#19 · Structural bio & protein design

Predictors for metal sites, molecular contacts and AlphaFold3 performance gaps

The dominant thread is computational representation of molecular structure for prediction and evaluation, though the four items span distinct applications rather than a single method. BioMetAll v2.0 extends metal-binding site prediction with metal-specific scoring and side-chain, first-coordination-sphere descriptors. A rank-based atom environment representation and fitness score (FS) discriminate native X-ray structures in molecular contact assessment. On the evaluation side, AlphaFold3 shows uneven performance and training-data dependence across ubiquitinated proteins, TCR-epitope, antibody-antigen, protein-RNA and protein-lipid tasks. NMR-AI takes a different route, combining predicted 1H/13C shifts with ECFP4 fingerprints to improve logP, logS, logD and pKa prediction, with up to 39% lower RMSE. The novelty lies in encoding local chemical environments, metal identity, and predicted spectroscopic information as model features.