Scoring produces a number for each candidate, but a number alone does not make a drug. A candidate also has to survive a set of filters that encode what medicinal chemists have learned about which molecules tend to work as medicines and which tend not to.
Drug-likeness filters are the most familiar. They are simple rules of thumb about size, how greasy the molecule is, and how many hydrogen-bonding groups it carries. The reasoning behind them is practical: a molecule that is too large or too greasy tends not to dissolve well enough to be absorbed, and one with too many polar groups tends not to cross membranes. These are tendencies, not laws, and many successful drugs break them — but they are cheap to apply and they remove a large fraction of hopeless candidates early.
Safety filters look for structural features that are known to cause trouble: groups that are chemically reactive, that are associated with toxicity, or that the body tends to convert into something harmful. Synthesizability asks a different question entirely — can a chemist actually make this molecule in a reasonable number of steps from available starting materials? A beautiful binder that would take twenty steps and exotic reagents to build is not a drug candidate; it is a curiosity.
The funnel is the right picture. Thousands of generated candidates enter, and each successive filter removes a large slice. What comes out the far end is a shortlist small enough for humans to examine and for a laboratory to test.