9:00

9:10

Welcome

9:10

9:20

Cluster updates

R. Schulz (LEI)

Overview and Updates from ALICE

9:20

9:30

Cluster updates

P. van Vliet (LUMC)

Overview and Updates from SHARK

9:30

9:50

User Talks Session 1

A. Georgios (LUMC, SHARK)

Multi-Objective Dual Simplex-Mesh Based Deformable Image Registration for 3D Medical Images on the GPU

Transferring information between images with large anatomical differences is an open challenge in medical image analysis. Existing methods require extensive up-front parameter tuning and have difficulty capturing large deformations and content mismatches. Using a multi-objective optimization approach with a dual-dynamic grid transformation model has previously proven effective at overcoming these issues while also producing a diverse set of high-quality registrations for 2D images. We successfully introduce the first method for multi-objective deformable image registration for 3D images, based on a new 3D dual-dynamic grid transformation model and using the Real-Valued Gene-pool Optimal Mixing Evolutionary Algorithm (RV-GOMEA) parallelized on a GPU. Our proof-of-concept prototype shows promising results on a synthetic and clinical registration problem. This talk is based on a conference paper published at the SPIE’22 Medical Imaging - Image Processing conference.

9:50

10:10

User Talks Session 1

F. van Meer (LEI, ALICE)

fMRIPrep on ALICE

fMRIPrep is an application for the robust preprocessing of task-based and resting-state functional MRI. The pipeline adapts the preprocessing steps depending on the dataset and requires minimal user input and should thus provide results as good as possible independently of scanner make or scanning parameters, which aids the reproducibility of results. In this talk I will introduce fMRIPrep and talk about my experiences running it on the ALICE cluster.

10:10

10:30

User Talks Session 1

H. Mei (LUMC, SHARK)

Reproducible sequencing analysis pipeline on SHARK using WDL, Cromwell and Singularity

Sequencing Analysis Support Core (SASC) at LUMC is a core facility to support all LUMC departments for their Next Generation Sequencing (NGS) data analysis. Every year, hundreds of projects and TBs of NGS data are processed by the SASC team. Therefore, high throughput and reproducible data analysis pipelines are required in order to process this large amount of sequencing data using the LUMC SHARK cluster. SASC has worked with several pipeline frameworks in the past such as Snakemake and Galaxy. In this presentation, we will explain our current pipeline framework, the biowdl project (https://github.com/biowdl ), and its ecosystem to execute data analysis on the SHARK cluster.

10:30

10:50

User Talks Session 1

D. Chao (LEI, ALICE)

ALICE use cases by a biologist / bioinformatician

As a biologist from beginning, I have been exploring the use and incorporation of ALICE HPC in my daily work to support my own research and also bring the power to the whole research group. In my talk I will be sharing my experiences with some real use cases to illustrate

• Setting up Conda environments

• Build “easy to modify” single task slurm batch script

• Use Singularity for containerised apps

• Use .Net app"

10:50

11:55

Break

11:00

11:13

User Talks Session 2

J. Jia (LUMC, ALICE)

Prediction of lung CT scores of systemic sclerosis by cascaded regression neural networks

Visually scoring lung involvement in systemic sclerosis (SSc) from CT scans plays an important role in monitoring progression. We proposed an automatic scoring framework that consists of two cascaded deep regression neural networks. The first (3D) network aims to predict the craniocaudal position of five anatomically defined scoring levels on the 3D CT scans. The second (2D) network receives the resulting 2D axial slices and predicts the scores, which represent the extent of SSc disease.

11:13

11:26

User Talks Session 2

W. Jespers (LEI, ALICE)

Large scale high-throughput prediction of ligand binding affinities in early drug discovery

Obtaining binding affinities of a potential drug for its protein target is an essential step in early-stage drug discovery. However, this is a costly and time-consuming process, which can benefit greatly from compound prioritizing before actual experiments are conducted. Here, I will show recent efforts of our group to accurately predict binding affinities using computational models running on ALICE. These approaches run at a fraction of the time and cost it takes to experimentally determine experimental affinities whilst maintaining near experimental accuracy. In addition, I will look towards the future and present upcoming plans using our approaches on ALICE.

