Figure 4.

Application of hallucination pipeline to generate large number of unique designs with improved binding as characterized by various interface metrics. (A) A pipeline for screening the hallucinated library for folding into the target structure and binding the target epitope. The screened library is characterized in silico for interfacial metrics to select designs for experimental testing. (B) Comparison of structure of forward folded hallucinated sequences with wildtype structure: (Left) Randomly selected forward folded structures of hallucinated designs (green) with DeepAb superposed with wildtype structure (black). (Right) The distribution of the H3 RMSD of the designs with respect to the wildtype. Dashed line at RMSD = 2 Å marks the threshold for selecting designs that fold into the wildtype structure. RMSDs were calculated on all heavy atom backbone residues (N, CA, C) excluding (O) consistent with RMSD reported in DeepAb (28). Distribution is shown for the motif-restricted hallucination I (95, 100A restricted) from Figures 3D, E . (C) Distribution of screened binding free energies against HER2 for the designs. The dashed line is the binding free energy of the wildtype Trastuzumab antibody. Popouts: (Left) Representative design with G at position 99 that exhibits binding energies better than or comparable to the wildtype. (Right) Representative design with an I at position 99 that exhibits binding energies significantly worse than the wildtype. (D) Sequence logos of screened libraries for the unrestricted hallucination and motif-restricted hallucination I (95 and 100A restricted). The unrestricted designs show strong preference for residues G and Y at positions 99 and 100A respectively. The motif-restricted (95 and 100A) mode additionally confirms the preference for Gly at position 99. Both restrictions (at positions 99 and 100A) are also observed in experiments. (E) Comparison of distributions of binding free energies for designs from two motif-restricted hallucinations: (gray) Motif-restricted hallucination with positions 95, 99, 100, 100A restricted to experimentally observed preferences, and (orange) motif-restricted hallucination with positions 99 and 100A restricted to hallucination-pipeline derived preferences from (D). Popout: Sequence logo for the screened library for designs from the latter hallucination. (F) Interface metrics for virtually screened designs from motif-restricted hallucination (99 and 100A restricted). A significant fraction of designs exhibit metrics better than the wildtype (boxed quadrant); top 25 designs (selected by binding free energy) highlighted in orange.