Peptides present an alternative modality to immunoglobulin domains for developing therapeutics that can either agonize or antagonize cellular pathways associated with diseases. However, peptides often suffer from poor chemical and physical stability, which hampers their therapeutic potential. De novo designed hyperstable peptides offer a solution to this challenge, while many naturally occurred venom exhibit Cysteine Knot Peptide (CKP) topology that qualify them as suitable drug-like scaffolds for peptide therapeutics. Previously, we established a robust platform for discovering peptide therapeutics by utilizing multiple CKPs as scaffolds. However, we observed that many hits obtained from CKP-based libraries lost the CKP topology and resulted in peptides with diminished stability. We hypothesized that specific sequence patterns within the peptide scaffolds played a crucial role in spontaneous folding into a stable topology, and thus, these sequences should have been avoided during randomization in the original library design. Our ultimate goal was to develop a method for systematically designing highly diverse CKP libraries for each scaffold while preserving the inherent stability of each scaffold, allowing for spontaneous folding in vitro. We generated a large-scale dataset from yeast surface display experiments to train a machine learning model. This tailored prediction model for peptide folding exhibited high accuracy. Using the insights gained from the prediction model, we designed a new generation of alternative peptide scaffold libraries that were optimized for enhanced efficiency in proper folding. Subsequent panning trials using these libraries yielded promising hits with desirable properties in folding and stability. In summary, our study presents a methodological advancement in the field of peptide therapeutics. By combining yeast surface display experiments with machine learning-based prediction models, we successfully designed alternative peptide scaffold libraries that exhibit improved folding efficiency and stability. These findings pave the way for the development of more effective and stable peptide-based therapeutics in the future.