High-throughput prediction of enzyme promiscuity based on substrate–product pairs