הצגת סמינר סטודנטים- לורנה בחית מאסטרנטית בהנחיית תמיר טולר ועבאדיין עודיי-בהנחיית פרופ' עופר ברנע אמריטוס

Understanding and Predicting Gene Coding Sequences with Natural Language Processing-Lorna Bahit

Abstract

Codon usage Bias, a phenomenon influenced by selective pressure, influences several aspects of protein production, including all gene expression steps and protein co-translational folding. However, traditional computational tools often fall short in capturing the complex interactions of codons across different species. To address this, our study employs a new version of BART (Bidirectional and Auto-Regressive Transformer) model, a state-of-the-art deep learning technique, to enhance the prediction of codon sequences. We trained BART on sequences from four different organisms—Saccharomyces cerevisiae, Saccharomyces castellii, Saccharomyces kudriavzevii, and Lachancea kluyveri—and improved its performance by incorporating information about orthologous codons. Orthologs, which are genes conserved across species due to common ancestry, offer valuable insights into conserved codon usage patterns. By integrating this evolutionary context, we were able to enhance the accuracy of our predictions, particularly in genes with highly conserved orthologs. Furthermore, our results suggest that OrthologBART is able to implicitly capture expression-related codon patterns, even in species without direct expression data, by learning from the evolutionary pressure on codon usage across gene regions. Notably, the model showed an ability to predict codons with higher confidence in regions of genes subject to regulatory pressures, such as the first and last codons in the coding regions. Using our tool, we were able to predict codons in new coding sequences more accurately than previous models, offering deeper insights into codon distribution both across and within genes.

We believe that our approach not only advances our understanding of codon selection mechanisms but may also provide a powerful tool for optimizing protein synthesis in diverse biological and industrial contexts.

Data analysis for intracranial parameter estimation in TBI patient-Oday Abdeen

Supervisor: Prof Emeritus Ofer Barnea

Abstract

To better understand intracranial dynamic in TBI patients, we developed a lumped parameter model which includes a representation of cerebral ISF flow within brain tissue and its interactions with CSF flow and cerebral blood flow (CBF). The model is based on an electrical analog circuit with four intracranial compartments: the (1) subarachnoid space, (2) brain, (3) ventricles, (4) cerebral vasculature and the extracranial spinal thecal sac. This model resulted in accurate predictions and elucidated intracranial fluid dynamics. To continue with real-time parameter-estimation of critical parameters that cannot be measured at this time, we reduced this model to a minimal model that consists of only 4 parameters. The values of the parameters are calculated using the model with the patient’s arterial pressure waveform as an input. Parameters are optimized to generate the ICP waveform that is also obtained from the patient. The optimization procedure is an exhaustive search within physiological boundaries. The study is performed at Hadassah Medical Center. It is expected to yield indications of the severity of Edema and provide an indication on intracranial elasticity and fluid output resistance – all vital parameters for patient management.