Harshini Suresha

PathSymetic

Neuro-Symbolic Causal Hypothesis Generation for Mechanistic Pathway Discovery

Year2025
LocationAAAI 2026 workshops
TagsGNN, Knowledge graphs, Causal inference, Reactome

Knowledge graphs, pathway ontologies, causal structure learning and GNNs combined to generate interpretable pathway hypotheses from genomic data; AUPR 0.81, Precision@10 of 0.94 and Hit@20 of 99.6%.

Mar 2025 – Aug 2025 at AAAI 2026 workshops.

Workflow figureOpen full size ↗
Figure for PathSymetic: Neuro-Symbolic Causal Hypothesis Generation for Mechanistic Pathway Discovery
Methods
  1. 01
    Knowledge integration

    Combined biological knowledge graphs and pathway ontologies into gene–gene, gene–pathway and pathway–pathway relationships.

    KEGG, Reactome, BioGRID
  2. 02
    Causal graphs

    Built gene–pathway causal graphs from multi-omics evidence with ontology-based constraints.

    Causal structure learning, Multi-omics
  3. 03
    Hybrid inference

    Paired graph neural networks and LLM-derived embeddings with rule-based symbolic reasoning to propose mechanistic hypotheses.

    Graph neural networks, LLM embeddings, Symbolic reasoning
  4. 04
    Ranking hypotheses

    Prioritised pathways with evidence-weighted ranking that combines symbolic constraints and learned representations.

    Evidence-weighted ranking
  5. 05
    Evaluation

    Benchmarked against GNN and pathway-enrichment baselines: AUPR 0.81, F1 0.77, Precision@10 0.94, Hit@10 / Hit@20 of 98.4% / 99.6%.

    AUPR, Precision@k, Hit@k
← Back