hypothesis.md

Hypothesis (probably false): You can use a "bag-of-words" model to compute the probability that a bound should break, and low probability bonds will be broken in symmetrical molecules.

• HOWTO: Compute the probability of each word for a bond; then, take the product. This is not numerically stable, so try:

  Sum(- log [word probability]) = - log (Product [word probability])
  

  This is the naïve Bayesian likelihood.

• Where to get words:

  • Matt's approach: Take 200 molecules, and pair them up randomly. This gives you (200 choose 2) = 19900 combinations. *5000 choose 2 = 12497500 Then, compute the Maximum Common Substructure for these 19900 and uniquify; this gives you ~3000 unique common substructures

    • Try on common molecules: Matt predicts even fewer MCSs
    • Does not preserve chirality

  • Prof. G's approach: Try natural products

    • Hypothesis: more = better (Matt disagrees)

• How to compute frequencies:

  • Use hasSubstructureMatch on some number of molecules

Supervised Learning Approach:

• Observation: any algorithm that suggests what bonds to cut based on a bag-of-words model will suggest symmetrical cuts with equal weight