Toward AI Virtual Cells for Hepatology: Representation, Generation, Dynamics, and Intervention in Single-Cell Models

It proposes organizing liver AIVC research into three complementary modeling routes: generative models for representing cell states, dynamics and transport models for inferring state transitions, and pretrained or foundation models for testing whether learned representations transfer across donors, etiologies, disease stages, and platforms. Whereas single-cell and spatial atlases show where cell states occur, this review places scattered modeling attempts into a unified framework, making the division of labor and interfaces among different technical routes comparable. This is a review-level conceptual organization based on component-level evidence from published models rather than new experimental data.

It positions perturbation-response prediction as a cross-cutting assessment of whether these layers can predict responses to untested genetic, chemical, inflammatory, or metabolic interventions. It elevates perturbation prediction from a single task to a cross-cutting criterion for whether the whole AIVC framework is usable, linking representation, generation, and dynamics models. The text notes that published models demonstrate retrospective response programs, which is a component-level demonstration rather than prospective validation.

It categorizes available evidence into direct liver validation, liver-included benchmarks, general single-cell evidence, and conceptual applications, and states explicitly that these models do not constitute a prospectively validated liver simulator. It gives readers a graded view of how strong a conclusion each type of evidence can support, preventing general single-cell evidence or conceptual applications from being equated with liver-level validation. The categorization comes from the review's synthesis of existing literature; no specific sample sizes or effect sizes are reported.

It proposes minimum evaluation requirements: hold-outs at the donor, etiology, stage, platform, and perturbation levels, with performance reported using response direction, recovery of differentially expressed genes and rare states, and calibrated uncertainty; claims about tissue- or function-level prediction additionally require independent spatial, histologic, metabolic, and functional readouts. It expands evaluation from single accuracy metrics to multi-level hold-outs and multimodal independent validation, offering an actionable reporting standard for future work. These are suggested standards proposed by the review; no empirical results applying them are reported in the text.

The review addresses researchers at the intersection of hepatology and single-cell modeling, and applies to settings such as planning AIVC-related research, designing evaluation schemes, and judging evidence strength; its proposed hold-out levels and reporting metrics are intended as an evaluation framework rather than completed validation results.

Readers should still watch: the text gives no specific model names, dataset sizes, or performance numbers, making it hard to judge maturity differences among the routes; the specific assignment of evidence to the four tiers requires the full text; and this reading is an incomplete text without figures or references, which may affect grasp of evidential detail.