Institute
General information about the institute, such as our mission statement, organizational structure, staff directory, history, directions, etc.

See More
Research
Scientific profile with all research groups, topics, collaborations, as well as columns on research at the institute.

See More
News
News and press releases about the institute, as well as a media archive.

See More
- News Overview
- Press Releases
Events
Overview of all events around the institute, such as talks, seminars, lectures, workshops, conferences and public events.

See More
Publications
Overview of all scientific publications of the institute, as well as our preprint and software repositories.

See More
Career
Information on open positions at the institute, benefits of working with us, graduate school, and postdoctoral supervision.

See More

Events
Graph Neural Networks Mini Meeting at MPI MiS

MPI für Mathematik in den Naturwissenschaften Leipzig (Leipzig)

Inselstr. 22
04103 Leipzig
Germany

G3 10 (Lecture hall)

conference

23.06.23

Graph Neural Networks Mini Meeting at MPI MiS

Overview

Graph Neural Networks (GNNs) are powerful models for processing graph structured data. These models connect with topics such as geometric deep learning, geometry of data, spectral graph theory, network science, topology, etc. GNNs find a diversity of applications, in areas where data can be associated with graphs, such as molecules and drug discovery. In this mini meeting we plan a few talks on theoretical aspects, practical aspects, and application areas of GNNs, with space for discussions.

Goals

1) Theory of GNNs, current topics and open problems. E.g., expressive power, rewiring, pre-coloring, higher order graphs, simplicial complexes.

2) Current and future applications of GNNs. E.g., chemistry, chemical reaction networks, drug discovery. What are currently the main challenges where GNNs can contribute.

3) Further topics: GNN variants, transformers, optimization, infinite limits

Registration

This will be an in-person workshop. If you are interested in participating, please register using the online form. Registration open till June 16, 2023.

