Idea: genetics-informed next steps for Delphi (rediscovery score, audit workflow, drug-target lift, debiasing, multi-ancestry)

[AR-Shicheng]

Hi gerstung-lab/Delphi team —

Thank you for releasing Delphi — it has become a cornerstone for modeling lifetime disease trajectories with generative transformers, and the open training/evaluation code has made serious follow-up work possible for many groups.

I recently read a related medRxiv paper from Maik Pietzner's group (Zeng, Kohleick, Beuchel, et al.; senior author Maik Pietzner):
"Genetic decoding reveals druggable biology implicitly learned by a medical-history foundation model"
https://www.medrxiv.org/content/10.64898/2026.07.28.26359117v1

They re-implement Delphi-2M-style medical-history modeling, then use human genetics to explain what the embeddings implicitly learned (including lipid and epithelial-alarmin / asthma biology, and some approved drug targets).

Given how much gerstung-lab/Delphi already contributes, a few ideas that might be interesting to explore:

1) Genetic rediscovery score next to checkpoint AUC

When comparing Delphi checkpoints, it could be interesting to report a simple "genetic rediscovery" score next to disease AUC — i.e., how well patient embeddings recover known lipid/asthma loci.

1. Export mean-pooled final-layer patient embeddings (as in Zeng/Pietzner et al.).
2. Run embedding GWAS + LDSC h² on a UKB genetic subset.
3. Score recovery of a small fixed sentinel set (e.g. LPA, PCSK9, APOE, alarmin/asthma loci).
4. Optionally use that score for model selection alongside AUC.

2) Minimal embeddings → genetics audit workflow

A minimal public workflow could help others: embeddings + genotypes in → h², loci, genes, and drug coloc out (the paper's analysis code is not public yet).

1. Take N×D embeddings and genotypes as inputs (plus a tiny synthetic demo if needed).
2. Run REGENIE → LDSC → SuSiE → FLAMES/PoPS → drug coloc.
3. Emit simple tables: h², sentinels, genes, drugs vs ICD-10 baseline.
4. Keep a one-page README so newcomers do not need the full paper.

3) Drug-target recovery lift alongside AUC

When comparing checkpoints, it may be useful to also report how many approved drug targets embedding GWAS recovers beyond single ICD-10 GWAS (the paper found a positive gap).

1. For each checkpoint, run embedding GWAS + drug coloc.
2. Run the same coloc pipeline on ICD-10 diagnosis GWAS.
3. Report Δdrugs = drugs found from embeddings but not ICD-10.
4. Compare models by AUC and Δdrugs together.

4) Debias embeddings using participation/education genetic axes

One idea is a light debiasing step: down-weight embedding directions that align with UKB participation/education loci, then check that disease biology still comes through.

1. Flag embedding GWAS hits tied to participation/SES/education.
2. Estimate those directions in embedding space.
3. Project them out (or lightly penalize in training).
4. Re-check disease AUC and lipid/asthma locus recovery.

5) Multi-ancestry check before using genetic rediscovery as a quality metric

Before treating "genetic rediscovery" as a model-quality score, it could be valuable to repeat embedding GWAS in more than one ancestry group (the Pietzner paper flags this limit clearly).

1. Keep one shared EHR model / embedding recipe.
2. Run embedding GWAS by ancestry group with adequate sample size.
3. Report shared vs private loci and drug-target recovery by group.
4. Check whether checkpoint rankings stay stable across groups.

Again, thanks to you and your team for your contribution to this field.