Over the past decade, it has become clear that olfactory input signals combinatorially encoded to olfactory receptor neurons (ORNs) undergo a substantial expansion of dimensionality and sparsification as they are relayed to Kenyon cells (KCs). The effects of such transformations on downstream processing have been studied theoretically using synthetic input and random projections, but how these transformations reshape real odor codes exhibiting significant correlation in a biological circuit remains unknown. Here, using previously published single-sensillum recordings spanning 23 ORN types responding to 110 monomolecular odorants and the wiring diagram of the \emph{Drosophila} brain revealed by a connectomic dataset, we quantify how the neural representation of each odorant is transformed and redistributed across the fly brain. Odor representations in the KC space show that consistent with experimentally measured sparsity in KCs, odor class discriminability is maximized at sparsity levels of 0.4\% to 6\%. Upon coarse-graining, the entropy of odor codes in KC layer is also greater than that in ORN layer, indicating that odor codes in KC representations are more uniformly redistributed which facilitates their linear decoding in the downstream circuits. Moreover, although behavioral valence is generally thought to be resolved only at the mushroom body output neurons, we find an evidence that aversive odors are already separable at the level of KCs. By combining real physiological and connectomic data, our study reveals how neural circuits reshape the geometry of sensory representations of natural odors to enable efficient olfactory signal processing in the fly brain.