TrendBio

#4 · ML / AI methods

Learned predictors score protein interactions, variant effects, and peptide activity

6 preprints · 1 published
Items per day, last 8 weekstoday: 7

The dominant thread is learned predictors that map sequence or structure to biomolecular properties, with a strong emphasis on benchmark construction and evaluation. SNAC-DB provides an ML-ready antibody and VHH-antigen complex database with 31-37% more structural diversity than SAbDab and benchmarks seven structure predictors. A separate benchmark study reports AlphaFold2 outperforming OpenFold3 on 271 protein-peptide complexes under CAPRI criteria, with evidence of memorization. SPPIDER-seq predicts partner-aware interaction sites from ESM-2 embeddings, reaching AUROC up to 0.797. UshEffect-3D applies gene-specific classification to USH2A missense variants, and GenPept-Curated-2025 offers homology-controlled antimicrobial peptide benchmarks. Another study tests AlphaFold2 Evoformer representations as features for sequence-based prediction tasks. CRC-Screen addresses DNA-synthesis hazard screening under taxonomic shift, a sequence-screening problem outside the structure-prediction thread. DDA-BERT, a published transformer for proteomics, extends the same learned-predictor pattern to mass spectrometry.

Published papers

Preprints

Topic history

DateRank
#4
#3
#2
#2
#3
#1
#2
#3
#2
#3
#2
#3
#3
#19
#2
#6
#12
#12
#4
#17
#16
#5
#5
#2
#4
#14
#13
#16