Jasdeep Singh

Using experimentally determined fibril structures from the Amyloid Atlas, with ordered residues labeled as core and unresolved residues from the same proteins used as matched non-core controls, this work encoded core and non-core regions with physicochemical descriptors and protein language model embeddings (ESM-2, ANKH, ProtT5) and evaluated them under protein-grouped cross-validation to build AmyloCore-ML, a sequence-based predictor of structural amyloid-core propensity; protein language models consistently outperformed physicochemical descriptors, with the best models reaching AUROC about 0.88 and AUPRC about 0.85, locked full-length protein scans localized experimental cores with ESM-2/ExtraTrees W21 achieving AUROC 0.833, AUPRC 0.751 and a mean peak distance of 12.

Using experimentally determined fibril structures from the Amyloid Atlas, with ordered residues labeled as core and unresolved residues from the same proteins used as matched non-core controls, this work encoded core and non-core regions with physicochemical descriptors and protein language model embeddings (ESM-2, ANKH, ProtT5) and evaluated them under protein-grouped cross-validation to build AmyloCore-ML, a sequence-based predictor of structural amyloid-core propensity; protein language models consistently outperformed physicochemical descriptors, with the best models reaching AUROC about 0.88 and AUPRC about 0.85, locked full-length protein scans localized experimental cores with ESM-2/ExtraTrees W21 achieving AUROC 0.833, AUPRC 0.751 and a mean peak distance of 12.