Biology subjects

Sailem, H.

Publications and source records attributed to Sailem, H..

8 recordsLinked to original sources

bioRxiv2026

Leveraging Foundation Models for the Characterisation of Small RNA Properties

ABSTRACT Small interfering RNAs (siRNAs) provide a promising therapeutic approach capable of selectively silencing disease-associated genes; however, achieving high efficacy and specificity while minimising off-target effects remains a significant challenge. Endogenous small RNAs, such as microRNAs (miRNAs) and PIWI-interacting RNAs (piRNAs), exhibit structural features supporting their functions and are biocompatible. Recent advances in RNA foundation models, such as RNA-FM, enable large-scale learning of sequence and structural representations of RNA sequences, offering a powerful framework for studying small RNA functions. Here, we leverage RNA-FM model alongside interpretable biological features to systematically compare endogenous small RNAs (miRNAs and piRNAs) with synthetic siRNAs. Biological features highlighted class-specific patterns: piRNAs showed significantly higher GC content and melting temperature than miRNAs and siRNAs, suggesting higher stability. Importantly, we mapped RNA-FM embeddings to interpretable features to better understand deep learning outputs and facilitate effective extraction of functionally relevant information. To support predictive and comparative analyses of small RNAs, we implemented these functionalities in RNAExplorer (www.rnaexplorer.com), a web-based application that allows analysing and visualising small RNA features interactively. Together, our integrative analysis provides a framework for understanding small RNA biology and improving siRNA therapeutic design strategies.

genetics↗

bioRxiv2025

ADAPTIVE MULTI-SCALE GRAPH TRANSFORMER FRAMEWORK FORHISTOPATHOLOGICAL IMAGES

Whole slide images (WSIs) contain hierarchical information from cellular to tissue architecture but their gigapixel scale poses major memory and computational challenges. Existing multi-scale graph and transformer models capture complex WSI features effectively but struggle with efficiency. We propose an Adaptive Multi-Scale Graph Transformer (AMGT) for WSI classification that addresses this limitation through two key modules: a Self-Guided Token Aggregation (SGTA) mechanism that fuses multi-resolution features to reduce redundancy, and a Prototypical Transformer (PT) that groups similar tokens into phenotype-representative prototypes with linear complexity. This design preserves essential spatial and semantic information, substantially lowering memory cost and improving interpretability by prototypical learning. AMGT achieves superior performance and efficiency, outperforming state-of-the-art models by 1.8% and 5.3% AUC on high-grade ovarian cancer and Camelyon16 datasets, respectively. These results demonstrate AMGTs capacity for scalable, interpretable multi-scale representation learning.

bioinformatics↗

bioRxiv2025

HRDPath: An Explainable Multi-Model Deep Learning Architecture for Predicting Homologous Recombination Deficiency from Histopathology Images

Homologous recombination deficiency (HRD) is a critical biomarker for guiding treatment decisions in high-grade serous tubo-ovarian carcinoma (HGSOC), a cancer with few reliable biomarkers. However, existing genomic-based tests for HRD are variable, expensive, and time-consuming. To this end, we developed HRDPath, a novel patient-level deep learning architecture that combines the strengths of two complementary models with a multi-task design, to predict genomically derived HRD status from whole slide images in HGSOC. HRDPath was comprehensively validated across three datasets and benchmarked against leading deep learning models. It achieved an AUC of 0.846, surpassing previously reported H&E-based HRD prediction results for HGSOC images by 0.09, and for the first time, reporting a specificity of 0.938, where accuracy significantly increased when multiple slides per patient were used. Our proposed patient-level approach and interpretability pipeline enhance model trustworthiness and reveal important clinical and biological insights into HRD-positive cancers, highlighting the associated morphological and pathological changes at the cellular and tissue levels. HRDPath is a potentially accessible and scalable digital biomarker that could improve ovarian cancer diagnosis and therapy selection.

bioinformatics↗

bioRxiv2024

TissuePlot: A Multi-Scale Interactive Visualization Tool for Spatial Data

AO_SCPLOWBSTRACTC_SCPLOWVisualization of spatial datasets is essential for understanding biological systems that are composed of several interacting cell types. For example, gene expression data at the molecular level needs to be interpreted based on cell type, spatial context, tissue type, and interactions with the surrounding environment. Recent advances in spatial profiling technologies allow measurements of the level of thousands of proteins or genes at different spatial locations along with corresponding cellular composition. Representing such high dimensional data effectively to facilitate data interpretation is a major challenge. Existing methods such as spatially plotted pie charts obscure underlying tissue regions and necessitate switching between different views for comprehensive interpretations. Here, we present TissuePlot, a novel method for visualizing spatial data. TissuePlot tackles the key challenge of visualizing multi-scale phenotypic data at molecular, cellular and tissue level in the context of their spatial locations. To this end, TissuePlot employs a transparent hexagon tesselation approach that utilizes object borders to represent cell composition or gene-level data without obscuring the underlying cell image. Moreover, we implement a multi-view interactive approach, to allow interrogating spatial tissue data at multiple scales linking molecular information to tissue anatomy and motifs. We demonstrate TissuePlot utility using mouse brain data from the Bio+MedVis Redesign Challenge 2024. Our tool is accessible at https://sailem-group.github.io/TissuePlot/.