Speakers

Pradeep Kr. Banerjee

MPI MiS

Nithya Bhasker

National Center for Tumor Diseases (NCT/UCC) Dresden

Tyler Derr

Vanderbilt University

Kedar Karhadkar

UCLA

Yunchao Lance Liu

Vanderbilt University

Jens Meiler

Leipzig University / Vanderbilt University

Guillermo Restrepo

MPI MiS

Stefanie Speidel

NCT Dresden

Yuguang Wang

SJTU

Bingxin Zhou

Shanghai Jiao Tong University

Program

09:00 - 09:15	Guido Montúfar (Max Planck Institute for Mathematics in the Sciences) Introduction
09:15 - 09:45	Yuguang Wang (SJTU) Applied harmonic analysis and particle dynamics for designing neural message passing on graphs Graph representation learning has broad applications applications from recommendation systems to drug and protein designs. In this talk, I will talk about using harmonic analysis and particle systems to design useful neural message passing with theoretically guaranteed separability and efficient computation. These message passings are proved to have strictly positive lower bounded Dirichlet energy and thus to circumvent the oversmoothing problem appearing in many spatial GNNs, when the node features are indistinguishable as the network deepens.
09:45 - 10:00	Coffee Break
10:00 - 10:30	Pradeep Kr. Banerjee (MPI MiS) On Oversquashing in Message Passing Graph Neural Networks The quality of signal propagation in message passing graph neural networks strongly influences their expressivity. In particular, for prediction tasks relying on long-range interactions, recursive aggregation of node features can lead to an undesired phenomenon called "oversquashing". We present a framework for analyzing oversquashing and propose a graph rewiring algorithm aimed at alleviating the same. Our algorithm employs a random local edge flip primitive motivated by an expander graph construction. We compare the spectral expansion properties of our algorithm with that of an existing curvature-based non-local rewiring strategy. Synthetic experiments show that while our algorithm in general has a slower rate of expansion, it is overall computationally cheaper, preserves the node degrees exactly and never disconnects the graph.
10:30 - 11:00	Kedar Karhadkar (Max Planck Institute for Mathematics in the Sciences) First-order spectral rewiring for addressing oversquashing in GNNs Graph neural networks (GNNs) are able to leverage the structure of graph data by passing messages along the edges of the graph. While this allows GNNs to learn features depending on the graph structure, for certain graph topologies it leads to inefficient information propagation and a problem known as oversquashing. This has recently been linked with the curvature and spectral gap of the graph. On the other hand, adding edges to the message-passing graph can lead to increasingly similar node representations and a problem known as oversmoothing. We propose a computationally efficient algorithm that prevents oversquashing by systematically adding edges to the graph based on spectral expansion. We combine this with a relational architecture, which lets the GNN preserve the original graph structure and provably prevents oversmoothing. We find experimentally that our algorithm outperforms existing graph rewiring methods in several graph classification tasks.
11:00 - 13:00	Lunch Break
13:00 - 13:30	Bingxin Zhou (Shanghai Jiao Tong University) Mutational Effect Prediction and Directed Evolution with Geometric Deep Learning The analysis of protein sequence–function relationships provides valuable insights for enzyme engineering to develop new or enhanced functions. Predicting the effects of point mutations in proteins allow researchers to dissect how changes in the amino acid sequence can impact the protein’s structure, stability, function, and interactions with other molecules. Despite the impressive performance exhibited by large self-supervised language models in protein sequence training and inference, their inherent incapability to account for spatial characteristics of protein structures—a key factor in understanding protein folding stability and intramolecular interactions—poses a significant limitation. In response, we introduce a lightweight gemoetric learning framework designed to analyze the microenvironment of wild-type proteins and suggest practical higher-order mutations tailored to a user-specific protein and its function of interest. The topological analysis of protein structures implemented in our framework alleviates the computational demands of model training and notably enhances the performance of sequence-based language models, particularly in zero-shot variant effect prediction tasks. Our framework exhibits substantial improvements in engineering candidate proteins across diverse assays and taxa, offering a promising tool for the advancement of protein engineering.
13:30 - 14:00	Nithya Bhasker (National Center for Tumor Diseases (NCT/UCC) Dresden) Application of Graph Neural Networks to surgical complication prediction Most complex tumour resections are associated with severe complications. These complications cause delays in adjuvant therapy resulting in worsened treatment response or death. Machine learning-based prediction of postoperative patient outcomes, before the surgery, enables individualised treatment planning and improved postoperative patient management. In this talk, I will describe such a surgical complication prediction framework and present the challenges associated with its practical implementation. Additionally, with the help of a case study, I will demonstrate the use of Graph Neural Networks for complication prediction and the challenges thereof.
14:00 - 14:30	Guillermo Restrepo (MPI MiS) Graph neural networks in chemistry, some of their applications and challenges Subjects of interest in chemistry involve estimating substance properties; given a target property, determining the substances holding the properties and finding ways to synthesize substances. Traditionally chemists have encoded substances, for the purposes of computation and prediction, into strings of characters, some of them Boolean, some others as integer vectors and often as alphanumeric characters. These machine readable substance representations were developed to avoid encoding the typical representation of substances used by chemists on a daily basis, namely the molecular structure(s) associated to the substances. Graph neural networks (GNNs) present a unique opportunity to overcome the problem of connecting molecular structures with estimations through string representations. GNNs allow for operating directly on molecular structures, since they are based on a graph representation of the input object and molecular structures correspond to graphs. Therefore, GNNs have begun to find important applications in chemistry. I will discuss some of these applications. Furthermore, I will challenge GNNs by shifting their input object. Instead of having graphs (molecules) as input, what happens if we now have high-order structures such as hypergraphs, which have been recognised as suitable models to understand and estimate the future of the chemical space — the collection of molecules and reactions to which GNNs have proven valuable? Can GNNs not only model and estimate substance and reaction properties but also predict the evolution of the chemical space?
14:30 - 14:45	Coffee Break
14:45 - 15:15	Yunchao Lance Liu (Vanderbilt University), Jens Meiler (Leipzig University/ Vanderbilt University) Interpretable Chirality-Aware Graph Neural Network for Quantitative Structure Activity Relationship Modeling in Drug Discovery In computer-aided drug discovery, quantitative structure activity relation models are trained to predict biological activity from chemical structure. Despite the recent success of applying graph neural networks (GNNs) to this task, important chemical information such as molecular chirality is ignored. To fill this crucial gap, we propose Molecular-Kernel Graph Neural Network (MolKGNN) for molecular representation learning, which features SE(3)-/conformation invariance, chirality-awareness, and interpretability. For our MolKGNN, we first design a molecular graph convolution to capture the chemical pattern by comparing the atom’s similarity with the learnable molecular kernels. Furthermore, we propagate the similarity score to capture the higher-order chemical pattern. To assess the method, we conduct a comprehensive evaluation with nine well-curated datasets spanning numerous important drug targets that feature realistic high class imbalance and it demonstrates the superiority of MolKGNN over other GNNs in Computer-Aided Drug Discovery (CADD). Meanwhile, the learned kernels identify patterns that agree with domain knowledge, confirming the pragmatic interpretability of this approach. Our codes are publicly available at https://github.com/meilerlab/MolKGNN.
15:15 - 15:45	Tyler Derr (Vanderbilt University) Towards Data-Centric Graph Learning for Real-World Applications In today's data-driven world, graphs play a crucial role in representing a significant portion of big data across diverse domains. However, real-world graph data often presents challenges in terms of quality. This talk emphasizes the need for data-centric graph learning, which has been overlooked in favor of model-centric approaches. We showcase recent advancements in recommender systems, neuroimaging, and drug discovery, highlighting the importance of data-centric approaches to address data quality challenges across diverse domains. By tackling data quality issues and optimizing graph datasets, we can significantly enhance the performance of graph neural networks in real-world applications. The talk concludes with insights into future directions and potential advancements in the field.

Graph Neural Networks Mini Meeting at MPI MiS

Speakers

Pradeep Kr. Banerjee

Nithya Bhasker

Tyler Derr

Kedar Karhadkar

Yunchao Lance Liu

Jens Meiler

Guillermo Restrepo

Stefanie Speidel

Yuguang Wang

Bingxin Zhou

Program

Scientific Organizers

Guido Montúfar

Pradeep Kr. Banerjee

Kedar Karhadkar

Administrative Contact

Katharina Matschke