Table 8.

A general overview of learning-based methods that have been applied to the task of designing functional peptides.

Task	Input data	Model	Output
Binders for Bcl-xL, Mcl-1, or Bfl-1	Set of 10,000 previously generated binders	SORTCERY [171]: Support vector machine for prediction, applied to a larger sequence library	36 high affinity binders for each target with Ki as low as 5.7$\pm$1.2 nM for Mcl-1, 1.32$\pm$0.18 nM for Bcl-xL, and 6.6$\pm$1.0 nM for Bfl-1
Finding peptide hits for Sfp and AcpS	Experimental data of known hits and fragments that don’t bind, each round new data was added	POOL [172]: Naive Bayesian	Selective binders for Sfp or AcpS after 4 rounds
Design of cell permeable peptides	600 PMO-miniprotein conjugates	RNN for generation and CNN for prediction [173]	Synthesized and characterized 12. Highly active for macromolecule delivery
Peptides with anticancer properties	Curated peptides from CancerPPD that target breast or lung cancer, pepCAST descriptors were used to find features	CPANN [174]: counter propagation artificial neural network for prediction and modLAMP for generation	6/15 peptides predicted to have anticancer activity indeed showed anticancer activity
Design of chitin binding peptides	21 million amino acid sequences taken from the Pfam database and 20000 sequences from the SCOPe database	Two Bi-LSTM each with 1024 hidden layers [175]	Interactive residues in the two predicted peptides matched experimentally known ones
Designing nonhemolytic AMPs	DBAASP database	RNN for classifying active, inactive, hemolytic, and nonhemolytic sequences [176]	12/28 were active