DevoGraph

Introduction

• DevoGraph is a GSoC 2025 project under the administration of INCF and DevoWorm. Our main goal is to provide examples and components that utlize (Temporal/Directed/...) Graph Neural Networks to model the developmental process of C. elegans.

Developers

• GSoC contributors: Jiahang Li, Wataru Kawakami, Himansuhu Chougule, Pakhi Banchalia, Mehul Arora, Pakhi Banchalia, Sushmanth Reddy, Jayadratha Gayen, Lalith Bharadwaj Baru.
• Mentors: Bradly Alicea, Jesse Parent, Himansuhu Chougule, Mehul Arora
• Additional contributors: Longhui Jiang, Gautham Krishnan

Contributions

Jiahang Li

• Design a KNN-based method constructing temporal graphs. The method is implemented in ./devograph/datasets/datasets.py. These temporal graphs are based on 3d positions of cell centroids and mimic cell developmental process of C. elegans. Each node represents a cell at a certain frame, and edges at the same frame connect neighbors according to KNN while edges across different frames connect mother and daughter cells. Please refer to ./stage_2/stage_2.ipynb to check more details.
• Refactor codes of constructing directed graphs initially implemented by cell-track-gnn. The re-implementation is in ./devograph/datasets/datasets1.py. This method gives each edge an direction implying the relationship between mother and daughter cells.
• Refactor codes of a directed GNN initially implemented by cell-track-gnn. The re-implementation is in ./devograph/models/ct.py. The GNN is based on directed graphs and incorporates information of nodes and edges to aggregate messages.
• Both of re-implementations above abstract the core logic, remove redundant and unrelated codes and unnecessary third-party frameworks, and finally provide easy-to-use APIs.
• Design the whole pipeline of DevoGraph presented in ./miscellaneous/GSoC 2022 22.1.pdf.
• Assign tasks to other participants.

Wataru Kawakami

• worked on image processing issues (Stage 1).

Longhui Jiang

• Refactor codes of pre-processing 2-D images(frames of videos) and converting them into location information of cells stored in .csv files (Stage 1). The re-implementation is based on cell-track-gnn.

Sushmanth Reddy

• incorporating DevoLearn models into DevoGraph, particularly for Stage 1.

Himanshu Chougule

• developed a customized RNN for creating graph embeddings, building out Topological Data Analysis tools and infrastructure.

Mehul Arora

• developed a Hypergraph model (HNN) of the embryo.

Lalith Baru

• extended the Hypergraph model (HNN) to include Neural Developmental Programs (NDP).

Pakhi Banchalia

• developed applications of k-mapper for Topological Data Analysis and Neural Developmental Programs.

Jayadrath Gayen

• developed the DevoTG (temporal graphs) approach to D-GNNs.