Multimodal deep learning connects polygenic metabolic disease risk to adipocyte regulatory circuits

Code for the analyses in Multimodal deep learning connects polygenic metabolic disease risk to adipocyte regulatory circuits, a variant-to-function (V2F) study of adipogenesis and metabolic disease. Most scripts run in protected environments against data that is not openly available, so this repository is primarily a durable record of the code rather than a runnable pipeline.

The study. Adipose-derived mesenchymal stem cells (AMSCs) from 37 donors, differentiating towards mature adipocytes, profiled across 269 samples for gene expression, LipocyteProfiler morphology, polygenic risk scores and cardiometabolic variants. The four analysis stages follow one thread from that dataset to a mechanism:

1. Integrate and generate hypotheses. A variational autoencoder integrates the modalities; in silico perturbations link insulin-resistance polygenic risk to SHC2 expression, and the C5orf67 variant rs3843467 to SHC2, MAP3K1 and SETD9.
2. Prioritise variants. Colocalization across four glycaemic traits and fine mapping of T2D and fasting insulin over C5orf67 narrow the signal to rs256903 and a set of variants in near-complete linkage with it.
3. Predict what the variants do. Sequence-to-function models place them in open chromatin next to AP-1 and SP/KLF motifs, and map their effect on MAP3K1 to adipose contexts.
4. Validate independently. A single-cell village of differentiating AMSCs confirms the motif enrichment and finds context-dependent eQTLs under glucose stress.

Where to find each part of the paper

Paper	Code
Fig. 1, Suppl. Figs. 1–3 — VAE integration and SHAP feature importance	1_data_integration/
Fig. 2a–g, Suppl. Figs. 4–5 — SHC2 as a convergence point, genotype–expression follow-up	1_data_integration/
Fig. 2h–j, Suppl. Figs. 6–7 — colocalization and fine mapping	2_colocalization_and_finemapping/
Fig. 3a — ChromBPNet accessibility and contribution scores	3_.../ChromBPNet/
Fig. 3b, Fig. 4, Suppl. Figs. 8–17 and 21 — Borzoi and AlphaGenome, and their cross-comparison	3_.../AlphaGenome_Borzoi/
Suppl. Figs. 18–20 — single-cell eQTLs and caQTLs	4_adipogenic_village_analyses/

Repository layout

1_data_integration/ — AMSC_MOVE.ipynb, the main notebook of the project: preprocessing, integration with MOVE, SHAP feature importance, the in silico perturbation analyses, and the downstream SHC2 / MAP3K1 genotype–expression analyses.

2_colocalization_and_finemapping/ — each stage reads from dat/ and writes to its own interim/ and results/, none of which are tracked.

• 2a_colocalization — SharePro_signed.ipynb builds the signed LD matrix, harmonises the four GWAS onto it, and runs SharePro across T2D, fasting insulin, fasting glucose and 2-hour glucose (res_signed is the three-trait model, res_signed_2h adds 2-hour glucose).
• 2b_finemapping — CARMA_T2D_signed.R fine-maps T2D and fasting insulin with CARMA; CARMA_seed_stability.R reruns it across 20 seeds and reports which conclusions survive, since CARMA's stochastic search sets no seed of its own and this locus holds many perfectly collinear variants.
• dat — what 2a writes and 2b reads: per-trait effect sizes and standard errors, the signed European LD matrix around rs3843467, and variant_order.tsv, its row index, which records the allele every sign and every beta refers to. The matrix carries no labels of its own, so those two belong together. variant_list.txt is the only input fixed in advance. The matrix comes from plink --r square on 1000 Genomes EUR genotypes rather than from LDlink, whose endpoints return only unsigned R² or D′ — SharePro and CARMA need the signed correlation to tell whether two correlated effects reinforce or cancel.

3_sequence_to_function_modeling/

• ChromBPNet — ChromBPNet-C5orf67.ipynb: bias-corrected ATAC-seq profile prediction on bulk visceral AMSCs, DeepLIFT/DeepSHAP contribution scores and in silico substitutions. The variant-centred regions and SNP list it takes as inputs are built in auxilliary_files/.
• AlphaGenome_Borzoi — C5orf67_MAP3K1_Borzoi_and_AlphaGenome.ipynb, both long-range models in one kernel so their results can be compared directly: Borzoi predictions of MAP3K1 across tracks with the hypergeometric test for adipose enrichment, AlphaGenome variant scoring, ISM and haplotype editing, and a cross-model section on why the two disagree on direction. The variant list is not typed in — it is derived at run time from the CARMA and SharePro outputs.

4_adipogenic_village_analyses/ — Adipogenic_village_analyses.v1.ipynb: the AMSC cell village 10X-Multiome analyses as presented in the manuscript — motif enrichment in peaks upregulated in the adipogenic subpopulation, and the eQTL / caQTL analyses around C5orf67.

configs/MOVE/ — the data, experiment and tasks YAML files used to run the MOVE models.

Running the analyses

This repository tracks the analysis code, the MOVE configuration and this Readme, and nothing else. Data and outputs are git-ignored — the dat/, data/ and auxilliary_files/ folders each stage reads from, and the interim/ and results/ folders it writes — so a fresh clone holds no inputs, figures or tables. Each stage rebuilds what the one before it produced, and stage 3 reads stage 2's output, so run them in order, or point a stage at a folder that already holds what it needs.

Also kept out of version control, and so present only in a working copy: the manuscript, local copies of the SharePro and CARMA packages, the ChromBPNet tracks exported for the WashU Epigenome Browser in Fig. 3a, and the superseded results from before the LD matrix carried a sign.

The AlphaGenome notebook needs an API key. Put a line reading ALPHA_GENOME_API_KEY=your-key in a .env file at the repository root; VS Code loads it into the kernel automatically, and .env is git-ignored. The notebook's Running locally section covers the rest.

Donor privacy and reproducibility

No donor or patient identifiers belong in this repository. The AMSC data is individual-level and stays in the protected environment where the notebooks are run. The code refers to donors only through placeholders such as <DONOR_ID> and through column names such as Patient_ID; cloud paths are likewise reduced to placeholders. Everything tracked here is variant-level. This matters most in 1_data_integration and 4_adipogenic_village_analyses, the two stages that touch donor-level data directly.

.gitignore blocks identifier files, MOVE working directories, access-controlled source data, API keys and all regenerable output. Before committing, check that nothing slipped through:

git add -An | sed 's/^add //' | tr -d "'" | grep -iE 'donor|patient|private|sumstat|\.vcf'

Strip notebook outputs before committing — they can carry identifiers, absolute paths and intermediate data, and they make diffs unreadable:

jupyter nbconvert --clear-output --inplace path/to/notebook.ipynb
# or once, to have it happen on every commit:
pip install nbstripout && nbstripout --install

Data availability

The AMSC data is not openly available. Transcriptomics from the LipocyteProfiler study are on GEO under GSE184089, and the high-content imaging data in the Cell Painting Gallery on the AWS Registry of Open Data under cpg0011. GWAS summary statistics used for colocalization and fine mapping:

• Fasting insulin — GCST90002238
• Two-hour glucose — GCST90002227
• Fasting glucose — GCST90002232
• Type 2 diabetes — DIAMANTE

Related code

MOVE is available as the move-dl pip package. Code and configuration for creating and running MOVE models are also at RasmussenLab/VAEs_for_biomedical_data_integration, with a tutorial notebook covering the main findings.