The research investigates whether the microstructure of securities holdings among European banks influences the sensitivity of the asset's price to financial shocks. We conjecture that the network structure of common holdings and potential transaction bottlenecks creates specific liquidity conditions that traditional models cannot capture. Leveraging granular supervisory data from the Securities Holdings Statistics (SHS) and the Centralised Securities Database (CSDB) for the Euro zone, we propose a machine learning approach specifically utilizing graph-based representations to calibrate securities price impact functions. By modeling the bipartite network of banks and securities, we aim to identify non-linear drivers of fire-sale price discounts that depend on who holds the security and the potential pool of buyers. Finally, we illustrate the implications of these network-dependent sensitivities within a recursive price-mediated contagion model. The findings provide a data-driven framework for assessment of price-mediated systemic risk, as well as calibrating liquidity haircuts in regulatory stress tests, offering a more precise assessment of counterbalancing capacity and systemic risk.