[NEW SUBMISSION] Confidence-Calibrated Delta scoring function

[EmmettBicker]

This PR adds in a new scoring method called Confidence-Calibrated Delta for ProSTT (K=2048) which raises the avg spearman on DMS_zero_shot from 0.507 to 0.523.

The intuitive change is going from

score = mutant score − wildtype score

to

score = mutant score − background penalty − uncertainty penalty + wildtype confidence bonus

Or more mathematically,

to

where l(a) = log p(a) and where background_weight and wildtype_confidence_weight are hyperparameters set to 0.35 and 0.2 respectively.

Here were my results:

• Average Spearman: 0.523
• Average NDCG: 0.773
• Average AUC: 0.785
• Average MCC: 0.408
• Average Top recall: 0.237

Selected Spearman breakdowns:

• Activity: 0.507
• Binding: 0.455
• Expression: 0.542
• Organismal fitness: 0.451
• Stability: 0.659
• Low MSA depth: 0.512
• Medium MSA depth: 0.517
• High MSA depth: 0.590

Takes ~3 minutes to run

I also independently validated average spearkman of K=1024 and K=4096

• 1024: 0.491
• 4096: 0.503
  But, I'm only interested in submitting 2048.

[EmmettBicker]

Hey @pascalnotin -- When you have time, would really appreciate a review of this