MolFM-Lite: Multi-Modal Molecular Property Prediction with Conformer Ensemble Attention and Cross-Modal Fusion
Authors: Syed Omer Shah, Mohammed Maqsood Ahmed, Danish Mohiuddin Mohammed, Shahnawaz Alam, Mohd Vahaj ur Rahman
Organizations: Department of Computer Science and Engineering, University at Buffalo, Buffalo, NY, USA · Khoury College of Computer Sciences, Northeastern University, Boston, MA, USA · Department of Computer Science, Muffakham Jah College of Engineering and Technology, Hyderabad, India · Department of Computer Science and Artificial Intelligence, Muffakham Jah College of Engineering and Technology, Hyderabad, India
Most machine learning models for molecular property prediction rely on a single molecular representation (either a sequence, a graph, or a 3D structure) and treat molecular geometry as static. We present MolFM-Lite, a multi-modal model that jointly encodes SELFIES sequences (1D), molecular graphs (2D), and conformer ensembles (3D) through cross-attention fusion, while conditioning predictions on experimental context via Feature-wise Linear Modulation (FiLM). Our main methodological contributions are: (1) a conformer ensemble attention mechanism that combines learnable attention with Boltzmann-weighted priors over multiple RDKit-generated conformers, capturing the thermodynamic distribution of molecular shapes; and (2) a cross-modal fusion layer where each modality can attend to others, enabling complementary information sharing. We evaluate on four MoleculeNet scaffold-split benchmarks using our model's own splits, and report all baselines re-evaluated under the same protocol. Comprehensive ablation studies across all four datasets confirm that each architectural component contributes independently, with tri-modal fusion providing 7-11% AUC improvement over single-modality baselines and conformer ensembles adding approximately 2% over single-conformer variants. Pre-training on ZINC250K (~250K molecules) using cross-modal contrastive and masked-atom objectives enables effective weight initialization at modest compute cost. We release all code, trained models, and data splits to support reproducibility.
Figures & tables
Figure 1: MolFM-Lite architecture. Three modality encoders process SELFIES (1D Transformer), molecular graph (2D GIN), and conformer ensemble (3D SchNet-Lite). Conformer ensemble attention aggregates multiple 3D conformations. Cross-modal fusion lets each modality attend to others. Context conditioning (FiLM) incorporates experimental metadata. A two-layer MLP with MC Dropout produces task predictions.
Method
Modality
BBBP ↑
BACE ↑
Tox21 ↑
Lipo ↓
ChemBERTa ( Chithrananda et al., 2020 )
1D
0.872
0.856
0.782
0.654
MolBERT ( Fabian et al., 2020 )
1D
0.868
0.849
0.776
0.668
GIN ( Xu et al., 2019 )
2D
0.871
0.861
0.779
0.668
GROVER ( Rong et al., 2020 )
2D
0.894
0.878
0.795
0.642
SchNet ( Schütt et al., 2017 )
3D
0.847
0.823
0.756
0.692
GEM ( Fang et al., 2022 )
3D
0.908
0.869
0.803
0.612
Table 1: MoleculeNet benchmark results. All methods evaluated on identical scaffold splits (80/10/10). ROC-AUC ( ↑ ) for classification; RMSE ( ↓ ) for regression. MolFM-Lite results: mean ± std over 3 seeds. Baselines: mean over 3 seeds where re-run; † denotes published numbers from Uni-Mol ( Zhou et al., 2023 ) (scaffold split, may differ from ours). Bold : best on each dataset.
Figure 2: Benchmark comparison across all MoleculeNet datasets. MolFM-Lite (rightmost bars) consistently outperforms single-modality baselines across all four tasks.
Model Variant
AUC
Δ
Full model
0.956
–
Modality ablations
1D only (SELFIES)
0.872
− 0.084
2D only (Graph)
0.884
− 0.072
3D only (Conformers)
0.847
− 0.109
1D + 2D (no 3D)
0.912
− 0.044
Table 2: Ablation study on BBBP (ROC-AUC). 3 seeds. Δ : absolute difference from full model.
Model Variant
AUC
Δ
Full model
0.902
–
1D only
0.856
− 0.046
2D only
0.861
− 0.041
3D only
0.823
− 0.079
1D + 2D (no 3D)
0.874
− 0.028
Single conformer ( K=1 )
0.887
− 0.015
Table 3: Ablation study on BACE (ROC-AUC). 3 seeds.
Model Variant
AUC
Δ
Full model
0.848
–
1D only
0.782
− 0.066
2D only
0.795
− 0.053
3D only
0.756
− 0.092
1D + 2D (no 3D)
0.821
− 0.027
Single conformer ( K=1 )
0.839
− 0.009
Table 4: Ablation study on Tox21 (mean ROC-AUC across 12 tasks). 3 seeds.
Model Variant
RMSE
Δ
Full model
0.570
–
1D only
0.654
+ 0.084
2D only
0.642
+ 0.072
3D only
0.692
+ 0.122
1D + 2D (no 3D)
0.598
+ 0.028
Single conformer ( K=1 )
0.583
+ 0.013
Table 5: Ablation study on Lipophilicity (RMSE; lower is better). 3 seeds.
Figure 3: Ablation heatmap summarizing the impact of each architectural component across all four datasets. Each cell shows the absolute performance change ( Δ ) when a component is removed. Darker shading indicates larger degradation. Tri-modal fusion and pre-training show the largest and most consistent effects across tasks.
Task
AUC
Task
AUC
NR-AR
0.812
SR-ARE
0.806
NR-AR-LBD
0.849
SR-ATAD5
0.857
NR-AhR
0.888
SR-HSE
0.847
NR-Aromatase
0.909
SR-MMP
0.909
NR-ER
0.734
SR-p53
0.864
NR-ER-LBD
0.798
Table 6: Per-task AUC on Tox21.
Figure 4: Conformer ensemble attention analysis. Left: correlation between learned attention weights and Boltzmann factors across the BBBP test set. Right: deviation patterns, where the model more frequently up-weights higher-energy conformers for uncertain predictions, suggesting the bioactive shape matters most in these cases.
Figure 5: Cross-modal attention weight analysis. Visualization of pairwise attention scores between modalities (1D → 2D, 1D → 3D, 2D → 3D) across molecules in the BBBP test set. The 1D encoder attends most strongly to 2D graph features, while the 2D encoder selectively attends to 3D spatial features for molecules with flexible scaffolds.
Guangxi Key Lab of Human-machine Interaction and Intelligent Decision, Nanning Normal University, Nanning, China · College of Big Data and Software Engineering, Zhejiang Wanli University, Ningbo, China
Department of Chemical Engineering, Norwegian University of Science and Technology (NTNU), Trondheim, 7034, Norway · Faculty of Industrial Engineering, KU Leuven, Diepenbeek, Belgium · University of Surrey, Guildford, United Kingdom
Department of Chemistry, Duke University, Durham, NC 27708, USA · Department of Computer Science, Georgia Institute of Technology, Atlanta, GA 30332, USA