ABLeS Participant: Australian Proteome Analysis Facility, Macquarie University

Project title

Scalable Multi-Modal Spatial Omics Integration via a novel mathematical algorithm

Collaborators and funding

Australian Proteome Facility, Macquarie University.

Contact(s)

• Ignatius Pang, Australian Proteome Analysis Facility (APAF), ignatius.pang@mq.edu.au

Project description and aims

Scientific Context & Challenge

Spatial biology has accelerated from single-modality experiments to complex multi-omics assays that measure spatial gene expression, targeted antibody panels, and untargeted metabolite/lipid distributions within the same tissue specimen. Reconstructing three-dimensional tissue structures and harmonizing cross-technology experiments requires robust pairwise and multi-slice registration.

Current computational approaches face severe limitations:

1. Extrinsic Metric Sensitivity: Methods relying on Cartesian Euclidean coordinates (e.g., standard Fused Gromov–Wasserstein / PASTE) fail when slices experience non-isometric shear or non-linear elastic stretching during histological sectioning.

2. Artificial Margin Discarding: Partial optimal transport methods (e.g., PASTE2) achieve nominal accuracy by mathematically dropping 15–20% of spots, selectively pruning ambiguous border cells and damaged margins rather than solving the true registration problem.

3. Cross-Modality Incompatibility: Methods that require shared gene features cannot align spatial transcriptomics with mass spectrometry imaging or antibody-based proteomics.

Project Aims

• Aim 1: Topological Data Analysis and Optimal Transport: Develop, implement, and optimise our novel algorithm that is robust to different types of stresses and damage applied to the tissue samples, including non-linear warping and tissue lacerations.

• Aim 2: Rigorous Combinatorial Benchmarking on Pawsey Setonix: Deploy a reproducible Nextflow DSL2 pipeline on Setonix to evaluate SC-FGW against 12 state-of-the-art algorithms that operate on spatial coordinates and molecular feature matrices (PASTE, PASTE2, MOSCOT, GPSA, SLAT, STalign, etc.) across:
  • 13 biological datasets (Human DLPFC, Andersson Breast Cancer, Mouse Olfactory Bulb, Stereo-seq Whole Embryo MOSTA, MALDI brain serial sections, and CODEX colorectal cancer).
  • 4 synthetic stress-testing regimes (anisotropic shear, planar rotation, non-linear elastic deformation, and tissue tearing).
  • Two resolution tiers (Tier 1 downscaled grids and Tier 2 native resolution exceeding 2,000 spots per slice).
• Aim 3: Formal Mathematical Verification: Provide machine-checked formal verification of all underlying optimal transport theorems, metric properties, and convergence bounds in the Lean 4 interactive theorem prover, guaranteeing zero algorithmic failure states.

High-Performance Computing Justification

Native-scale spatial alignment problems require optimizing quadratic assignment problems over thousands of cells with non-convex manifold constraints. Individual full-scale tasks require up to 32 cores and 128 GB RAM to compute toplogical data analysis and optimal transport calculation iterations. The full multi-donor combinatorial benchmark requires 12,516 individual process executions (120,000–250,000 core-hours). Pawsey Setonix’s high-speed Lustre scratch filesystem and distributed Slurm scheduling via Nextflow DSL2 provide a viable infrastructure capable of completing this workload at scale.

How is ABLeS supporting this work?

This work is supported through the production bioinformatics scheme provided by ABLeS. The supports includes storage and compute resources.

Expected outputs enabled by participation in ABLeS

Participation in ABLeS on Pawsey Setonix directly enables the generation, publication, and FAIR dissemination of the following research outputs:

1. Open-Source Scientific Software:
   • ** The Python Package**: Production-grade, vectorized implementation supporting PyTorch and POT backends, released under an open-source MIT license on PyPI and GitHub.
   • Containerized Nextflow DSL2 Pipeline: An automated, portable, and nf-core compliant workflow containerized with Singularity and Docker, enabling Australian and international researchers to run robust spatial omics alignment workflows on institutional clusters or cloud environments.
2. FAIR-Compliant Benchmark Data & Artifacts:
   • Standardized Spatial Benchmark Corpus: Curated, preprocessed AnnData objects across 13 tissue systems with standardized cell annotations and synthetic distortion sets deposited on public data repositories (NCBI GEO and BioStudies).
   • Comprehensive Evaluation Archive: Master benchmark records (15,000 evaluated configurations), CSV performance summaries, and runtime Pareto frontiers deposited with permanent, citable DOIs on Zenodo in compliance with FAIR data principles.
3. High-Impact Scientific Publications:
   • A lead manuscript submitted to a peer-reviewed computational biology journal (e.g. Bioinformatics), citing Australian BioCommons and Pawsey Setonix support.
   • A companion mathematical formalization repository with machine-checked proofs in Lean 4 verifying the novel algorithm in spatial data science.
4. Australian Research Capability Building:
   • Workflows, tutorials, and documentation will be shared with the Australian BioCommons community through collaborative webinars, computational biology workshops, and BioCommons workflow registries, empowering Australian researchers to analyze complex spatial multi-omics datasets with state-of-the-art national HPC tools.

These details have been provided by project members at project initiation. For more information on the project, please consult the contact(s) or project links above.