Table of Contents

TL;DR: Biased memos from ASHG 2026 talks/posters.

Some new works on molecular QTL methods

• Using multiple PC across ancestry
  • Linear mixed effect model...
  • Isn't PC correlated with cell types?
  • Are we imply recapitulating cell type QTL?
  • Expression PCs are computed from expression and also track cell type, so they sit downstream of both genotype and expression: a collider.
  • Adjusting for a collider opens the path G → C ← E, so PC correction can create or distort the genotype–expression association rather than only remove confounding.

graph LR
  G[Genotype] -->|eQTL effect| E[Gene expression]
  G --> C[Cell type / expression PCs]
  E --> C

• Poisson SuSiE

  • Poisson GLM and variational inference
  • Not convinced why a Poisson model can do better.

• Developmental GTEx

  • Use OpenTargets for validation
  • What disease prediction? Expression heritability.

• Genetic control of Treg suppression

  • Co-culture of Treg and CD4+ T-cell
  • CRISPR KO CTLA4 and IL2RA to confirm
  • 3,600 Treg eQTL and co-localization
  • Cytokine up/down regulation via different mechanisms
  • Recapitulate the findings via scRNA-seq with KO
  • Highlight myeloid differentiation
  • IKZF1 KO $\to$ destabilized Treg

• CCHC study

  • Brownsville, TX
  • Admixture with EUR, Latino, American Indigenous
  • Deeply phenotyped
  • 3,106 pQTLs across 1,361 proteins

• Single-molecule footprinting

  • Non-coding eQTLs
  • eQTLs explain 10%
  • caQTL explains more
  • SMF as molecular phenotypes
  • SMF-QTLs
  • Allele specific
  • open chromatin region vs. footprint
  • What a single variant can do vs. associated within LD
  • SNP is usually associated within motif
  • Isn't it a bit circular within tight LD?

Germline and somatic interaction

• Cancer somatic mutations with germline mutations

  • Finish data (iCAN/FiMM)
  • PRS on Breast, Lung, Colorectal
  • Mutational signatures vary across cancers
  • e.g., SBS4 ~ smoking and lung cancer
  • PRS association with somatic mutation signatures
  • e.g., liver function PRS ~ colorectal cancer signatures
  • What about cancer PRS vs. other PRS? Confounded? Collider bias?

• Multi-ancestry germline-somatic linkage

  • Genome-wide germline haplotype from panel sequencing
  • MSK, tumour cohorts and normal EHR
  • Gusev et al. Genome Medicine
  • "Driver" GWAS replicated across cohorts in germline and somatic
  • MDM4 TWAS associated with TP53 somatic mutations
  • MDM4 CRISPR-CAS9 KO didn't affect neighbouring genes
  • Risk allele specific TMPRSS2-ERG fusion in EAS prostate cancer

• Cancer pathogenic/likely pathogenic

  • Pathogenic Germline Variants in Cancer Susceptibility Genes
  • All of US
  • Median age at first cancer diagnosis inversely correlated with some genes mutation occurrence and their cancer prevalence
  • Several key genes still associated with cancer after accounting for germline-driven age effects (?)

EHR-based computational and AI approach to clinical genetics

• Phenotype risk scores

  • Goal: cardiac amyloidosis, hereditary V1421, Mount Sinai
  • ICD9 and ICD10 log odds ratio (heart traits?)
  • True label: TTR genetic testing
  • Workflow involving clinicians give out clinking alert
  • Strong comorbidity... ?

• Multimodal agentic knowledge graph

  • Amyloid cardiomyopathy
  • Phenotypic heterogeneity
  • Didn't explain what the graph is about...
  • Multimodal data $\to$ patient graph $\to$ differential diagnosis agent $\to$ specialist
  • Local open weight model
  • Gather textual data and process raw ECG
  • Build a disease agnostic patient graph
  • Ask a local agent to spot differential or missed diagnosis
  • What is the underlying cause of "differential" diagnosis?
  • GLM was the most sensitive agent.

• Assessing diagnostic delays

  • 1 in 20 genetic disorders in the US
  • Curated diagnostic delay PTs followed by mixed effect regression
  • 1.4 years median delay; 2.8 years mean delay

• Predict cancer type progression using LLM

  • Dermatomyositis (DM), autoimmune, 3-7x higher cancer risk
  • Myositis-specific antigen (MSA) tests 0.2%
  • TriNetX cohort
  • EHR = a sequence of chronological ICD-9/10 codes
  • Text embedding... why do we need to use pretrained LLM (gemma)?
  • During the pretraining, fine-tune?
  • Transfer learning (feature engineering), and then train a classifier
  • Help prioritize high actionable PT groups