the-art Graph Neural Network-based SLR frameworks often transform sign videos into a sequence of graphs formed by hand and body joints. We show that learning sign representations using only body topology results in suboptimal solutions. We also show that training the graph-based SLR frameworks with a domain-specific hand topology can single-handedly help reach state-of-the-art recognition performance. Therefore, fusing the representations learned from the topologies of multiple visual cues is essential to a generalizable and robust SLR framework. However, integrating multiple cues in an SLR framework becomes challenging due to the difference in the spatial and temporal dynamics of hand and upper body movements. This study exposes the differences in the temporal patterns of hand and body representations extracted from the Spatio- Temporal Graph Convolutional Network (ST-GCN) using self-attention. Considering these distinctions, we propose a multi-cue cross-attention framework to enable hand and body cues to attend to each other differently while fusing their representations for SLR. Thus, the proposed attention-based fusion framework allows us to explain how individual articulators affect the model’s predictions for correctly classified and misclassified signs. Our experiments on two benchmark isolated sign language datasets, BosphorusSign22k and AUTSL, show that the proposed framework’s performance of isolated SLR is on par with state-of-the-art while highlighting the benefit of choosing domain-specific hand graph topologies and fusing multiple cues for SLR. Furthermore, our cross-attentive approach for fusion of upper-body and hand cues improves recognition accuracy by around 1% and 3% on the respective datasets over hand-only models, while making the SL recognition interpretable, and demonstrates the complementary interactions between visual cues.

Alzheimer’s disease (AD) is a neurodegenerative disorder that leads to cerebral atrophy, impacting memory and cognitive abilities. A precursor to AD known as Mild Cognitive Impairment (MCI) shows subtle symptoms that do not overwhelm the patients’ daily activities. MCI patients might eventually progress to AD in later stages. Early detection of the conversion is a vital step in preventative treatment planning. However, conversion detection is challenging due to the rarity of conversion visits in public datasets and the unknown nature of the conversion. This study aims to improve conversion detection with an attention-based architecture designed to encode input biomarkers and time into a shared space where time and attribute embeddings are fused with attention. Temporal information is incorporated as a separate modality with time embeddings to capture the correlation between time and feature significance for the model’s predictions. Experiments with widely used public databases (TADPOLE and NACC) show encouraging performance in conversion detection. In TADPOLE, a conversion recall of 74.3%, significantly outperforming baseline models such as logistic regression (36.9%) and Long Short-Term Memory networks (62.3%), is reported while maintaining an area under the curve (AUC) score of 82.0%. In NACC, our model demonstrates a competitive conversion recall of 71.6% and an AUC of 82.6%. The experimental results highlight the contribution of the attention between time and attributes to MCI-AD conversion recall. The experimental analyses hold promise for assisting physicians in designing targeted preventative treatment strategies for at-risk individuals. The implementation of the proposed method is available at https://github.com/ALLab-Boun/FATE-Net.

Graph Neural Networks (GNNs) set the state-of-the-art in representation learning for graph-structured data. They are used in many domains, from online social networks to complex molecules. Most GNNs leverage the message-passing paradigm and achieve strong performances on various tasks. However, the message-passing mechanism used in most models suffers from over-smoothing as a GNN's depth increases. The over-smoothing degrades GNN's performance due to the increased similarity between the representations of unrelated nodes. This study proposes an adaptive channel-wise message-passing approach to alleviate the over-smoothing. The proposed model, Channel-Attentive GNN, learns how to attend to neighboring nodes and their feature channels. Thus, much diverse information can be transferred between nodes during message-passing. Experiments with widely used benchmark datasets show that the proposed model is more resistant to over-smoothing than baselines and achieves state-of-the-art performances for various graphs with strong heterophily. Our code is at https://github.com/ALLab-Boun/CHAT-GNN.

Deep Edge Intelligence targets the deployment of deep learning algorithms at the edge devices to make them contribute to the overall intelligence in the Internet of Things. While training deep networks requires computational resources, edge devices usually lack high computational power. Decentralized online learning methods provide a solution for gathering limited information from edge devices to collectively improve prediction performance. However, methods that require multiple rounds of data transfer between edge components for an acceptable accuracy increase not only the training time but also the risk of communication errors due to connectivity failures that result in packet delays and losses. Besides, every device at the edge may not experience the same training data due to limitations set by physical constraints such as the location of devices and environmental factors. To overcome these bottlenecks, this study proposes a convolutional ensemble learning approach, coined EdgeConvEns, that facilitates training heterogeneous weak models on edge and learning to ensemble them by fusing the limited knowledge obtained from edge where data on edge are heterogeneously distributed. In this setup, it is possible for an edge device to barely train a shallow network that fits its resources. Losses that occur while transferring their feature representations to the edge server are handled by variational autoencoders, which learn to fill in missing representations. The last stage is the proposed convolutional ensemble model at the edge server which uses all representations during training and inference to boost the prediction performance. The overall proposed approach can also be employed in setups where edge devices can finely tune pre-trained inferences. To accelerate the training on edge devices, a flexible chained-tiling method is also proposed and demonstrated on field programmable gate array (FPGA) devices with various computational capacities. Extensive experiments demonstrate that EdgeConvEns can outperform the state-of-the-art performance with fewer communications and less data in training scenarios selected from the literature.

Adversarial Training (AT) aims to alleviate the vulnerability of deep neural networks to adversarial perturbations. However, the AT techniques struggle to maintain the performance on natural samples while improving the deep model’s robustness. The absence of perturbation diversity in generated during the adversarial training degrades the generalizability of the robust models, causing overfitting to particular perturbations and a decrease in natural performance. This study proposes an adversarial training framework that augments adversarial directions from a single-step attack to address the trade-off between robustness and generalization. Inspired by feature scattering adversarial training, the proposed framework computes a principal adversarial direction with a single-step attack that finds a perturbation disrupting the inter-sample relationships in the mini-batch during adversarial training. The principal direction obtained at each iteration is augmented by sampling new adversarial directions within a region spanning 45 degrees from the principal adversarial direction. The proposed adversarial training approach does not require extra backpropagation steps in adversarial direction augmentation. Therefore, generalization of the robust model is improved without posing an additional burden on the feature scattering adversarial training. Experiments on CIFAR-10, CIFAR-100, SVHN, Tiny-ImageNet, and The German Traffic Sign Recognition Benchmark consistently improve the accuracy on adversarial with an almost pristine natural performance.

Early diagnosisof Alzheimer's disease plays a crucial role in treatment planning that might slow down the disease's progression. This problem is commonly posed as a classification task performed by machine learning and deep learning techniques. Although data-driven techniques set the state-of-the-art in many domains, the scale of the available datasets in Alzheimer's research is not sufficient to learn complex models from patient data. This study proposes a simple yet promising framework to predict the conversion from Mild Cognitive Impairment (MCI) to Alzheimer's Disease (AD). The proposed framework comprises a shallow neural network for binary classification and a single-step gradient-based adversarial attack to find an adversarial progression direction in the input space. The step size required for the adversarial attack to change a patient's diagnosis from MCI to AD indicates the distance to the decision boundary. The patient's diagnosis at the next visit is predicted by employing this notion of distance to the decision boundary. We also present a potential application of the proposed framework to patient subtyping. Experiments with two publicly available datasets for Alzheimer's disease research imply that the proposed framework can predict MCI-to-AD conversions and assist in subtyping by only training a shallow neural network.