For MD, I would not choose a pose only because it has the lowest docking score. Docking scores are useful for ranking poses inside the same setup, but they are not a reliable estimate of what will remain stable during a force-field MD simulation.

A practical way to choose starting structures is:

1. Look at the clusters first. A highly populated cluster usually means the search repeatedly found a similar binding mode, which is often more robust than a one-off very-low-score pose.
2. Within the main cluster(s), inspect the representative or lowest-energy pose manually: pocket fit, key H-bonds/salt bridges, hydrophobic contacts, metal coordination if relevant, and obvious clashes.
3. If the top-scoring pose and the most-populated cluster disagree, consider running short MD tests from both rather than forcing a single choice.
4. Before MD, convert the docked ligand/receptor carefully, assign ligand parameters, check protonation/tautomers, and verify that the ligand coordinates were not altered during conversion.
5. After minimization/equilibration, check whether the ligand leaves the pocket immediately or relaxes into a plausible interaction pattern.

So if "cluster 1" is both well populated and chemically plausible, using its best or representative pose is reasonable. If it is a singleton cluster with a slightly better score, I would be cautious and compare it against poses from larger clusters.

I work on ProteinIQ's AutoDock Vina tool, and this is one of the recurring caveats we try to make clear: docking can suggest poses to test, but MD setup and pose inspection still determine whether the follow-up simulation is meaningful.