Single-cell technologies have increased the resolution at which disease biology can be examined, yet computational target-discovery strategies still frequently treat cell type as the principal biological unit. This practice can obscure therapeutically important variation because cells assigned to the same lineage may occupy divergent activation, differentiation, stress, resistance, or disease-associated states. The unresolved problem is therefore not simply how to classify cells more accurately, but how to determine which state-dependent molecular relationships are sufficiently reproducible, mechanistically credible, and contextually bounded to support therapeutic hypotheses. This theory article proposes State-Anchored Target Discovery as a conceptual framework for organizing such evidence. The proposed construct represents a candidate target through a state–transition–context relationship: the cellular state in which the target is implicated, the disease- or treatment-associated transition it may influence, and the biological context within which the relationship is expected to hold. It further separates state markers from causal regulators, distinguishes predicted trajectories from experimentally confirmed transitions, and treats annotation uncertainty, cross-cohort replication, and perturbational evidence as integral parts of the target claim rather than downstream technical checks. The central contribution is an evidence architecture that connects single-cell heterogeneity to target prioritization and patient-stratification hypotheses without reducing scientific adequacy to a single clustering, integration, prediction, or benchmark measure. The construct remains theoretical. It cannot establish causality, druggability, safety, developability, clinical utility, or regulatory acceptability without additional experimental and pharmaceutical evidence. Its value lies in defining a more precise analytical object for therapeutic target discovery and in clarifying the validation boundaries that should govern state-centered computational claims.