Research map / Probabilistic and scientific modeling / Molecules and proteins
FlexMol: A Flexible Toolkit for Benchmarking Molecular Relational Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces FlexMol, an open-source toolkit for dynamically building and benchmarking molecular relational learning models from modular drug/protein encoders and interaction layers, with DTI benchmarks on DAVIS and BIOSNAP.
Problem
Benchmarking molecular relational learning (MRL) is hard because the model space is huge, and existing frameworks like DeepPurpose are limited in input types (mostly protein sequences) and rigid dual-encoder-plus-MLP architectures.
Approach
A modular API that lets users compose preset drug encoders (16), protein sequence encoders (13), protein structure encoders (9), and interaction layers (7) into arbitrary architectures, supporting both sequence and structure inputs and reporting many metrics under a unified trainer.
Key results
- Experiment 1.9 (GCN drug encoder + ESM-GCN protein encoder) achieved the best baseline-level FlexMol results, ROC-AUC 0.916 on DAVIS and 0.913 on BIOSNAP.
- A custom model combining GCN, ESM-GCN, PocketDC encoders and self-attention reached ROC-AUC 0.926 on DAVIS and 0.920 on BIOSNAP, outperforming all listed baselines.
- FlexMol can generate 71,368 distinct DTI models when limiting to four encoders with concatenation only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the PocketDC protein encoder to the custom model (ablation) | Custom Model · Davis | PR-AUC: 0.408 ± 0.016 -> 0.427 ± 0.014 | Helped |
| Add the PocketDC protein encoder to the custom model (ablation) | Custom Model · BIOSNAP | PR-AUC: 0.909 ± 0.001 -> 0.919 ± 0.003 | Helped |
| Add self-attention over the three encoder outputs (ablation) | Custom Model · Davis | PR-AUC: 0.364 ± 0.019 -> 0.427 ± 0.014 | Helped |
| Add self-attention over the three encoder outputs (ablation) | Custom Model · BIOSNAP | PR-AUC: 0.906 ± 0.004 -> 0.919 ± 0.003 | Helped |
| Use ESM features in the protein GCN encoder instead of standard GCN features (ablation) | Custom Model · Davis | PR-AUC: 0.342 ± 0.013 -> 0.427 ± 0.014 | Helped |
| Use ESM features in the protein GCN encoder instead of standard GCN features (ablation) | Custom Model · BIOSNAP | PR-AUC: 0.901 ± 0.002 -> 0.919 ± 0.003 | Helped |
| Add Daylight alongside CNN for drug encoding | Experiment 1.1 CNN drug encoder · BIOSNAP | PR-AUC: 0.874 ± 0.004 -> 0.901 ± 0.004 | Helped |
| Add GCN alongside Morgan for drug encoding | Experiment 1.4 Morgan drug encoder · Davis | PR-AUC: 0.330 ± 0.025 -> 0.388 ± 0.009 | Helped |
| Add GCN alongside Morgan for drug encoding | Experiment 1.4 Morgan drug encoder · BIOSNAP | PR-AUC: 0.879 ± 0.002 -> 0.899 ± 0.006 | Hurt |
| Use Gated-Fusion instead of concatenation | Experiment 1.11 (GAT + PubChem drug encoders, AAC protein encoder) · Davis | ROC-AUC: 0.884 ± 0.007 -> 0.888 ± 0.007 | Helped |
| Use Cross-Attention instead of concatenation | Experiment 1.13 (Transformer drug and protein encoders) · BIOSNAP | PR-AUC: 0.880 ± 0.005 -> 0.905 ± 0.002 | Helped |
Surprising findings
- Adding encoders or interaction layers does not improve every metric or dataset; e.g., Experiment 1.6 improved on DAVIS but dropped on BIOSNAP versus 1.5.
Limitations
- Did not exhaustively explore the combinatorial model space; only 14 example configurations were benchmarked.
- Benchmarks are limited to DTI experiments on two datasets in the main text; DDI and PPI results are deferred to the appendix.
- Improvements from additional encoders or interaction layers are not consistent across metrics and datasets.
Experimental setup
Datasets: DAVIS BIOSNAP
Benchmarks: DAVIS BIOSNAP
Metrics: ROC-AUC PR-AUC
Baselines: LR DNN GNN-CPI DeepDTI DeepDTA DeepConv-DTI MolTrans 3DProt-DTA
Builds on: DeepPurpose MolTrans Therapeutics Data Commons
Closest papers by meaning
- ModuLM: Enabling Modular and Multimodal Molecular Relational Learning with Large Language Models (NeurIPS 2025 · similarity 0.88)
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.87)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.86)
- Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (ICLR 2024 · similarity 0.86)
- 3D Interaction Geometric Pre-training for Molecular Relational Learning (NeurIPS 2025 · similarity 0.86)
- Learning to engineer protein flexibility (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.