Fig. 3

Overview of experiments on simulated datasets. We simulated three types of protein libraries: (i) peptide insertion, (ii) AAV recombination, and (iii) avGFP random mutagenesis libraries (Table 1). We simulated three types of sequencing datasets: (i) short reads (for all libraries), (ii) long reads (for all libraries except the peptide insertion libraries, where short reads cover the entire region of interest), and (iii) hybrid reads (for the AAV recombination library). We employed three types of architectures: (i) linear, (ii) fully-connected neural network, and (iii) convolutional neural network—using a classification head for MBE and a regression head for wLER. For short-read and hybrid-read datasets, we only used convolutional neural networks because only they can operate on variable-length sequences