1. Research Problem & Scientific Motivation
In computational ecology and systems biology, ecological food webs represent complex directed graphs where vertices denote biological taxa (species, functional groups) and edges represent energy-flow interactions (predation, parasitism, mutualism).
Traditional tabular machine learning and Convolutional Neural Networks (CNNs) assume Euclidean geometric structures—fixed grid representations with spatial locality. However, food webs are inherently non-Euclidean: they exhibit variable node degree distributions, arbitrary graph sizing, and permutation invariance. When an apex predator or keystone species suffers decline, energy dissipation cascades across topological pathways that linear regression models cannot predict.
2. Methodology & Mathematical Formulation
To capture these dependencies, we investigated two principal Graph Neural Network architectures implemented in PyTorch Geometric (PyG):
A. GraphSAGE (Sample and Aggregate)
Rather than computing full graph Laplacians (which scale quadratically and require transductive retraining), GraphSAGE learns inductive neighborhood aggregation functions:
h_v^(k) = σ(W · Concat(h_v^(k-1), AGGREGATE({h_u^(k-1), ∀u ∈ N(v)})))
By sampling a fixed size of 1-hop and 2-hop ecological neighbors, GraphSAGE generalizes to previously unseen marine ecosystems without re-evaluating the full adjacency matrix.
B. Graph Attention Networks (GAT)
Not all ecological interactions carry equivalent energetic consequence. GAT introduces multi-head self-attention coefficients $\alpha_{ij}$ to weight the relative importance of connected prey and predator nodes:
α_ij = softmax_j(LeakyReLU(a^T [W h_i || W h_j]))
3. The Literature Extraction Bottleneck & Multi-LLM Ensembles
A fundamental barrier in biological graph modeling is data sparsity: empirical trophic interaction observations reside locked inside unstructured scientific monographs and PDF papers published over the past 40 years.
Multi-LLM Synthesis Pipeline
We engineered an automated data extraction and validation pipeline using a multi-LLM ensemble combining Gemini, DeepSeek, Anthropic, and OpenAI:
- Document Parsing: Chunking academic biological papers and extracting textual interaction tables.
- Schema Enforcement: Forcing structured JSON edge lists (Source Species, Target Species, Biomass Flow, Interaction Type, Confidence Score).
- Consensus Voting: Multi-model cross-verification to prevent hallucinated predator-prey links.
4. Key Results & Research Takeaways
This work proved that combining relational graph architectures with deterministic LLM extraction pipelines bridges the gap between unstructured scientific literature and high-accuracy predictive modeling.