bioinformatics↗

bioRxiv2023

Decoding mTOR signalling heterogeneity in the tumour microenvironment using multiplexed imaging and graph convolutional networks

Evaluating the contribution of the tumour microenvironment (TME) in tumour progression has proven a complex challenge due to the intricate interactions within the TME. Multiplexed imaging is an emerging technology that allows concurrent assessment of multiple of these components simultaneously. Here we utilise a highly multiplexed dataset of 61 markers across 746 colorectal tumours to investigate how complex mTOR signalling in different tissue compartments influences patient prognosis. We found that the signalling of mTOR pathway can have heterogeneous activation patterns in tumour and immune compartments which correlate with patient prognosis. Using graph neural networks, we determined the most predictive features of mTOR activity in immune cells and identified relevant cellular subpopulations. We validated our observations using spatial transcriptomics data analysis in an independent patient cohort. Our work provides a framework for studying complex cell signalling and reveals important insights for developing mTOR-based therapies.

cancer biology↗

bioRxiv2022

Efficient Bayesian inference for mechanistic modelling with high-throughput data

Bayesian methods are routinely used to combine experimental data with detailed mathematical models to obtain insights into physical phenomena. However, the computational cost of Bayesian computation with detailed models has been a notorious problem. Moreover, while high-throughput data presents opportunities to calibrate sophisticated models, comparing large amounts of data with model simulations quickly becomes computationally prohibitive. Inspired by the method of Stochastic Gradient Descent, we propose a minibatch approach to approximate Bayesian computation. Through a case study of a high-throughput imaging scratch assay experiment, we show that reliable inference can be performed at a fraction of the computational cost of a traditional Bayesian inference scheme. By applying a detailed mathematical model of single cell motility, proliferation and death to a data set of 118 gene knockdowns, we characterise functional subgroups of gene knockdowns, each displaying its own typical combination of local cell density-dependent and -independent motility and proliferation patterns. By comparing these patterns to experimental measurements of cell counts and wound closure, we find that density-dependent interactions play a crucial role in the process of wound healing.

bioinformatics↗

bioRxiv2021

PERK-dependent reciprocal crosstalk between ER and non-centrosomal microtubules coordinates ER architecture and cell shape

The architecture of the endoplasmic reticulum (ER) is tightly controlled as a key determinant of its function. Its dynamics are linked to those of the cytoskeleton, but our understanding of how this coordination occurs and what its functional relevance is, is limited. We found the Unfolded Protein Response (UPR) transducer EIF2AK3/PERK is essential for acute stress-induced peripheral redistribution and remodeling of the ER, through eIF2a phosphorylation and translation initiation shutdown. PERK-mediated eIF2a phosphorylation can be bypassed by blocking ribosome activity; by depleting microtubule-anchoring ER proteins such as REEP4, p180/RRBP1 and Climp63/CKAP4; or by disrupting the microtubule cytoskeleton. Notably, specific disruption of non-centrosomal microtubules, but not centrosome depletion, relieved blockade of ER redistribution in PERK-deficient cells. Conversely, PERK deficiency stabilized non-centrosomal microtubules, promoting polarized protrusiveness in epithelial cells and neuroblasts. We propose that PERK coordinates ER architecture and homeostasis with cell morphogenesis by coupling ER remodeling and non-centrosomal MT dynamics.

cell biology↗

bioRxiv2019

KDML: a machine-learning framework for inference of multi-scale gene functions from genetic perturbation screens

Characterising context-dependent gene functions is crucial for understanding the genetic bases of health and disease. To date, inference of gene functions from large-scale genetic perturbation screens is based on ad-hoc analysis pipelines involving unsupervised clustering and functional enrichment. We present Knowledge-Driven Machine Learning (KDML), a framework that systematically predicts multiple functions for a given gene based on the similarity of its perturbation phenotype to those with known function. As proof of concept, we test KDML on three datasets describing phenotypes at the molecular, cellular and population levels, and show that it outperforms traditional analysis pipelines. In particular, KDML identified an abnormal multicellular organisation phenotype associated with the depletion of olfactory receptors and TGF{beta} and WNT signalling genes in colorectal cancer cells. We validate these predictions in colorectal cancer patients and show that olfactory receptors expression is predictive of worse patient outcome. These results highlight KDML as a systematic framework for discovering novel scale-crossing and clinically relevant gene functions. KDML is highly generalizable and applicable to various large-scale genetic perturbation screens.

systems biology↗