GLOM TGITO: multiscale protein modelling and multiview landscape-guided sequence discovery

GLOM-TGITO improves prediction and prioritization for measured ppluGFP2 sequences, achieving RMSE 0.170/0.247 and yielding 96.3%/97.3% bright Top 100 variants.

• Why it matters: Offline protein design can over-select high-scoring sequences that lack measured support due to surrogate-predictor errors, so better representations and local evidence are needed.
• What they did: The method links a global-local fluorescence predictor to multiview landscape-guided optimization, evaluated on 31,434 measured ppluGFP2 sequences with depth-holdout testing and 900 peak-region searches.
• The result: Compared with strongest SaProt-650M controls (0.247/0.324 RMSE), GLOM produced 2,594 distinct candidates and selected Top 100 with 2.9–8.2% lower mean representation deviation, passing 38 joint novelty-reliability filters versus 11–16.