In this letter, we present a novel discrete Fourier transform (DFT)-based channel estimator for orthogonal frequency-division multiplexing (OFDM) systems. The conventional DFT-based estimator zeroes out noise-dominant values in the transform domain, which suffers from a poor performance in the presence of channel leakage. Some improved estimators have been proposed, but these estimators all focus on reducing the leaked energy. We propose to estimate the channel leakage by using the partial minimum mean square error (MMSE) filtering. When the channel delay is sufficiently less than the duration of an OFDM symbol, the partial MMSE filtering can be simplified. Simulation results show that the proposed channel estimator achieves much better performance than conventional estimators and well approaches the MMSE estimator.