Transmission control protocol (TCP) is a reliable transport layer protocol widely used in the Internet over decades. However, the performances of existing TCP congestion control algorithms degrade severely in modern heterogeneous networks with random packet losses, packet reordering and congestion. In this paper, we propose a novel TCP algorithm named TCP-ACC to handle all three challenges mentioned above. It integrates 1) a real-time reorder metric for calculating the probabilities of unnecessary Fast Retransmit (FRetran) and Timeouts (TO), 2) an improved RTT estimation algorithm giving more weights to packets that are sent (as opposed to received) more recently, and 3) an improved congestion control mechanism based on packet loss and reorder rate measurements. Theoretical analysis demonstrates the equilibrium throughput of TCP-ACC is much higher than traditional TCP, while maintaining good fairness with regard to other TCP algorithms in ideal network conditions. Extensive experimental results using both network emulators and real network show that the algorithm achieves significant throughput improvement in heterogeneous networks as compared with other state-of-the-art algorithms.