ScanNet

ScanNet: an interpretable geometric deep learning model for structure-based protein binding site prediction. Tubiana J, Schneidman-Duhovny D, Wolfson HJ. Nat Methods. 2022 Jun;19(6):730-739.

ScanNet: A web server for structure-based prediction of protein binding sites with geometric deep learning. Tubiana J, Schneidman-Duhovny D, Wolfson HJ. J Mol Biol. 2022 Oct 15;434(19):167758.

Method Overview

• each atom is characterized by the spatial arrangement (in local coords) and properties of its neighbors
• atom vector: how well the atom fits each “filter” (trained salient pattern)
• atom vectors for a residue are pooled and combined with residue properties (residue type, sequence conservation, ...)
• each residue is characterized by the spatial arrangement (in local coords) and properties of its neighbors
• residue vector: how well the residue fits each trained residue filter
• values are locally smoothed and converted to probabilities

Atom Patterns and Exemplars

• a. NH at center, O in quadrant opposite the -NH- covalent bonds (N-H-O hydrogen bond)
• c. C near methyl group and aromatic ring
• e. Backbone C or O without nearby NH (gray means absence)
• f. Solvent-exposed methionine sidechain

Residue Patterns and Exemplars

• Sequence logo height by strength of match to pattern
• Residues colored by category: negative-charge red, hydrophobic black, aromatic gold, etc.
• Oval-to-circular pie chart: blue exposed (a. hydrophobic surface, positive correlate of binding), gray buried (b. hydrophobic core, negative correlate of binding)

Predicting Protein-Protein Binding Sites

• Investigated four splits of 20K proteins with annotated binding sites from Dockground DB (a → d: ablated datasets with at least one training-set example with 70% sequence ID to the test sequence, or in same CATH “H” group, or in same “T” group, or none of those)
• Comparator 1: handcrafted set of geometric, chemical and evolutionary features for each amino acid (machine learning training in minutes, vs. ScanNet training 1-2 hours)
• Comparator 2: structural homology, aggregate probabilities in structural superposition of homologs (1 month!)

Predicting Protein-Protein Binding Sites

• Results for full dataset
• ScanNet is best overall
• Comparator 3: Masif, a surface-based geometric deep learning model trained on different dataset
  [paper: Nat Methods 2020] [github]

Predicting B-Cell Epitopes

• SAbDab structural antibody database of 3,756 protein chains with annotated BCEs (796 clusters at 95% sequence identity)
• applied to MD snapshot of SARS-CoV-2 spike trimer with one monomer in the RBD-open conformation
• representative antibodies from solved complexes are superimposed
• RBD-open monomer surface colored white → blue by increasing BCE probability
• high probability assigned to 8 epitopes previously described, but also around Asn residues expected to be glycosylated: N657 (mislabeled Asp in figure) and two others when predicting for the NTD domain alone (not shown)

ScanNet v. AlphaFold Multimer

• AlphaFold-Multimer outperformed ScanNet when using partner-protein sequence information, but was comparable when that was omitted (upper right, 6pnq with increasing ScanNet probability blue → red, AF predictions in green)
• ScanNet identified all the main epitopes of the SARS-CoV-2 spike RBD, whereas AlphaFold-Multimer predictions for six antibody complexes only identified one epitope (lower right, AF predictions in green)

ScanNet Server Options

• Fetch from PDB or AlphaFold DB, or upload local file (PDB, mmCIF)
• Choose one of three separately trained classes of binding site:
  • protein-protein
  • protein-antibody
  • protein-disordered protein
• Monomer or multimer prediction
• Whether to use a multiple sequence alignment
  (slowest step, but recommended except for epitope prediction or a designed protein)

Examples colored blue → red by increasing probability:

ScanNet Server Results

• Results are emailed to the user as a zip file containing:
  • the PDB file with binding site probabilities in the B-factor field
  • csv file containing the binding site probabilities
  • ChimeraX command script (.cxc)
  • Chimera python script (.py)

[Chimera-Related Software] [ChimeraX-Related Software]

ScanNet Server Results

• 121p (Ras) took <15 seconds, and the resulting email reported runtime 0.0s
• ScanNet finds the expected switch regions for effector binding
• results zip, the ChimeraX session I made for the figure on the right

• monomer #1.1, biounit #1.1 and #1.2
• surf hide
• example complex 1bkd, chain R ras, chain S Son of Sevenless
• open 1bkd; mm #2 to #1.1
• select using chain table in Log, seq features "short sequence motif"

Application Paper

Reduced B cell antigenicity of Omicron lowers host serologic response. Tubiana J et al., Cell Rep. 2022 Oct 18;41(3):111512.

• 15 sequence differences (Omicron vs. Wuhan) modeled individually and in combination(s)
• most of the mutations decreased antigenicity
• agreement with expt (antibody titers, pseudovirus neutralization)