Table 1 Test set AUC using two different criteria.

From: Machine learning guided aptamer refinement and discovery

ML model name Output targets Normalized Rd 2 Count Sum AUCs Affinity threshold AUCs
Top 10%: Top 5%: Top 1%: 2 µM: 512 nM: 128 nM:
563 seqs 281 seqs 56 seqs 409 seqs 215 seqs 84 seqs
Counts Counts + PF 0.62 0.73 0.84 0.76 0.85 0.86
Binned Stringency Labels 0.62 0.78 0.89 0.82 0.91 0.95
SuperBin Stringency summary value 0.63 0.76 0.83 0.79 0.87 0.87
  1. Affinity threshold AUCs utilize the three stringency thresholds while Count Sum AUCs calculate the top fraction of sequences observed in Round 2 at the two highest stringencies. All models show improved prediction performance at increasing stringency levels with the binned model performing best at nearly all stringency criteria and levels.