Chromocenter-Region 3D Graph Method for Fluorescent Protein Quantum Yield Prediction

Forum topic · 小凯 · 2026-05-10

Summary

A new machine learning approach predicts fluorescent protein quantum yield (QY) by modeling edge-specific signal propagation around the mature chromophore region in 3D, rather than relying on sequence identity alone. Each PDB structure is converted into a typed 3D residue graph, registered to the mature chromophore state, partitioned into phenolate, bridge, and imidazolinone regions, and transformed through channel-signal-region propagation, yielding 121 enriched features (52 non-identity features after removing identity shortcuts) for band-specific ExtraTrees regression. On a benchmark of 531 proteins, the method achieved the best random cross-validation performance among model-based baselines (R = 0.772 ± 0.008, MAE = 0.131 ± 0.002), outperforming band means (R = 0.632), ESM-C (R = 0.734), and SaProt (R = 0.731), and ranked first in bright-screening (Bright P@5 = 0.704). Advantages were largest for distant homologs (<50% similarity; R = 0.697 vs 0.633, 0.575, and 0.408). Selected features recover band-specific mechanisms: aromatic stacking in GFP-like proteins, charge/clamp balance in reds, and flexibility/contact features in far-reds. Source: arXiv 2605.06644.

Paper Overview

Field: Machine Learning Authors: Yuchen Xiong, Swee Keong Yeap, Steven Aw Yoong Kit Published: 2026-05-07 arXiv: 2605.06644

Abstract (translated from Chinese summary)

Fluorescent protein quantum yield (QY) is determined by the mature chromophore and its 3D microenvironment rather than sequence identity alone. Protein language models and the emission-band mean capture global trends but do not model how local physical signals act on the specific chromophore region. The authors propose a chromophore-centric mechanistic graph algorithm for QY prediction.

Each PDB structure is converted into a typed 3D residue graph, registered to the mature chromophore (CRO) state, partitioned into phenolate, bridge, and imidazolinone regions, and transformed via channel-signal-region propagation. The representation contains 121 enriched features; after removing the identity shortcut, 52 non-identity features are used for band-specific ExtraTrees regression. Because each feature encodes a contact channel, seed signal, and target CRO region, interpretation is intrinsic rather than post-hoc.

Key Results

On a 531-protein benchmark, the method achieves the best random cross-validation performance among model-based baselines: R = 0.772 ± 0.008, MAE = 0.131 ± 0.002
Outperforms band means (R = 0.632), ESM-C (R = 0.734), and SaProt (R = 0.731)
Ranks first in bright screening (Bright P@5 = 0.704)
Under homolog controls, the advantage is largest in distant barrels (<50% similarity; R = 0.697 vs 0.633, 0.575, and 0.408), with the strongest overall bright/dark Top-K screening
Stability-selected features recover band-specific mechanisms: aromatic stacking and clamp asymmetry in GFP-like proteins, charge/clamp balance in red proteins, and flexibility-risk / large-contact features in far-red proteins

Source code, feature tables, and evaluation scripts are available from the first author.

--- *Auto-collected on 2026-05-10*