Large Scale Analysis of Linear Interacting Peptides in Model Organisms

Identifying which residues within intrinsically disordered regions (IDRs) mediate binding remains an open problem. This thesis compares two predictors on the bind-ing-IDR task: LIPNet, a CNN trained on ProtT5 embeddings, and AlphaFold-bind-ing, which scores residues by combining pLDDT confidence and solvent accessi-bility from AlphaFold2 structures. The comparison uses two independent references: the CAID3 binding-IDR bench-mark and a custom DisProt-derived set whose positive class unites GO binding an-notations with IDPO disorder-to-order transitions. On CAID3, LIPNet reaches an AUC-ROC of 0.616 while AlphaFold-binding falls to chance level (0.502; DeLong p = 2.7×10⁻³⁴); on DisProt the gap narrows but holds (0.563 vs 0.511; p = 1.6×10⁻⁵³). Absolute values remain modest for both methods, confirming the difficulty of the task. Sequence and structural characterisation show that the regions LIPNet recov-ers correctly are enriched in order-promoting and aromatic residues: the composi-tional signature of coupled folding-and-binding; and are more ordered and helical than the mostly coil-like reference. Applied to the full proteomes of Homo sapiens, Escherichia coli K12 and Saccharomyces cerevisiae with a proteome-calibrated cut-off, LIPNet's binding-rich proteins share a structural/ribosomal enrichment (a length-bias artefact) but show organism-specific signatures consistent with the ex-pansion of disordered binding with organismal complexity.