Fig. 8: Train losses and metrics vs. training steps.
From: Structure prediction of alternative protein conformations
The losses have been smoothed with an exponential moving average (step size = 100). The lDDT CA increases continuously, although almost all performance is reached in the first 10,000 steps (a). First, the MSA (e) and distogram losses (d) saturate, followed by the structural module loss (b) and plDDT loss (c). The total loss (f) is the sum of distogram, MSA, predicted lDDT (plDDT) and structural module losses.
