MIND: Multimodal Integration with Neighbourhood-aware Distributions

Abstract

Revised and tested implementation of MIND, replacing the initial 0.0.1 code on main. The previous implementation remains available at v1.0.0. First semver release. Addresses the SMARTbiomed/software-review checklist (Issue SMARTbiomed/software-review#2). All public API entry points from 0.0.1 are preserved. Added MIND.data.make_toy_dataset — pure-numpy synthetic generator that requires no network access; used in tests and the new quickstart tutorial. examples/reproduce_results.py — single-script reproduction CLI (python reproduce_results.py --dataset {synthetic,CCMA,CCLE,TCGA}). examples/quickstart_toy.ipynb — fast tutorial notebook driven by make_toy_dataset, geared at users applying MIND to their own data. MIND._validation — input-validation helpers; surfaced via MIND.MIND.__init__. Full pytest suite covering the model, data loaders, and validation. .github/workflows/ci.yml — automated CI on Linux + macOS, Python 3.10–3.12. docs/ — Sphinx skeleton (autodoc + napoleon) plus a written user guide (docs/user_guide.md) covering input format and hyperparameter tuning. .readthedocs.yaml, pyproject.toml, CITATION.cff, AUTHORS.md, CHANGELOG.md. download_if_missing flag on every load_* helper — allows fail-fast behaviour for users who don't want network calls. Changed Package layout: MIND_model.py → MIND/model.py + MIND/layers.py + MIND/_train.py. MIND_data.py → MIND/data.py. The public re-exports in MIND/__init__.py are unchanged, so from MIND import MIND and the existing from MIND import get_*, load_* imports keep working. Type hints added to every public function, method and attribute. NumPy-style docstrings written for every public symbol. MIND.data.get_data now uses os.path.join (was string concatenation that assumed a trailing /). README rewritten: jargon-free summary, fixed typos ("Sofrware", "User needs ot provide"), added "Citing" section, comparison-to-alternatives table, and a network-free quickstart. Fixed Stale device = 'cuda' if torch.cuda.is_available() else 'cpu' line in MIND_model.py that ran before import torch. Toy-dataset construction no longer leaves any patient missing from every modality (previously caused a divide-by-zero in the appearance denominator).