For a fixed "pattern" graph $G$, the $\textit{colored $G$-subgraph isomorphism problem}$ (denoted $\mathrm{SUB}(G)$) asks, given an $n$-vertex graph $H$ and a coloring $V(H) \to V(G)$, whether $H$ contains a properly colored copy of $G$. The complexity of this problem is tied to parameterized versions of $\mathit{P}$ ${=}?$ $\mathit{NP}$ and $\mathit{L}$ ${=}?$ $\mathit{NL}$, among other questions. An overarching goal is to understand the complexity of $\mathrm{SUB}(G)$, under different computational models, in terms of natural invariants of the pattern graph $G$. In this paper, we establish a close relationship between the $\textit{formula complexity}$ of $\mathrm{SUB}$ and an invariant known as $\textit{tree-depth}$ (denoted $\mathrm{td}(G)$). $\mathrm{SUB}(G)$ is known to be solvable by monotone $\mathit{AC^0}$ formulas of size $O(n^{\mathrm{td}(G)})$. Our main result is an $n^{\tildeΩ(\mathrm{td}(G)^{1/3})}$ lower bound for formulas that are monotone $\textit{or}$ have sub-logarithmic depth. This complements a lower bound of Li, Razborov and Rossman (SICOMP 2017) relating tree-width and $\mathit{AC^0}$ circuit size. As a corollary, it implies a stronger homomorphism preservation theorem for first-order logic on finite structures (Rossman, ITCS 2017). The technical core of this result is an $n^{Ω(k)}$ lower bound in the special case where $G$ is a complete binary tree of height $k$, which we establish using the $\textit{pathset framework}$ introduced in (Rossman, SICOMP 2018). (The lower bound for general patterns follows via a recent excluded-minor characterization of tree-depth (Czerwiński et al, arXiv:1904.13077).) Additional results of this paper extend the pathset framework and improve upon both, the best known upper and lower bounds on the average-case formula size of $\mathrm{SUB}(G)$ when $G$ is a path.