Predictive adequacy and measurement dependence of metabolic representations in controlled synthetic systems

Abstract

Background: A metabolic representation may predict adequately, appear compact or change repeatably without satisfying the other properties. Evaluating these properties requires explicit reference information, measurement conditions and definitions of change. Methods: Six stipulated stochastic scenarios generated twenty independent confirmation cohorts each, with 5,115 synthetic individuals per cohort and 1,023 held out. Three representation families (selected current and longitudinal descriptors [Phi], a larger selected collection of measurements and historical descriptors, and an outcome guided learned embedding of longitudinal inputs) were evaluated against complete available longitudinal history for prediction at year 25 from information available through year 20. Forty two adequacy groups used overall and context specific relative loss margins with 2,850 simultaneous upper bounds. Information accounting distinguished displayed coordinates, upstream values and nominal input fields. Separate procedures examined current remeasurement, natural coordinate tracking and controlled endpoint responses. Results: Five groups satisfied every adequacy requirement; none of the twelve Phi groups did. Of 393 failed predicates, 321 had estimates within the margin but upper bounds above it. The conventional panel embedding displayed eight coordinates but depended on 72 upstream engineered values and additional baseline, context, support and missingness inputs. Remeasurement stability did not ensure lower prediction error, and changing the future scoring reference could reverse a method ordering without changing predictions. Natural tracking yielded 25,440 defined associations, including 442 negative values, and 1,440 undefined associations. Controlled responses agreed with their assigned directions when numerically resolved, but included mixed dose sequences and unsupported coordinates. Of 47,127,708 eligible stochastic response records, 86.331% were smaller than one matched measurement standard deviation; 96.587% of records exceeding that reference retained their corresponding category. Conclusions: Adequacy, compactness, repeatability and responsiveness describe different properties of a representation. Failure to certify adequacy does not necessarily demonstrate excess loss, and movement without a category change does not establish clinical importance or subsequent predictive harm. Evaluation should retain complete input requirements, measurement references and unsupported cases. These synthetic findings do not establish a sufficient human physiological representation or therapeutic benefit.