Multi-RF Fusion with Multi-GNN Blending for Molecular Property Prediction

Zacharie Bugaud

cs.AIMar 21, 2026

Multi-RF Fusion with Multi-GNN Blending for Molecular Property Prediction

Authors: Zacharie Bugaud

Abstract

Multi-RF Fusion achieves a test ROC-AUC of 0.8476 +/- 0.0002 on ogbg-molhiv (10 seeds), placing #1 on the OGB leaderboard ahead of HyperFusion (0.8475 +/- 0.0003). The core of the method is a rank-averaged ensemble of 12 Random Forest models trained on concatenated molecular fingerprints (FCFP, ECFP, MACCS, atom pairs -- 4,263 dimensions total), blended with deep-ensembled GNN predictions at 12% weight. Two findings drive the result: (1) setting max_features to 0.20 instead of the default sqrt(d) gives a +0.008 AUC gain on this scaffold split, and (2) averaging GNN predictions across 10 seeds before blending with the RF eliminates GNN seed variance entirely, dropping the final standard deviation from 0.0008 to 0.0002. No external data or pre-training is used.

Figures & tables

Config	Count	Feature Set	max_features	Criterion
Primary	5	fcfp_ext (4,263-d)	0.20	entropy
Secondary	3	fcfp_ext (4,263-d)	0.18	entropy
Tertiary	2	fcfp_ext (4,263-d)	0.22	entropy
Crit. diversity	1	fcfp_ext (4,263-d)	0.20	log_loss
Feat. diversity	1	fcfp_ext_2k (6,311-d)	0.20	entropy

Table 1: Random Forest model configurations in the Multi-RF ensemble. All models share n_estimators

=

20,000, min_samples_leaf

=

3, and class_weight

=\{0{:}1,1{:}25\}

Method	Test ROC-AUC	External Data
Multi-RF Fusion + Multi-GNN	$\mathbf{0.8476\pm 0.0002}$	No
HyperFusion ( Zhang, 2024 )	$0.8475\pm 0.0003$	No
PAS+FPs ( Wei et al., 2021 )	$0.8420\pm 0.0015$	No

Table 2: Comparison with top methods on the ogbg-molhiv leaderboard.

Seed	42	179	316	453	590	727	864	1001	1138	1275
AUC	0.8480	0.8479	0.8474	0.8478	0.8476	0.8477	0.8474	0.8474	0.8474	0.8476

Table 3: Per-seed test ROC-AUC for the full method (Multi-RF + GIN-VN @ 6% + GIN-VN-deep @ 6%).

Component	Test ROC-AUC	$\boldsymbol{\Delta}$
Full method (Multi-RF + 2 GNNs)	$0.8476\pm 0.0002$	—
Single GNN (GIN-VN @ 12%)	$0.8475\pm 0.0002$	$-0.0001$
RF only (no GNN)	$0.8461\pm 0.0002$	$-0.0015$
Per-seed GNN @ 7% (no deep ensemble)	$0.8467\pm 0.0011$	$-0.0009$
max_features=sqrt (conventional)	$0.8396\pm 0.0005$	$-0.0080$

Table 4: Ablation study showing the contribution of each component.

Architecture	Individual Test AUC	Cross-Seed Test AUC	Contribution
GIN-VN (5L, 300d)	$0.7657\pm 0.0109$	0.7930	Primary
GIN-VN-deep (8L, 256d)	$0.7613\pm 0.0101$	0.7897	Complementary
GIN-VN-sum (5L, 300d)	$0.7583\pm 0.0204$	0.7808	Marginal
FP-MLP (3L, 1024d)	$0.7668\pm 0.0160$	0.7853	Redundant with RF
GCN-VN (5L, 300d)	$0.7407\pm 0.0188$	0.7626	Harmful
GIN-VN-large (5L, 600d)	$0.7558\pm 0.0148$	0.7809	Marginal

Table 5: GNN architecture comparison. “Individual” is mean

\pm

std over 10 seeds; “Cross-Seed” is the AUC of averaged predictions.

