Going up to 5000 sequences per week is losing dynamic range here. It looks like all the action is definitely within the first 2000 sequences per week. Change the x-axis to span from 0 to 2000 sequences per week. Make datapoints every 100 along this line (will be 20 points along the x axis rather than the current 11 which seems fine). If this is computationally too demanding can reduce to datapoints every 200.
Average out estimates in panels B-D over 10 replicates. Right now stochastic noise makes it difficult to pick out a pattern. If this is computationally too demanding can reduce to 5 replicates.
A few recommended updates to Figure 5: