Biology subjects

Abir, A. R.

Publications and source records attributed to Abir, A. R..

8 recordsLinked to original sources

bioRxiv2026

ConnectoFM: A Foundation Model for Learning the Language of the Connectome

Accurate reconstruction of neural circuits from electron microscopy (EM) data is central to connectomics, yet modern datasets are now so large and heterogeneous that manual annotation and dataset-specific model retraining have become major challenges. While recent EM foundation models provide general visual representations, they are not specifically tailored to the connectomics domain, where preserving fine membrane boundaries and synaptic structures is essential to mitigate topological and connectivity errors. Here, we present ConnectoFM, the first foundation model for connectomics, pretrained on a diverse corpus of 1.7 million unlabeled EM images drawn from six species and 25 subdomains. ConnectoFM combines masked image modeling with contrastive alignment to learn robust visual representations directly from large-scale connectomics data. These representations organize EM images into biologically meaningful clusters across species, brain regions, developmental cohorts, and acquisition domains. Using frozen pretrained features with lightweight decoder heads, we transfer ConnectoFM to three important downstream tasks: binary segmentation, multiclass cell typing, and instance segmentation. Across 29 diverse datasets, including established benchmarks, ConnectoFM consistently outperforms existing EM foundation models and state-of-the-art methods that require task-specific training from scratch. With only 10% labeled data, ConnectoFM surpasses the baselines trained on 100% annotation budget, showing the superiority of ConnectoFM in low-data regimes. Improvements of ConnectoFM are especially pronounced for challenging and biologically important targets, including membranes, mitochondria, vesicles, post-synaptic densities and synapses, and remain strong in low-label settings. Extension to 3D volumetric segmentation and qualitative comparisons further show that ConnectoFM enables more accurate and biologically faithful performance across downstream tasks. These results establish ConnectoFM as a generalizable and data-efficient foundation model for connectomics and provide a scalable route towards more reliable neural circuit reconstruction.

neuroscience↗

bioRxiv2025

MExConn: A Mechanistically Interpretable Multi-Expert Framework for Multi-Organelle Segmentation in Connectomics

Electron microscopy (EM) provides subcellular resolution which has made it a critical tool in fields such as cellular biology and connectomics. However, manual annotation of subcellular organelles in these EM images is extremely labor-intensive and impractical at scale. While computational segmentation methods have been developed, most existing approaches are limited to segmenting a single organelle at a time, neglecting the inherent shared information present in EM images containing multiple organelles. To address this, we present MExConn, the first known interpretable multi-expert U-Net architecture in the connectomics field that employs a shared encoder and multiple decoder heads to simultaneously segment multiple organelles from the same input EM image. MExConn significantly outperforms five baselines, including single-organelle model and four state-of-the-art connectomics segmentation models in all evaluation metrics, reducing the Variation of Information by up to 33.54% on average across organelles. A key novelty of our approach is that MExConn offers mechanistic interpretability by revealing that the shared encoder learns shared representations essential for accurately segmenting multiple organelles. Through systematic analysis of encoder gradients with respect to each decoder output, we identify channel-wise importance profiles and reveal that many encoder channels are jointly essential for all organelles, while others are organelle-specific. Rigorous experiments on three connectomics datasets demonstrate the effectiveness of MExConn in both segmentation performance and interpretability, establishing it as a principled approach for multi-organelle analysis in connectomics. The source code is publicly available at https://github.com/abrarrahmanabir/MExConn.

neuroscience↗

bioRxiv2025

MACS: Multi-Domain Adaptation Enables Accurate Connectomics Segmentation

Connectomics aims to map the brains neural wiring by segmenting cellular structures from high-resolution electron microscopy (EM) images. Manual labeling and proofreading remain a major bottleneck for accurate extraction of microstructures. While computational models have advanced automated segmentation, they typically require training from scratch on each dataset, demanding substantial annotated data. Domain adaptation methods address this by transferring knowledge from a labeled source to a less-annotated target. However, existing approaches are limited to adaptation from a single source domain. This overlooks the potential benefits of integrating information from multiple diverse domains, motivating the development of multidomain adaptation. To address this, we propose MACS, the first known multi-domain adaptation framework that combines knowledge from multiple heterogeneous source domains to learn segmentation in the target domain, and employs active learning to efficiently select the most informative target samples for annotation. MACS uses information-theoretic weighting to combine source domains, and introduces a novel and efficient Bayesian Laplace approximation for uncertainty estimation. Our extensive experiments across nine connectomics datasets demonstrate that MACS consistently and substantially outperforms state-of-the-art models, even under limited annotation budgets, with a mean improvement of 5.89% at the lowest annotation budget and 27.72% at the highest annotation budget. In-depth analyses further reveal that MACS offers mechanistic interpretability by quantifying and explicitly upweighting the most transferable source domains for each target. The preprocessed datasets and the source code of MACS are publicly available at http://github.com/abrarrahmanabir/MACS.

neuroscience↗

bioRxiv2025

RNA-EFM : Energy based Flow Matching for Protein-conditioned RNA Sequence-Structure Co-design

