Figure 2.

Performance evaluation of HydrogelFinder‐GPT and chemical structures of self‐assembling peptides. A) Performance comparison of generative models with different training strategies over valid, unique, novel and active on generation task (Details see Section 4.2 and Table 1). B) The generative capacity of model structural diversity under different training strategies. C) UMAP visualization of the chemical space distribution of candidates generated by HydrogelFinder‐GPT and generated by training set without self‐assembling small molecules. D) The statistical distributions of peptide‐based candidates. The peptide‐based candidates have 111 sequences (blue), of which 9 molecules can self‐assembly (Orange), while 8 molecules failed (Green). E) Chemical structures and identification numbers (IDNs) of the nine peptides selected from the candidate library that are able to self‐assembly.