Graph Neural Networks For Biological Knowledge Discovery
Résumé
Background: In the era of modern biology, the exponential growth of biological data calls for new
methods that can aggregate and extract meaningful insights from this large accumulation. Biological
data is inherently complex and multimodal, making it well-suited for graph-based representation[1].
This work explores the application of graph neural networks (GNNs) to the task of biological knowledge
discovery using multimodal biological graphs.
Results: GNNs have shown success in learning meaningful representations from graph- structured
data, making them a natural fit for tackling challenges in the biological domain[2]. We leverage the
expressive power of GNNs to address knowledge discovery through the task of link prediction, where
the model is trained to rank the probability of an existing link’s existence against negative samples.
Furthermore, strategies for leveraging additional modalities, such as biological sequences from both
protein and genes through the use of Large Language Models are investigated to enhance the
robustness and performance of the GNN model. The proposed methods are benchmarked on real-
world datasets with a focus on Japanese Rice (Oryza Sativa sub. Japonica).
While GNNs in their naive configuration do not outperform traditional embedding-based techniques
such as TransE[3], DistMult[4] or ComplEx[5], they are far more memory efficient. The addition of
node features extracted from language models shows that this gap can be greatly reduced.
Conclusion: In order to discover hidden knowledge in the graph, GNNs are trained to predict missing
links between biological entities. We showcase the potential of this method to capture relations
between multimodal data on real-world datasets from species such as the Japanese Rice. Additionally,
we explore the potential of large language models in encoding biological sequence information for
downstream GNN processing, showing that through the usage of adequate node features, GNNs can
compete with traditional geometric models. Ultimately, this contribution to the field of network
biology aims to aid in the discovery of novel biological insights and the advancement of our
understanding of complex biological systems.
Origine | Fichiers produits par l'(les) auteur(s) |
---|