Edge-Specific Signal Propagation on Mature Chromophore-Region 3D Graphs for Fluorescent Protein Quantum Yield Prediction

Forum topic · 小凯 · 2026-05-10

Summary

This paper proposes a chromophore-centric mechanistic graph algorithm for predicting the quantum yield (QY) of fluorescent proteins. Fluorescent protein QY is determined by the mature chromophore and its 3D microenvironment rather than sequence identity alone. The method converts each PDB structure into a typed 3D residue graph, registers it to the mature CRO state, partitions it into phenolate, bridging, and imidazolinone regions, and applies channel-signal-region propagation. The representation yields 121 enriched features; after removing identity shortcuts, 52 non-identity features feed band-specific ExtraTrees regression, making interpretation intrinsic rather than post hoc. On a 531-protein benchmark, the approach achieves the best random cross-validation among model-based baselines (R = 0.772 ± 0.008, MAE = 0.131 ± 0.002), exceeding band-mean (R = 0.632), ESM-C (0.734), and SaProt (0.731), and ranks first in bright screening (Bright P@5 = 0.704). Gains are largest in distant homology buckets (<50% similarity). Selected features recover band-specific mechanisms such as aromatic stacking in GFP-likes, charge/clamp balance in reds, and flexibility risks in far-reds. Source code, feature tables, and evaluation scripts are available from the first author.

Overview

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

Abstract

Fluorescent protein quantum yield (QY) is determined by the mature chromophore and its three-dimensional microenvironment, not by sequence identity alone. Protein language models and emission-band means capture global trends, but do not model how local physical signals act on specific chromophore regions.

The authors propose a chromophore-centric mechanistic graph algorithm for QY prediction:

1. Each PDB structure is converted into a typed 3D residue graph. 2. The graph is registered to the mature CRO (chromophore) state. 3. It is partitioned into phenolate, bridging, and imidazolinone regions. 4. Channel–signal–region propagation is applied, yielding 121 enriched features.

After removing identity shortcuts, 52 non-identity features are used in band-specific ExtraTrees regression. Because each feature encodes a contact channel, seed signal, and target CRO region, interpretation is intrinsic rather than post hoc.

Results

Benchmark (531 proteins): best random cross-validation performance among model-based baselines — R = 0.772 ± 0.008, MAE = 0.131 ± 0.002.
Baselines: outperforms band-mean (R = 0.632), ESM-C (R = 0.734), and SaProt (R = 0.731).
Screening: ranks first with Bright P@5 = 0.704.
Homology control: the advantage is largest for distant buckets (<50% similarity): R = 0.697 vs. 0.633, 0.575, and 0.408 for baselines, with the strongest overall bright/dark Top-K screening.

Interpretation

Stability-selected features recover band-specific mechanisms:

GFP-like proteins: aromatic stacking and clamp asymmetry.
Red proteins: charge/clamp balance.
Far-red proteins: flexibility risk and large-contact signatures.

Availability

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

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