The Evidence Layer for Scientific AI

 Quick Explanation Copied

 Long Explanation

What Dynema Is

Dynema is a computational method for genome-wide, cell-state-dependent eQTL mapping at single-cell resolution. It uses a Poisson GLM with cluster-robust variance estimators (CRVE) and score/Lagrange multiplier tests, implemented in Julia. The motivating problem: pseudobulk eQTL approaches average away cell-state-dependent regulatory effects, and existing single-cell mixed models (scPME, CellRegMap, scLME) are both computationally heavy and statistically miscalibrated Dynema methods overview.

Evidence 1: Statistical Calibration

Under a genotype-permuted null in TBRU memory T cells (259 donors, ~500,089 cells), Dynema's analytical and bootstrap p-values were <0.01 in only 0.8% and 0.7% of interaction tests — close to the nominal 1%. Inflated false positives hit scPME (2.6%), CellRegMap (4.6%), and scLME (3.9%) Type I error calibration benchmark.

Evidence 2: Speed and Scaling

On ~500,000 cells, Dynema completed a multi-context interaction test in ~1.4 seconds per variant-gene pair — 10.5× faster than scPME and 1442× faster than CellRegMap Runtime benchmark vs competitors. Because the method is applied test-by-test per variant-gene pair, this speed is what makes genome-wide scanning feasible.

Evidence 3: Biological Results

Applied to two independent datasets (TBRU: 849 total eGenes; OneK1K: 985), Dynema found single-context interaction effects replicating across cohorts — allelic direction concordance 0.92–0.98 and z-score correlations 0.84–0.88 for cell-state canonical variates CV1–CV3 Cross-cohort replication of interaction eQTLs. Critically, after conditioning on the lead main eQTL, 71 (27%) and 117 (43%) interaction eGenes remained significant in TBRU and OneK1K — many cell-state effects are genetically distinct from main effects Conditional independence of interaction eGenes. At TSPAN32, the interaction effect colocalized with seropositive rheumatoid arthritis GWAS (PP.H4 = 0.99 TBRU, 0.98 OneK1K) while the main effect did not TSPAN32 RA colocalization result.

Limitations and What Could Change These Conclusions

• CRVE assumes donor independence — related individuals are out of scope; robustness can degrade in small samples (partially addressed by score bootstrapping) Method limitations.
• Only resting peripheral T cells were analyzed; other cell types, tissues, and activated/disease states are untested, and calibration beyond two T cell cohorts is unknown.
• The TSPAN32 colocalization is correlational — it implies shared causal variants, not proven causal mechanisms.
• Falsification: a null benchmark in another large single-cell eQTL dataset with >1% of Dynema interaction tests significant at p<0.01 would undermine the calibration claim.

Code: Dynema.jl and Dynema_analysis.