<p>Quantitative structure–activity relationship (QSAR) has been broadly applied in the computational peptidology area to tackle the regression modeling of multivariate statistical correlation between the sequence/structure and activity/function of bioactive peptides (BAPs). However, traditional peptide characterization methods such as amino acid descriptor (AAD) and one-hot encoding (OHE) are normally based on residue building blocks through the peptide sequence order, which would lead to a significant issue that the number of resulting features is relevant to peptide length, thus causing an inconsistency in feature vector dimensions for length-varying peptides, which cannot be directly engaged in QSAR modeling. Nowadays, only the auto-cross covariance (ACC) that was first proposed thirty years ago is available to treat the dimensional inconsistency issue, which, however, may have some disadvantages such as dimensional explosion and low interpretability. In addition to ACC, we describe a second (alternative) post-peptide characterization strategy termed residue descriptor-distance vector (RDDV) to scale the feature vectors of length-varying peptides. The RDDV includes two subclasses: RDDV(I) and RDDV(II), post-used for processing single- and multiple-AAD characterizations of peptide samples, respectively. They are employed to develop predictive QSAR models based on an in-house curated <i>ScBAPqad</i> dataset consisting of seven large-scale, length-varying BAP sample sets with experimentally measured quantitative activity values. We also perform systematic comparison of RDDV with ACC as well as global descriptor (GD) on these sample sets by combining different structural characterization strategies and machine learning methods, and discuss the applicability domains, limitations and future works of the proposed strategy.</p>