In molecular communication (MC) systems, the \textit{expected} number of molecules observed at the receiver over time after the instantaneous release of molecules by the transmitter is referred to as the channel impulse response (CIR). Knowledge of the CIR is needed for the design of detection and equalization schemes. In this paper, we present a training-based CIR estimation framework for MC systems which aims at estimating the CIR based on the \textit{observed} number of molecules at the receiver due to emission of a \textit{sequence} of known numbers of molecules by the transmitter. Thereby, we distinguish two scenarios depending on whether or not statistical channel knowledge is available. In particular, we derive maximum likelihood (ML) and least sum of square errors (LSSE) estimators which do not require any knowledge of the channel statistics. For the case, when statistical channel knowledge is available, the corresponding maximum a posteriori (MAP) and linear minimum mean square error (LMMSE) estimators are provided. As performance bound, we derive the classical Cramer Rao (CR) lower bound, valid for any unbiased estimator, which does not exploit statistical channel knowledge, and the Bayesian CR lower bound, valid for any unbiased estimator, which exploits statistical channel knowledge. Finally, we propose optimal and suboptimal training sequence designs for the considered MC system. Simulation results confirm the analysis and compare the performance of the proposed estimation techniques with the respective CR lower bounds.