A network-based deep learning framework for the in silico prioritization of candidate synthetic-lethal interactions.
Traditional oncology relies on broad treatment protocols that help only a small subset of patients. Cancer is biologically heterogeneous, and treating it as a single disease limits progress. Precision oncology responds to this reality by aligning therapy with the specific molecular state of each tumor. Synthetic lethality enables this shift by allowing cancer-specific targeting instead of broad cytotoxic exposure. In this setting, accuracy in drug discovery is essential.
Synthetic lethality is a well-established concept in cancer biology. A cell can tolerate disruption of one gene, but disruption of a specific paired gene leads to cell death. In cancer, one of these genes is often already altered by mutation or deletion. Targeting its partner exposes a vulnerability that is selective to tumor cells, reducing damage to healthy tissue and improving therapeutic durability.
Genes act within complex, interconnected systems. Mutations propagate through networks rather than isolated pathways. We represent this complexity using a large-scale Knowledge Graph containing 1.4 million biological relationships. Graph Neural Networks operate on this structure to infer gene interactions and predict synthetic lethal links that have not yet been observed in CRISPR screens, extending discovery beyond current experimental limits.
The initial focus is glioblastoma, one of the most aggressive human cancers, where survival outcomes have seen little improvement. The anchor is PTEN (phosphatase and tensin homolog), a frequently mutated or deleted tumor suppressor in glioblastoma. PTEN loss reshapes signaling, metabolism, and survival pathways, making it a biologically grounded entry point for uncovering actionable vulnerabilities.
The model is built on high-quality, large-scale datasets. Functional dependency signals come from DepMap Public 25Q4, integrating both CRISPR and RNAi screens across diverse cancer contexts. This is combined with curated synthetic lethality knowledge from SynLethDB 2.0, forming a structured network of known and candidate genetic interactions.
On this foundation, we apply knowledge-aware Graph Neural Networks with attention-based link prediction. The model learns from network structure and biological context, weighting evidence based on relevance and consistency. The output is a set of synthetic lethal predictions that are mechanistically plausible and experimentally testable. The scope remains narrow by design, prioritizing depth and clarity over breadth.
Median survival for glioblastoma remains 15 months. This has not meaningfully changed in decades. 36% of patients carry PTEN mutations or deletions. The human protein interaction network is too large for manual reasoning, and trial-and-error drug discovery cannot keep pace with its complexity. Precision computation provides a path forward by converting biological complexity into actionable insight.
Independent research initiative at the intersection of AI and Oncology.