11:26

11:39

User Talks Session 2

P. Mody (LUMC, SHARK)

Streamlining deep learning on SHARK

Although deep learning libraries have matured over the years, there is still much streamlining that a deep learner needs to perform while working on clusters. This presentation shall discuss the various commands, tools and editors for use on LUMC's SHARK cluster that can make it easy for a deep learner to focus on model development rather than workflow optimization. The talk will follow a wiki page which will be made public.

11:39

11:52

User Talks Session 2

S. von Grebmer zu Wolfsthurn (LEI, ALICE)

Using ALICE to explore language processing in the brain: the application to EEG data 

A fundamental tool in language research is electroencephalography (EEG), which is a technique whereby brain potentials are measured directly from the scalp via electrodes to explore the underlying cognitive mechanism, for example of language processing. In our study, we collected EEG data from human participants to investigate the event-related brain potentials (ERPs) evoked in a simple grammatical error detection task. EEG data is typically collected in the form of voltage amplitudes from multiple scalp locations at a sampling interval of 2 ms for the duration of a particular task. Given the characteristic size of an EEG dataset with often more than 100 million observations, a computer cluster such as ALICE is an ideal platform to not only perform data manipulations, but also complex statistical analyses. EEG research has seen a recent shift from performing robust but simple analyses on averaged EEG data, to more sophisticated methods such as linear mixed effects models and generalized additive models at a single-trial basis. These more complex models require a large amount of computational power, which can be easily accommodated by ALICE. This talk outlines how we used ALICE to explore the neural correlates or language processing, applied to an EEG dataset.

11:52

12:05

User Talks Session 2

M. de Jong (LEI, ALICE)

Reaction-diffusion equations on GPUs with ALICE

The sinoatrial node in the heart is crucial because it generates action potentials that initiate our heartbeat. To better understand the electrophysiological processes and the propagation of this action potential, we resort to an in silico approach to model the tissues. The propagation of the action potential is modelled by a reaction-diffusion equation. Especially the 'reaction part' of this equation can be parallellized intuitively using GPUs. Because our local workstations contain merely a single graphics card, using an HPC-cluster like ALICE is essential to do multiple simulations. This talk will mostly focus on my motivation to use ALICE on the one hand, and my experiences in doing so on the other hand. These experiences also include some of my recommendations for new users.

12:05

12:18

User Talks Session 2

E. Tsingos (LEI, ALICE)

Modelling mechanical interactions with a fibrous extracellular matrix using a hybrid cellular Potts model and molecular dynamics framework.

Animal cells are surrounded by extracellular matrix (ECM) fibers spun into vast protein networks with non-trivial topology and bulk mechanics. Mechanical reciprocity between cells and the ECM plays a central role in fundamental processes such as cell migration and tissue morphogenesis. Current mathematical modelling approaches suffer from a key limitation: They either model ECM fiber networks in detail but simplify cells, or, vice-versa, they model cells in detail but average ECM structure. As a result, existing models do not fully capture key mechanobiological phenomena such as cell-induced strain-stiffening and contact guidance.

To address this gap, we introduce a cell-based model based on the cellular Potts formalism hybridized with a discrete ECM fiber model based on molecular dynamics. We model fibers as bead-spring chains with linear elastic potentials between consecutive beads and linear bending potentials between consecutive bead triplets. Fibers can be mechanically coupled via crosslinkers, and cells link to fibers via discrete focal adhesion-like sites.
We simulate a common experimental setup in the field of cellular biomechanics consisting of an isolated contractile cell embedded in a fiber network. We demonstrate how several parameters including fiber number, fiber crosslinks, and number of adhesion sites affect network strain and local fiber density, showing that the underlying dynamics of the model suffice to reproduce published experimental observations.

12:18

12:28

Break

12:28

12:58

Q&A with ALICE and SHARK Team

12:58

13:00

Closing