Signal, noise, and sampling: How pool size and replication shape metabolomic inference

Metabolomics provides a direct readout of physiological state and is increasingly used in evolutionary and systems biology. In small organisms such as Drosophila melanogaster, metabolomic analyses typically require pooling individuals to obtain sufficient material, yet pool sizes vary widely across studies with little justification. How pooling and biological replication influence metabolome characterization and the detection of biological signal remains poorly understood. Here, we evaluate the effects of pool size and biological replication on metabolomic profiles and signal detection using two complementary experimental designs. In the first, we assess how pooling (5, 50, or 100 individuals) influences metabolomic structure and reproducibility in inbred and outbred populations. In the second, we test how pool size interacts with systematic variation in replicate number to affect detection of diet-associated metabolite changes under a high-sugar perturbation. Pool size strongly influenced metabolomic profiles, with samples pooled at five individuals consistently differing from larger pools, while profiles from 50 and 100 individuals were more similar. Larger pools improved reproducibility in a dataset-dependent manner. In the dietary experiment, smaller pool sizes substantially reduced sensitivity, leading to loss of true diet-associated metabolites without increasing false discoveries. Replicate downsampling further revealed that both pool size and biological replication jointly determine signal retention, with smaller pools accelerating the loss of detectable metabolites under reduced replication. Across all analyses, the ability to detect metabolite signals was strongly dependent on effect size and variability. Metabolites with larger and more stable effect estimates were consistently retained, whereas those with smaller or more variable effects were rapidly lost under reduced sampling. Linear mixed-effects modeling confirmed that detection probability is governed by a balance between biological signal strength and measurement variability, with pool size and replication jointly modulating this relationship. More broadly, our results demonstrate that metabolomic inference is governed by the interplay of signal, noise, and sampling design, with pool size and replication jointly shaping the detectability, stability, and interpretation of biological signals.