Scholarlink

<p>Therapeutic anti-diabetic peptides (ADPs) are an emerging class of clinically-relevant biologics with the potential to manage glucose (glycaemic) levels. However, the majority of existing computational pipelines are either narrowly focused or unverified. Here, we present a principled, end-to-end, and provenance-aware framework for the de novo generation, filtering and classification of ADPs. In terms of design, candidate peptides are derived from three complementary strategies: guided modification of functional motifs, recombination of conserved bioactive fragments, and a hybrid generative engine. Candidates are then subjected to biochemical triage (net charge, hydrophobicity, Boman index), homology screening, and APD-style predictors/calculators. For classification, we use a blend of (1) interpretable biochemical descriptors (net charge, hydrophobicity, Boman index), and (2) sequence-derived representations learned by a CNN-with-attention backbone, to parse local motifs and longer-range context. Classifier heads are automatically tuned with an Optimized Tree-structured Parzen Estimator (OptimizedTPE). Training used 238 experimentally validated ADPs with homology-aware splits (Train/Val/Internal-Test positives: 167/24/47) and a curated negative pool at a 2:1 ratio; additionally, Train-only weak-label augmentation added 412 screened positives (and matched negatives) for robustness. We report an evaluation on an independent, external panel of 180 peptides, fully disjoint from the training data in both source and time. On this unseen set, the model achieves ≈ 98.75% accuracy (F1 ≈ 0.985, precision ≈ 0.99, recall ≈ 0.98, specificity ≈ 0.99, ROC AUC ≈ 0.99). This suggests high sensitivity to true ADPs while tightly controlling false positives in the setting of realistic class imbalance. Taken together, these results make the framework a candidate for a reproducible, biologically-grounded, in silico screening layer for metabolic peptide therapeutics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

De novo generation and in silico screening of anti-diabetic peptide candidates via a deep learning–attention framework with physicochemical feature fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月12日

查看全文
  • Zahra Rahmani Asl,
  • Khosro Rezaee,
  • Mojtaba Ansari,
  • Hadi Zare-Zardini,
  • Hossein Eslami

摘要

Therapeutic anti-diabetic peptides (ADPs) are an emerging class of clinically-relevant biologics with the potential to manage glucose (glycaemic) levels. However, the majority of existing computational pipelines are either narrowly focused or unverified. Here, we present a principled, end-to-end, and provenance-aware framework for the de novo generation, filtering and classification of ADPs. In terms of design, candidate peptides are derived from three complementary strategies: guided modification of functional motifs, recombination of conserved bioactive fragments, and a hybrid generative engine. Candidates are then subjected to biochemical triage (net charge, hydrophobicity, Boman index), homology screening, and APD-style predictors/calculators. For classification, we use a blend of (1) interpretable biochemical descriptors (net charge, hydrophobicity, Boman index), and (2) sequence-derived representations learned by a CNN-with-attention backbone, to parse local motifs and longer-range context. Classifier heads are automatically tuned with an Optimized Tree-structured Parzen Estimator (OptimizedTPE). Training used 238 experimentally validated ADPs with homology-aware splits (Train/Val/Internal-Test positives: 167/24/47) and a curated negative pool at a 2:1 ratio; additionally, Train-only weak-label augmentation added 412 screened positives (and matched negatives) for robustness. We report an evaluation on an independent, external panel of 180 peptides, fully disjoint from the training data in both source and time. On this unseen set, the model achieves ≈ 98.75% accuracy (F1 ≈ 0.985, precision ≈ 0.99, recall ≈ 0.98, specificity ≈ 0.99, ROC AUC ≈ 0.99). This suggests high sensitivity to true ADPs while tightly controlling false positives in the setting of realistic class imbalance. Taken together, these results make the framework a candidate for a reproducible, biologically-grounded, in silico screening layer for metabolic peptide therapeutics.