Table 2.

Innovative models for the design and development of PROTAC linkers.

Model	Method	Principle	Characteristic	Ref.
PROTAC-INVENT	Reinforcement learning (RL)	The model is jointly trained using the RL approach, aiming to direct PROTAC structure generation towards predefined 2D and 3D properties.	It can not only generate PROTAC SMILES, but it can also calculate their possible three-dimensional bind conformations in conjunction with POI and E3 protein.	163
DeLinker	Graph-based deep generative model (Gated Graph Neural Network (GGNN))	The model combines structural knowledge with machine learning approaches, utilizes two fragments or partial structures, and generates a molecule that incorporates both.	This is the inaugural molecular generative model that incorporates 3D structural information directly into the design process, successfully demonstrating its efficacy and applicability in PROTAC design.	165
FFLOM	Deep generative model (graph convolutional network (GCN))	The model typically adjusts the fragment and generation lengths, optimizing the local fragment while preserving the dominant region and its conformation.	During the PROTAC design process, the model not only replicates the experimentally verified baseline molecule but also generates numerous novel structures with superior binding affinity scores.	167
AIMLinker	Deep learning (deep neural network (DNN))	The model retrieves structural data from input fragments and creates linkers for their integration.	It can rapidly design and generate linkers to create drug-like PROTACs with improved chemical properties.	170
Link-INVENT	Reinforcement learning (RNN)	The model takes two molecular subunits with a defined exit vector and a pair of warheads. It then produces a linker and outputs the connected molecule in SMILES format.	The model generates optimal linkers that connect molecular subunits and fulfill various goals, making it easier to use the model in practice for PROTAC design.	168
3DLinker	Conditional VAE-based generative model (vector neuron networks)	The model has the ability to predict anchor atoms and generates invariant graphs and equivariant absolute coordinates of linkers given two 3D fragments.	The model has a significantly higher rate of recovering molecular graphs and accurately predicting the 3D coordinates of all atoms.	164
ShapeLinker	Reinforcement learning (RNN)	The model implements fragment-linking using reinforcement learning on an autoregressive SMILES generator.	The model generates linkers that satisfy relevant 2D and 3D criteria, achieving top-notch results in generating novel linkers under a predefined linker conformation.	169
DRlinker	Deep reinforcement learning (transformer neural network)	The model manages the linking of fragments by exploring the desired chemical space for novel molecules with anticipated properties.	The model is effective for various tasks, ranging from controlling linker length and logP to optimizing the predicted bioactivity of compounds and addressing various multiobjective challenges.	166
PROTAC-RL	Reinforcement learning (RL) (transformer neural network)	The model, utilizing generative deep learning, incorporates a robust transformer architecture along with memory-enhanced RL to create highly effective PROTACs. It accepts E3 ligands and warheads as input and generates linker sequences that produce chemically feasible PROTACs with superior properties.	It not only enables the rational design of PROTACs in limited resource settings but also guides the selection of PROTACs with the most favorable pharmacokinetics.	172
DeepPROTAC	Deep neural network model (graph convolutional network)	In the model, ligand and ligand binding pockets are represented using graphs and fed into GCNs for feature extraction. Linkers are represented using SMILES notation to generate features for rational PROTAC design.	It not only can achieve rational PROTAC design but also can estimate the biodegradation activity of the resulting PROTAC based on the targeted POI and E3.	173