Torsions with pairwise energies

PB

• Slides - https://docs.google.com/presentation/d/1taA5O5VHGYC90CjKaLWgfXLLrGK1lUqt0PgXAteVSLM/edit?usp=sharing

• Presentation from Bill last year on the conformer energy mismatch issue - https://docs.google.com/presentation/d/1XG6AudhWvKjkW5Ia-aeyKMPiv-Bhp8XD/edit?usp=sharing&ouid=112699581003689115831&rtpof=true&sd=true

• OMSF talk from PB which has the definition of Swope Fingerprint - Slides 28 to 30 are relevant - https://docs.google.com/presentation/d/19it7iGyOz0LlNNdSuDMjcfMD7niIUKcpBWncpPxD5qo/edit?usp=sharing

• No improvement wrt Sage 2.1.0, possibly due to not including the optgeo target

• Concordant pair: agreement in the rank (that conformer 3 is lower than conformer 4 in energy), discordant pair disagrees

• BS – note that with the order loss, this assumes that conformer i in MM is similar to conformer i in QM

• PB – with the torsion energy targets there’s a restraint of 1 kcal/mol so the geometry shouldn’t relax much

• AMI – what’s the goal?

  • (General) – getting relative conformer energies right, appropriately penalising large geometry shifts and conformer ranking

  • DLM – how do we change the fitting process to fix the relative energies without hurting other things (RMSD)

• BS – LPW suggested augmenting training set with spurious minima

  • LW – not going too well

  • PB – suggests iterative fit along same lines. The other approach is building a new target and a new loss function

  • BW – does FB do the single-point QM data generation of new MM structures, or is the iteration in human time?

    • PB – the latter – I have scripts