Designing novel peptides with amyloid-<i>β</i> binding and clearance potential using BiLSTM and molecular dynamics.

Abstract

Generative artificial intelligence is transforming de novo biomolecular design, yet developing models that reliably generate functional, target-specific peptides remains a significant challenge. Here, we introduce and validate a novel two-stage Bidirectional Long Short-Term Memory (BiLSTM) framework for the generative design of short, functional peptides. Our AI pipeline is trained on full-length proteins annotated with specific Gene Ontology (GO) terms related to amyloid-<i>β</i> (Aβ) interaction and is fine-tuned on experimentally validated peptide fragments to capture local functional motifs within a global protein context. As a proof-of-concept, we applied this framework to generate peptides targeting Aβ42, a key pathological agent in Alzheimer's disease. From 1,000 AI-generated sequences, 25 candidates were shortlisted using biophysical filters (GRAVY, instability index, Shannon entropy), and 11 were prioritized via sequence similarity analysis, designated as AI-Designed Novel Peptides (ADNP1-ADNP11). Structural modeling (AlphaFold2) and docking (pyDockWEB) against Aβ42 identified ADNP7 as the top candidate, exhibiting a highly favorable docking score (-63.33 kcal/mol), with interactions localized to Aβ's aggregation-prone regions. All-atom molecular dynamics simulations (20 ns) confirmed complex stability, and MM/PBSA analysis yielded a strong binding free energy (-50.6 kcal/mol), driven primarily by hydrophobic and aromatic interactions involving PHE12 and TRP50 in ADNP7. This work demonstrates that our fine-tuned BiLSTM architecture can successfully generate novel, stable peptide sequences with high predicted binding affinity for a therapeutically relevant target. While the training data included proteins associated with Aβ clearance (GO:0097242), only binding interactions were computationally validated; clearance potential remains a hypothesis for future experimental testing. This study establishes a generalizable, AI-driven pipeline for functional peptide design, with broad applicability across therapeutic discovery and synthetic biology.