Table 8 Open in new tab Method for...

Name

Class

Model

Training
dataset

PearsonR

RMSE

Description

Ref

GearBind

Antibody

Geometric
GNN

SKEMPI v2
(6k mutations)
PDB
(123k struct.)

SM: 0.62

SM: 1.40 Å

Strengths: Use of contrastive learning to detect destabilizing mutations.
Limitations: Mutant structure generation time should be improved.

[168]

GeoPPI

Antibody

GAT

PDB-BIND [169]
3DComplexes [170]
(13k mutations)

SM: 0.58
MM: 0.74

SM: 2.01 Å
MM: 2.21 Å

Strengths: Self-supervised learning to reconstruct the coordinates of the perturbed side chains.
Limitations: Lower performance for single mutation compared with the other two presented methods.

[171]

Shan et al.

Antibody

Transformer

SKEMPI V2.0
(5k mutations)

SM: 0.65
MM: 0.59

SM: -
MM: -

Strengths: The attention network learns to identify key residue pairs near the protein interface that contribute to binding affinity.
Limitations: Operates at the residue level and does not consider the atom level.

[73]

Table 8

Open in new tab

Method for antibody affinity maturation compared in terms of Pearson correlation coefficient (PearsonR) and Root Mean Square Error (RMSE). The methods are compared for single mutation (SM) and multiple mutation (MM) using S1131 [172] and M1707 [173] respectively, with the exception of GearBind. The ’Application’ is not shown in the ’Description’ column as all methods predict mutational effects on binding affinity. (structures = struct.)

Name	Class	Model	Training dataset	PearsonR	RMSE	Description	Ref
GearBind	Antibody	Geometric GNN	SKEMPI v2 (6k mutations) PDB (123k struct.)	SM: 0.62	SM: 1.40 Å	Strengths: Use of contrastive learning to detect destabilizing mutations. Limitations: Mutant structure generation time should be improved.	[168]
GeoPPI	Antibody	GAT	PDB-BIND [169] 3DComplexes [170] (13k mutations)	SM: 0.58 MM: 0.74	SM: 2.01 Å MM: 2.21 Å	Strengths: Self-supervised learning to reconstruct the coordinates of the perturbed side chains. Limitations: Lower performance for single mutation compared with the other two presented methods.	[171]
Shan et al.	Antibody	Transformer	SKEMPI V2.0 (5k mutations)	SM: 0.65 MM: 0.59	SM: - MM: -	Strengths: The attention network learns to identify key residue pairs near the protein interface that contribute to binding affinity. Limitations: Operates at the residue level and does not consider the atom level.	[73]

Name	Class	Model	Training dataset	PearsonR	RMSE	Description	Ref
GearBind	Antibody	Geometric GNN	SKEMPI v2 (6k mutations) PDB (123k struct.)	SM: 0.62	SM: 1.40 Å	Strengths: Use of contrastive learning to detect destabilizing mutations. Limitations: Mutant structure generation time should be improved.	[168]
GeoPPI	Antibody	GAT	PDB-BIND [169] 3DComplexes [170] (13k mutations)	SM: 0.58 MM: 0.74	SM: 2.01 Å MM: 2.21 Å	Strengths: Self-supervised learning to reconstruct the coordinates of the perturbed side chains. Limitations: Lower performance for single mutation compared with the other two presented methods.	[171]
Shan et al.	Antibody	Transformer	SKEMPI V2.0 (5k mutations)	SM: 0.65 MM: 0.59	SM: - MM: -	Strengths: The attention network learns to identify key residue pairs near the protein interface that contribute to binding affinity. Limitations: Operates at the residue level and does not consider the atom level.	[73]

This Feature Is Available To Subscribers Only