The traditional pipeline for developing a new pharmaceutical compound is a decade-long endeavor. Bringing a single de novo therapeutic to market typically requires over ten years of research, billions of dollars in investment, and exhaustive clinical trials to establish safety and efficacy. As noted by Arash Keshavarzi Arshadi et al., this standard timeline is fundamentally incompatible with the immediate needs of global health crises, such as the COVID-19 pandemic, or acute conditions like sepsis and cardiac injury.
To bypass this latency, researchers are increasingly turning to computational drug repurposing. By identifying new therapeutic indications for existing, FDA-approved drugs, the scientific community can skip early-stage safety trials and move directly to efficacy testing. The goal is to transform a multi-year development cycle into a high-throughput screening process that yields results in days or weeks.
Relational Reasoning via Knowledge Graphs
One of the most powerful ways to identify repurposing candidates is to treat biological data as a massive, interconnected network. A Knowledge Graph (KG) represents entities—such as drugs, proteins, genes, and diseases—as nodes, with their biological interactions serving as edges.
The intuition here is that a "missing" edge between a drug and a disease might actually represent a hidden therapeutic potential. Charilaos Kanatsoulis et al. demonstrated this with TeX-Graph, a framework using coupled tensor-matrix factorization to identify Dexamethasone as a top candidate for COVID-19 before its clinical efficacy was widely recognized. Similarly, Xiangxiang Zeng et al. developed CoV-KGE, which mines millions of scientific publications to build a graph capable of predicting which approved drugs might inhibit viral activity.
By applying automated reasoning to these graphs, systems can move beyond simple keyword matching. Luis Tari et al. showed that by encoding fine-grained molecular interactions—such as whether a drug activates or inhibits a specific protein—models can achieve high recall in identifying treatments for complex diseases like cancer.
Matching Biological Signatures
Another approach focuses on the "fingerprint" of a disease. When a virus infects a cell or a heart attack occurs, the cell's internal signaling is "re-wired." Researchers can measure this change through transcriptomics—the study of which genes are being turned on or off.
The strategy, as described by Fuhai Li 0001 et al., is to find a drug that induces a biological signature that is the "inverse" of the disease signature. If a virus causes a specific set of inflammatory genes to spike, a drug that naturally suppresses those same genes might be an effective treatment. Nehmé El-Hachem et al. used this molecular fingerprinting to identify over 50 candidates for dampening the inflammatory response in severe COVID-19 cases.
This method is not limited to viruses. Liang Yu et al. used bioinformatics-driven repositioning to identify Linifanib—originally a cancer drug—as a potent treatment for sepsis by blocking inflammatory cell death. Similarly, Bingrui Li et al. combined transcriptomic fingerprints with chemical structure analysis to discover that the antipsychotic drug Pimozide could be repurposed for non-small cell lung cancer.
Deep Learning and Structural Simulation
While knowledge graphs and transcriptomics look at high-level relationships, deep learning allows researchers to simulate the physical interaction between a drug and a viral protein at the atomic level.
Haiping Zhang et al. utilized a Dense Fully Convolutional Neural Network (DFCNN) to rapidly screen millions of candidates against the 3C-like protease of SARS-CoV-2, completing in one day a process that typically takes weeks. To achieve even greater scale, the authors of Language models for the prediction of SARS-CoV-2 inhibitors pre-trained a BERT-based language model on 9.6 billion molecules, using high-performance computing to reduce pre-training time from days to hours.
These models act as a "virtual filter," narrowing down billions of possibilities to a handful of high-probability candidates. For instance, Kamonpan Sanachai et al. combined virtual screening with molecular dynamics to identify the cancer drug lapatinib as a potent inhibitor of the SARS-CoV-2 main protease.
Bridging the Gap in Emergencies
The ultimate value of these computational methods lies in their ability to function with incomplete data. During the early stages of an outbreak, comprehensive databases do not yet exist. Dokyoon Kim et al. developed a "complementary linkage" method that anchors a new, disconnected disease node into a massive "backbone" knowledge graph using only early-stage clinical reports and preprints.
By leveraging existing safety data and high-speed computational screening, these diverse approaches—ranging from tensor decomposition (Charilaos Kanatsoulis et al.) to multiscale interactome analysis (Haotian Cui)—provide a roadmap for responding to medical emergencies. They shift the paradigm from designing new molecules to intelligently searching the vast library of existing ones, significantly reducing the latency between the emergence of a threat and the deployment of a treatment.
Key Foundational Work
Deep Learning Based Drug Screening for Novel Coronavirus 2019-nCov (2020) — contributed 2 concept elements to this direction.
Repurpose Open Data to Discover Therapeutics for COVID-19 using Deep Learning (2020) — contributed 1 concept elements to this direction.
Integrative Transcriptome Analyses Empower the Anti-COVID-19 Drug Arsenal (2020) — contributed 1 concept elements to this direction.