Integration Hub

Everything in this section supports the phased escalation strategy documented in the Integration Plan (Nov 2025). Use it as the connective tissue between per-modality preprocessing, harmonization, and experiment execution.

Overview

This hub provides end-to-end guidance for integrating genetics, brain, and behavioral data using foundation models.

📋 Key Resources

• Integration Strategy — High-level playbook: covariates to regress, projection dims, escalation triggers
• Design Patterns — Late fusion → two-tower → MoT → BOM escalation logic
• Multimodal Architectures — Detailed patterns from BAGEL, MoT, M3FM, Me-LLaMA, TITAN
• Embedding Policies — Naming conventions and PCA dimensionality guidelines
• Benchmarks — Prior benchmark targets to compare against

🔬 Analysis Recipes

Copy-ready runbooks for common integration tasks:

• CCA + Permutation — Test gene-brain associations before heavy fusion
• Prediction Baselines — Gene-only vs Brain-only vs Late fusion
• Partial Correlations — Control for covariates with logistic regression

🧬🧠 Modality Features

Concrete instructions for extracting and harmonizing features:

• Genomics — Genetics embeddings, RC-equivariance, gene attribution
• sMRI — FreeSurfer ROIs, PCA compression, site harmonization
• fMRI — Functional connectivity, BrainLM/SwiFT embeddings, preprocessing

🎨 Integration Cards

Comprehensive multimodal fusion guidance:

• Ensemble Integration — Model stacking, averaging, meta-learning
• Oncology Multimodal Review — Early/intermediate/late fusion taxonomy
• Multimodal FM Patterns — Architectural patterns from state-of-the-art FMs

Quick Start

Before running any analysis, grab the relevant strategy IDs and log them with your experiment configs:

# Show sMRI baseline recipe
python scripts/manage_kb.py ops strategy smri_free_surfer_pca512_v1

# Inspect harmonization metadata (e.g., MURD)
python scripts/manage_kb.py ops harmonization murd_t1_t2

# Show rs-fMRI preprocessing stack
python scripts/manage_kb.py ops strategy rsfmri_swift_segments_v1

This keeps downstream reports auditable even when raw datasets (e.g., UKB) cannot be shared.

Integration Phases

We follow a phased escalation strategy to avoid premature complexity:

Phase	Status	Pattern	Trigger	Documentation
Phase 1	✅ Active	Late Fusion	Baseline	Integration Plan
Phase 2	🚧 Prep	Two-Tower Contrastive	CCA p<0.001, ΔAUROC>5%	Integration Plan
Phase 3	⏳ Future	Unified Multimodal (MoT/BAGEL/LLM-Bridge)	ΔAUROC>10%, cross-modal reasoning	Integration Plan

Navigation Guide

For Late Fusion Workflows (Phase 1)

1. Read Integration Strategy
2. Pick analysis recipe: CCA, Prediction, or Partial Correlations
3. Extract features: Genomics, sMRI, fMRI
4. Review Ensemble Integration card for stacking strategies
5. Run analysis with logged strategy IDs

For Multimodal Architecture Design (Phase 2+)

1. Read Design Patterns for escalation logic
2. Study Multimodal Architectures for BAGEL/MoT/M3FM/Me-LLaMA/TITAN patterns
3. Review Multimodal FM Patterns integration card
4. Consult Oncology Multimodal Review for fusion taxonomy
5. Check Integration Plan decision table for recommended pattern

For Adding New Integration Strategies

1. Start from Integration card template
2. Review existing cards for structure and style
3. Document mechanics, use cases, caveats, and BOM integration
4. Add to models/integrations/ directory
5. Update mkdocs.yml navigation

Key Principles

✅ Late fusion first — Preserve modality-specific signal under heterogeneous semantics
✅ Unimodal baselines — Establish gene-only and brain-only performance before multimodal claims
✅ Covariate control — Z-score + residualize vs age/sex/site before interpreting effects
✅ Reproducibility — Log embedding strategy IDs, harmonization methods, CV folds
✅ Phased escalation — Only escalate when data and compute justify the complexity

Read full integration plan →