Explore four ways to encode automatically generated molecular features.
Change the molecule and target vector length, then look for the failure mode.
Design challenge:
the representation should have the same length for every molecule, handle new features,
let all generated features contribute, and preserve structural differences.
64
Used by strategies B and D.
This is the molecule whose automatically generated local structural features are being encoded below.
Heavy atoms
–
Simple proxy for molecule size
Unique features
–
Automatically generated local features
Output length
–
Feature inclusion
–
Fraction allowed to influence the representation
Distinct feature retention
–
How many distinct features remain distinguishable
What does the selected molecule become?
What is a collision?
feature A→position 12←feature B
Two different structural features have been assigned to the same vector position.
Because this is a bit vector, that position can only be 0 or 1. If either feature is present, the bit becomes 1.
If both are present, it is still just 1. We therefore lose the ability to tell those two features apart from that bit alone.
Using more positions usually makes these collisions less frequent.
Teaching model: the local structural features were precomputed from the SMILES using one fixed neighborhood size.
Strategy C is intentionally a coarse property summary, so students can see that a fixed-length representation can still lose local structural identity.