Knowing which peptide can be detected during a mass spectrometry based proteomics analysis is a valuable information. Detectability prediction models predict detectability based on peptides amino acid sequences. Since peptide detectability varies substantially across different instruments, acquisition methods, and experimental conditions, sequence-based models alone cannot account for these sources of variability. Recent state-of-the-art methods address this transferability limitation by fine tuning specific models for each experimental setup. Nonetheless, these methods rely on significantly large training detectability datasets attaining 300k peptides, which incurs substantial costs both in terms of acquisition and processing times. In this study we propose a complementary method to infer detectability dataset from a single DIA spectrum. Such datasets can then be used to fine-tune prediction models from limited raw data while improving transferability to any specific setup. The associated goal is to further promote the use of detectability models in proteomics pipelines by cutting the underlying costs. For instance, we show that filtering search library based on predicted detectability simultaneously improve peptide identification and reduce computing time.