Fig. 5: Performance of the goal-directed benchmarks for T16.SMPO with different training epochs.
From: t-SMILES: a fragment-based molecular representation framework for de novo ligand design
For t-SMILES family, random(R) and goal-directed(G) reconstruction are evaluated. “TS” in TSBR, TSBG, TSSR, TSSG, TSMR and TSMG means TSDY models. “B, M, S” in t-SMILES codes mean BRICS, MMPA and Scaffold based fragmentation algorithm. “SM” in SMG means SMILES model, “DS” in DSG means DSMILES model, “SF” in SFG means SELFIES model. “G” in SMG, DSG and SFG means more training rounds with 100. 10 or 15 random candidates are selected to calculate scores and the top-scoring one is chosen for output. The TSMG model yields significantly higher results. Further comprehensive experiments with different GPUs in SI.B.3.4 indicate that t-SMILES models exhibit high levels of repeatability with higher scores. Experiments against TSSA and TSID in SI.B.3.5 indicate that, similar to TSDY, TSSA models can outperform baseline models as well. This allows t-SMILES models to be used for the exploration of the limits in goal-oriented tasks.
