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

Abstract

Motivation: Offline protein design must seek high-activity variants while retaining support for their predicted performance. The challenge is to explore multiple promising regions without confining the search to average-performing sequences. We test whether multiscale representations and local evidence can guide this search. Results: GLOM-TGITO links a global-local fluorescence predictor to multiview landscape-guided optimization. Across 31,434 measured ppluGFP2 sequences, GLOM achieved random/depth-holdout RMSEs of 0.170/0.247 versus 0.247/0.324 for the strongest SaProt-650M controls. Averaged across three training seeds, 96.3%/97.3% of prediction-ranked Top 100 test variants were measured bright (log10 fluorescence >=4.2). Search from 900 peak regions yielded 2,594 distinct candidate sequences. The selected Top 100 showed 2.9-8.2% lower mean representation deviation than four comparator pipelines; 38 versus 11-16 met joint novelty-reliability filters. Deviation identified measured test errors >0.1 (AUROC 0.773). The first molecular-dynamics run of the selected Top 30 identified smaller surface hydrophobic patches in 20 candidates relative to their starts, with representative variants showing more persistent chromophore contacts or narrower torsional fluctuations.