Table 3.

Complex-based models. Summary of recent work using a protein-ligand complex for active molecule or binding affinity prediction. The year of publication, the name of the authors or the model, the complex encoding and the machine/deep learning model(s) are shown in the respective columns. Classification (class.) implies predicting e.g. hit or non-hit, whereas regression (reg.) evaluates e.g., $pI{C}_{50}$ values. CNNs, coupled with 3D grids, have become frequent in state-of-the-art studies.

Year	Name	Complex Encoding 1	ML/DL Model	Framework
2010	Sato et al. [84]	IFP	SVM, RF, MLP	class.
2016	Wang et al. [135]	IFP	Adaboost-SVM	class.
2019	Li et al. [136]	IFP	MLP	class.
2018	gnina [90]	3D grid	CNN	class.
2018	KDEEP [91]	3D grid	CNN	reg.
2018	Pafnucy [89]	3D grid	CNN	reg.
2018	DenseFS [137]	3D grid	CNN	class.
2019	DeepAtom [92]	3D grid	CNN	reg.
2019	Sato et al. [138]	3D grid	CNN	class.
2019	Erdas-Cicek et al. [94]	3D grid	CNN	reg.
2019	BindScope [93]	3D grid	CNN	class.
2018	PotentialNet [96]	graph	GGNN	reg.
2019	Lim et al. [95]	graph	GANN	class.
2017	TopologyNet [97]	topol.	CNN	reg.
2019	Math-DL [139]	topol.	GAN, CNN	reg.
2018	Cang et al. [140]	topol.	CNN	reg.
2016	DeepVS [99]	atom contexts	CNN	class.
2019	OnionNet [141]	atom pairs	CNN	reg.
2020	Zhu et al. [98]	atom pairs	MLP	reg.

¹ Abbreviations: IFP: interaction fingerprints, topol.: algebraic topology.