Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation

Table 1 Summary of all objectives/tasks used in this work and for which experiment (see Fig. 1)

Experiment	Aim	Objective type	Objective target	Performance measure
1	Compare REINVENT and AHC for varying values of σ	Docking	DRD2	Docking score & uniqueness
2	Compare REINVENT and AHC against different target systems	Docking	DRD2	Docking score & uniqueness
		Docking	OPRM1	Docking score & uniqueness
		Docking	AGTR1	Docking score & uniqueness
		Docking	OX1R	Docking score & uniqueness
3	Investigate and identify optimal DF and respective parameters for use with AHC	Similarity	Aripiprazole	Tanimoto similarity, uniqueness & wall time
		Isomer	C₁₁H₂₄	Isomer score, uniqueness & wall time
		Similarity & PhysChem (MPO)	Osimertinib	MPO score, uniqueness & wall time
4 & 5	Benchmark AHC to other commonly used RL strategies	PhysChem	Heavy atoms	# Heavy atoms, validity, uniqueness & wall time
		Similarity	Risperidone	Tanimoto similarity, validity, uniqueness & wall time
		Activity	DRD2	Predicted activity, validity, uniqueness & wall time
		Docking	DRD2	Docking score, validity, uniqueness & wall time
		Dual activity (MPO)	DRD2 & DRD3	Average predicted activity, validity, uniqueness & wall time
		Selectivity (MPO)	DRD2 > DRD3	Average predicted activity, validity, uniqueness & wall time

ISSN: 1758-2946