Inês Dutra

Abstract

Abstract

Abstract

Abstract
This paper presents a novel approach to the computational representation of cellular lines using transformer-based embeddings. By leveraging state-of-the-art natural language processing techniques, we generate context-aware embeddings from biomedical literature from the PubMed database, offering a more nuanced and biologically relevant representation of cellular lines compared to traditional methods like TF-IDF and SVDD. We applied these embeddings to cluster cellular lines, using the elbow method to identify a set of distinct clusters that reflect biologically meaningful relationships. To evaluate the quality of these clusters, we employed the Topic Coherence metric, achieving a coherence score of 0.395, indicative of moderate consistency across clusters. The results demonstrate the potential of transformer-based models to improve drug discovery by identifying shared characteristics between cellular lines, enabling more accurate drug response predictions and advancing personalized medicine. This method offers an interesting improvement in the precision of cellular line modeling, paving the way for more efficient drug repositioning and targeted therapies in cancer research.

Abstract
Background Intrusion Tolerant Systems (ITS) aim to maintain system security despite adversarial presence by limiting the impact of successful attacks. Current ITS risk managers rely heavily on public databases like NVD and Exploit DB, which suffer from long delays in vulnerability evaluation, reducing system responsiveness.Objective This work extends the HAL 9000 Risk Manager to integrate additional real-time threat intelligence sources and employ machine learning techniques to automatically predict and reassess vulnerability risk scores, addressing limitations of existing solutions.Methods A custom-built scraper collects diverse cybersecurity data from multiple Open Source Intelligence (OSINT) platforms, such as NVD, CVE, AlienVault OTX, and OSV. HAL 9000 uses machine learning models for CVE score prediction, vulnerability clustering through scalable algorithms, and reassessment incorporating exploit likelihood and patch availability to dynamically evaluate system configurations.Results Integration of newly scraped data significantly enhances the risk management capabilities, enabling faster detection and mitigation of emerging vulnerabilities with improved resilience and security. Experiments show HAL 9000 provides lower risk and more resilient configurations compared to prior methods while maintaining scalability and automation.Conclusions The proposed enhancements position HAL 9000 as a next-generation autonomous Risk Manager capable of effectively incorporating diverse intelligence sources and machine learning to improve ITS security posture in dynamic threat environments. Future work includes expanding data sources, addressing misinformation risks, and real-world deployments.