Ribonucleic acids (RNAs) are essential biomolecules involved in gene regulation and molecular recognition. Designing RNA molecules that can bind specific protein targets is crucial for therapeutic applications but remains challenging due to the structural flexibility of RNA and the laborious nature of experimental techniques. We propose RNA-EFM, a novel Energy-based Flow Matching framework for protein-conditioned RNA sequence and structure co-design. RNA-EFM integrates biophysical constraints, including the Lennard-Jones potential and sequence-derived free energy, to generate low-energy and biologically plausible RNA conformations. By incorporating an idempotent refinement strategy for iterative structural correction, RNA-EFM consistently outperforms existing baselines, achieving lower RMSD, higher lDDT, and superior sequence recovery across multiple evaluation splits.

bioinformatics↗

bioRxiv2024

UnCOT-AD: Unpaired Cross-Omics Translation Enables Multi-Omics Integration for Alzheimer's Disease Prediction

Alzheimers Disease (AD) is a progressive neurodegenerative disorder, posing a growing public health challenge. Traditional machine learning models for AD prediction have relied on single omics data or phenotypic assessments, limiting their ability to capture the diseases molecular complexity and resulting in poor performance. Recent advances in high-throughput multi-omics have provided deeper biological insights. However, due to the scarcity of paired omics datasets, existing multi-omics AD prediction models rely on unpaired omics data, where different omics profiles are combined without being derived from the same biological sample, leading to biologically less meaningful pairings and causing less accurate predictions. To address these issues, we propose UnCOT-AD, a novel deep learning framework for Unpaired Cross-Omics Translation enabling effective multi-omics integration for AD prediction. Our method introduces the first-ever cross-omics translation model trained on unpaired omics datasets, using two coupled Variational Autoencoders and a novel cycle consistency mechanism to ensure accurate bidirectional translation between omics types. We integrate adversarial training to ensure that the generated omics profiles are biologically realistic. Moreover, we employ contrastive learning to capture the disease specific patterns in latent space to make the cross-omics translation more accurate and biologically relevant. We rigorously validate UnCOT-AD on both cross-omics translation and AD prediction tasks. Results show that UnCOT-AD empowers multi-omics based AD prediction by combining real omics profiles with corresponding omics profiles generated by our cross-omics translation module and achieves state-of-the-art performance in accuracy and robustness. Source code is available at https://github.com/abrarrahmanabir/UnCOT-AD

bioinformatics↗

bioRxiv2024

DeepRNA-Twist: Language Model guided RNA Torsion Angle Prediction with Attention-Inception Network

RNA torsion and pseudo-torsion angles are critical in determining the three-dimensional conformation of RNA molecules, which in turn governs their biological functions. However, current methods are limited by RNAs structural complexity and flexibility, as it can adopt multiple conformations, with experimental techniques being costly and computational approaches struggling to capture the intricate sequence dependencies needed for accurate predictions. To address these challenges, we introduce DeepRNA-Twist, a novel deep learning framework designed to predict RNA torsion and pseudo-torsion angles directly from sequence. DeepRNA-Twist utilizes RNA language model embeddings, which provides rich, context-aware feature representations of RNA sequences. Additionally, it introduces 2A3IDC module (Attention Augmented Inception Inside Inception with Dilated CNN), combining inception networks with dilated convolutions and multi-head attention mechanism. The dilated convolutions capture long-range dependencies in the sequence without requiring a large number of parameters, while the multi-head attention mechanism enhances the models ability to focus on both local and global structural features simultaneously. DeepRNA-Twist was rigorously evaluated on benchmark datasets, including RNA-Puzzles, CASP-RNA, and SPOT-RNA-1D, and demonstrated significant improvements over existing methods, achieving state-of-the-art accuracy. Source code is available at https://github.com/abrarrahmanabir/DeepRNA-Twist

bioinformatics↗

bioRxiv2024

BioLLMNet: Enhancing RNA-Interaction Prediction with a Specialized Cross-LLM Transformation Network

Existing computational methods for the prediction of RNA related interactions often rely heavily on manually crafted features. Language model features for bio-sequences has gain significant popularity in proteomics and genomics. However, during interaction prediction, how language model features from different modalities should be combined to extract the most representative features is yet to be explored. We introduce BioLLMNet, a novel framework that introduces an effective combination approach for multi-modal bio-sequences. BioLLMNet provides a way to transform feature space of different molecules language model features and uses learnable gating mechanism to effectively fuse features. Rigorous evaluations show that BioLLMNet achieves state-of-the-art performance in RNA-protein, RNA-small molecule, and RNA-RNA interactions, outperforming existing methods in RNA-associated interaction prediction.

bioinformatics↗

bioRxiv2024

LOCAS: Multi-label mRNA Localization with Supervised Contrastive Learning

Traditional methods for mRNA subcellular localization often fail to account for multiple compartmentalization. Recent multi-label models have improved performance, but still face challenges in capturing complex localization patterns. We introduce LOCAS (Localization with Supervised Contrastive Learning), which integrates an RNA language model to generate initial embeddings, employs supervised contrastive learning (SCL) to identify distinct RNA clusters, and uses a multi-label classification head (ML-Decoder) with cross-attention for accurate predictions. Through extensive ablation studies and multi-label overlapping threshold tuning, LOCAS achieves state-of-the-art performance across all metrics, providing a robust solution for RNA localization tasks.

bioinformatics↗