When a biological system shifts from health to disease, thousands of molecular variables change simultaneously. MicroRNAs (miRNAs), long non-coding RNAs (lncRNAs), and circular RNAs (circRNAs) may spike or drop in concentration, creating a dense web of statistical correlations. Traditional computational models excel at identifying these patterns, mapping out vast networks of non-coding RNA-disease associations.
However, as a comprehensive benchmark by Zhou Huang et al. reveals, standard association predictors suffer from a fundamental limitation: they cannot distinguish between an RNA that actively drives pathogenesis and one that is merely differentially expressed as a downstream consequence.
In therapeutics, this distinction is everything. If you design a drug to knock down a "passenger" RNA, the disease progresses untouched. If you target a "driver"—a causal agent—you disrupt the disease mechanism itself. The emerging frontier in computational biology focuses on moving past mere correlation to explicitly isolate these causal relationships.
Grounding Causality in Functional Experiments
To train a model to recognize causality, researchers first need a clean, supervised signal. What makes an RNA-disease relationship causal rather than associative? The answer lies in the nature of the underlying experimental evidence.
[Association Database] ───► Co-occurrence / Differential Expression (95% of data)
│
▼ (Strict Filtering Criteria)
[Causal Database] ───► Gain-of-Function / Loss-of-Function Validation
Yuanxu Gao et al. established that roughly 20% of known miRNA-disease associations possess true causal properties. To isolate this subset, the authors manually reviewed literature and filtered out weak associations using strict criteria. A link is only deemed causal if it is backed by direct functional validation:
- Gain-of-function experiments: Artificially overexpressing the RNA directly induces or accelerates the disease state.
- Loss-of-function experiments: Knocking out or silencing the RNA halts or reverses the disease progression.
Crucially, associations where an RNA merely enhances or reduces a drug's efficacy are excluded. This rigorous curation strategy was extended to other non-coding RNAs by Kaiwen Jia et al., who updated the LncRNADisease database to explicitly annotate causal links for lncRNAs and circRNAs. These curated datasets provide the ground-truth labels needed to move from association mapping to causal inference.
How Models Propagate and Infer Causality
Because functional experiments are labor-intensive, verified causal links are scarce. Computational models must therefore take these sparse "causal seeds" and infer new ones across biological networks. They achieve this by leveraging different dimensions of molecular similarity.
Network Label Propagation
In the MDCAP framework developed by Yuanxu Gao et al., known causal links act as source nodes on a heterogeneous network connecting miRNAs and diseases. A label propagation algorithm allows the "causal probability" to flow through the network graph. If an uncharacterized miRNA shares highly similar disease-association profiles with a known causal driver, the algorithm propagates the causal label to it, effectively utilizing the topology of the network to infer direct influence.
Sequence Similarity Metrics
Network topology alone can sometimes reflect shared downstream effects rather than shared causes. To introduce a more intrinsic biological prior, Zhou Huang et al. developed LE-MDCAP, which incorporates Levenshtein distance to measure the sequence similarity of miRNAs. The intuition here is structural: miRNAs with similar nucleotide sequences tend to bind to the same target messenger RNAs (mRNAs). By clustering miRNAs based on sequence-level similarity, the model can infer that if one miRNA causally disrupts a pathway, its sequence-similar counterparts are likely to possess the same causal capacity.
Deep Graph Embeddings and Imputation
Biological networks are notoriously noisy and incomplete, often suffering from research biases where well-studied diseases have many documented links and rare diseases have few. To prevent these gaps from misguiding causal inference, Qiong Zhou et al. introduced DNI-MDCAP.
[Incomplete Association Graph] ──► [Deep Graph Embedding] ──► [Imputed/Cleaned Graph] ──► [Causal Prediction]
This approach uses deep graph embeddings to perform network imputation, filling in missing background associations before attempting to predict causality. By smoothing out the underlying graph topology first, the semi-supervised learning framework can more accurately differentiate between direct causal pathways and indirect, non-causal correlations.
The Horizon of Causal Discovery
The transition from correlational biomarkers to causal networks is reshaping clinical applications. For example, Akram Yazdani et al. demonstrated that constructing transcriptomic-causal networks allows researchers to derive robust gene signatures that accurately stratify colorectal cancer patients for targeted therapies. Because these signatures capture actual drivers of treatment response rather than reactive downstream effects, they prove far more reproducible across independent patient cohorts.
Despite these advances, scaling causal discovery remains a major bottleneck. As highlighted in a systematic survey of biomedical pre-trained language models by Benyou Wang et al., even state-of-the-art text models still primarily extract surface-level correlations from medical literature. Teaching computational architectures to automatically recognize and reason about causality—rather than just cataloging co-occurrences—remains the critical next step toward fully automated therapeutic target identification.
Key Foundational Work
A Computational Model to Predict the Causal miRNAs for Diseases (2019) — contributed 4 concept elements to this direction.
Annotation and curation of the causality information in LncRNADisease (2020) — contributed 4 concept elements to this direction.
Benchmark of computational methods for predicting microRNA-disease associations (2019) — contributed 3 concept elements to this direction.