ClinicalCritic

EVA's 'Universal' Immune Model — Overstated Generality Claims

EVA: Towards a universal model of the immune system

Read the original source

What the paper says

EVA establishes a universal foundation model for immunology with clear scaling laws and strong performance across the drug development pipeline.

The Critique

The 'universal' claim is wildly overstated. The model is trained almost exclusively on human and mouse data—two species with extensive genomic resources. This isn't 'universal' immunology; it's 'well-funded model organism' immunology. They claim to capture 'complex phenotypes emerging from multicellular interactions' but their architecture primarily processes gene expression and histology separately, with limited modeling of true cellular dynamics or spatial tissue organization. The 58% PPV for drug efficacy sounds impressive until you realize it's on a hand-curated set of 28 drugs with known outcomes—a far cry from predicting efficacy for novel targets. Most critically, they don't validate on non-Western populations or diverse genetic backgrounds, where immune responses can differ dramatically. A 'universal' model should work across populations, not just across technical assays on convenience samples.

Why It Matters

If pharma companies adopt EVA for drug discovery assuming it's 'universal,' they may miss disease mechanisms specific to underrepresented populations. The model's blind spots become the field's blind spots.

What They Missed

They completely ignore the fundamental challenge of immune system variability—HLA diversity, microbiome interactions, environmental exposures, and trained immunity effects that vary across geography and socioeconomic status. A truly universal immune model needs to handle this heterogeneity, not just harmonize technical platforms.