Fig. 2: Quality of computational model and properties of experimentally tested design variants.

[A, Model Quality] The computational model was evaluated for quality by comparison to published metrics of β-lactamase structure and function. Predicted residue-residue interactions (top L model couplings that are at least 5 positions apart in primary sequence) are compared with known structural contacts of WT TEM-1 as determined by X-ray crystallography (PDB: 1XPB³⁴). Dots indicate contacts either in the model (couplings) or PDB structure. Green bar on y-axis (left) and x-axis (top) indicates positions that are aligned in the MSA used for model inference, and are therefore included in the model (n = 252). Dashed gray bars indicate position numbers that are skipped in the PDB structure (positions 239 and 253). Over 80% of the top predicted interactions are structural contacts. [B, Design Properties] (Predicted fitness, left): Predicted fitness changes (ΔEVH) between WT TEM-1 and designs (bottom) as well as the natural homologs of β-lactamases (top, from the multiple sequence alignment used to generate the model). (Diversity, right): Table of similarities and general properties of each design. Identity to natural homologs was determined using blastp (Methods). Core residues were defined as having a relative surface accessibility (ACC field of DSSP analysis of PDB 1XPB divided by residue size) of less than 0.2. [C, Design Sequences] Multiple sequence alignment (MSA) of all designs, with amino acid changes relative to WT TEM-1 colored by new residue properties (standard colors: green, hydrophobic and glycine (G); blue, negative charge; red, positive charge; light blue, polar). Conservation and logo for each position in the multiple sequence alignment of natural homologs are displayed above the design sequences. Empty columns in the logo and lowercase letters in the WT TEM-1 sequence indicate unaligned positions in the MSA and model. Source data are provided in the Source Data file.