Hyperparameter	Search Range
max_features	$\{0.05,0.08,0.10,0.12,0.15,0.18,\mathbf{0.20}^{\ast},0.22,0.25,0.30\}$
class_weight (pos)	$\{10,15,20,\mathbf{25}^{\ast},30,50\}$
min_samples_leaf	$\{1,\mathbf{3}^{\ast},5\}$
n_estimators	$\{5000,10000,15000,\mathbf{20000}^{\ast}\}$
criterion	$\{\text{gini},\mathbf{entropy}^{\ast},\text{log\_loss}\}$
GNN blend weight (per arch.)	$\{0.03,0.04,0.05,\mathbf{0.06}^{\ast},0.07,0.08,0.10\}$

Table 6: Hyperparameter search ranges and selected values (

\ast

Molecular Property Prediction

Molecular property prediction estimates chemical, physical, and biological properties from molecular structure and context.

See all 197 papers

MolGraphBench: A Benchmark of GNN Architectures for Molecular Regression Tasks
Feb 24, 2026Rajan, Ishaan GuptaGraph Neural Networks Molecular Property Prediction
School of Interdisciplinary Research (SIRe), IIT Delhi, New Delhi-110016, India · Department of Biochemical Engineering and Biotechnology (DBEB), IIT Delhi, New Delhi-110016, India
Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints
Sep 14, 2026Gwang-Hyeon Yun, Jong-Hoon Park, Bing Hu +3Graph SSL Hierarchical Representation Learning
Yonsei University · Wonju-si, Republic of Korea · University of Waterloo +1
MMGNN: Multi-level, multi-color graph neural networks for molecular property prediction
Jun 18, 2026Trung Nguyen, Duc Duy NguyenMessage Passing Neural Networks Molecular Property Prediction
The Bredesen Center, University of Tennessee, Knoxville, TN 37996, USA · Department of Mathematics, University of Tennessee, Knoxville, TN 37996, USA

Feb 24, 2026J/K move · Enter open · S save

MolGraphBench: A Benchmark of GNN Architectures for Molecular Regression Tasks

Rajan, Ishaan Gupta

School of Interdisciplinary Research (SIRe), IIT Delhi, New Delhi-110016, India · Department of Biochemical Engineering and Biotechnology (DBEB), IIT Delhi, New Delhi-110016, India

Graph Neural Networks Molecular Property Prediction Molecular Representation Learning

Open paper AskPDF arXiv

Molecules are often represented as SMILES strings, which can be readily converted to hand-crafted descriptors or fingerprints (FP) for molecular property prediction. Research has demonstrated that SMILES can be converted to molecular graphs

G = (V, E)

, with atoms as nodes

(V)

and bonds as edges

(E)

. These molecular graphs can subsequently be used to train graph neural networks (GNN) models. Despite the recent surge in application of GNN (existing and novel architectures) for molecular property prediction, a rigorous benchmark is still lacking. We propose MolGraphBench, a comprehensive benchmark of four commonly used GNN models for molecular property prediction. Benchmarking results demonstrate graph convolutional network (GCN) and graph isomorphism networks (GIN) as the optimal GNN architectures for molecular graph regression tasks, based on absolute performance, training efficiency, transfer learning and prediction quality. The study also indicates the non-complementary nature of molecular fingerprints in the fusion (GNN-FP) framework. Furthermore, our GNN models achieved performance superior or comparable performance to current state-of-the-art GNN baselines across three datasets (GCN with RMSE of

0.518

on B3DB, GIN-FP with RMSE of

1.022

on FreeSolv and GIN with MAE of

63.783

on RT datasets). Findings from this study indicate that type of GNN-layer, should be treated as a tunable hyperparameter rather than a fixed design choice to achieve superior performance.

Abstract

Molecular Property Prediction

Explore similar work

MolGraphBench: A Benchmark of GNN Architectures for Molecular Regression Tasks

Multi-View Molecular Representation Learning with Hierarchical Graphs and Contextualized Fingerprints

MMGNN: Multi-level, multi-color graph neural networks for molecular property prediction