Research map / Large language models / Agents, code and math

ModuLM: Enabling Modular and Multimodal Molecular Relational Learning with Large Language Models

NeurIPS 2025 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks · 0 citations (OpenAlex)

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.

Proposes ModuLM, a modular framework for building LLM-based molecular relational learning models from interchangeable 1D/2D/3D molecular encoders, interaction layers, and LLM backbones, and benchmarks configurations on DDI, SSI, and CSI datasets.

Problem

Benchmarking LLM-based molecular relational learning (MRL) is hard because the model space (LLM backbones, molecular structure encoders, interaction layers) is large, and no existing LLM framework supports flexible molecular input formats (1D/2D/3D) with dynamic architecture switching.

Approach

A JSON-configured modular framework that assembles LLM-based MRL models from interchangeable 2D graph encoders, 3D conformation encoders, interaction layers, alignment modules, and LLM backbones, plus three incremental pretraining text strategies, enabling over 50,000 configurations and fair comparison.

Key results

• The best 3D-encoder configuration (1.8: Uni-Mol with DeepSeek-1.5B) reaches 0.966 AUC-ROC on ChChMiner and 0.190 MAE on CombiSolv, outperforming MolTC (0.964 and 0.237).
• A custom model built on ModuLM (Uni-Mol with reweighted attention, interaction, and CoT) reaches 0.968 accuracy on ChChMiner and 0.288 RMSE on CompSol, improving on configuration 1.8.
• Larger backbones do not consistently outperform smaller ones; DeepSeek-series models often perform better.

Measured effects reported in the paper

Surprising findings

• Larger LLM sizes did not necessarily yield better MRL performance; the authors hypothesize that stronger generalization limits task-specific fine-tuning adaptation.
• Several configurations (e.g., 1.8 on DDI) outperform the MolTC baseline on most DDI metrics, while MolTC remains best on ZhangDDI AUC-ROC and DeepDDI AUC-ROC.

Limitations

• Benchmark covers only a subset of the 50,000+ possible configurations; only 15 models are evaluated.
• Results are reported with direct inference for most experiments, and the custom-model gains are small relative to standard deviations.
• Several reported baselines are reproduced under the authors' setup rather than from original papers, and dataset descriptions and split details are sparse.

Experimental setup

Datasets: DrugBank ZhangDDI ChChMiner DeepDDI TWOSIDES Chromophore MNSol CompSol Abraham CombiSolv FreeSolv CombiSolv-QM

Benchmarks: ChChMiner ZhangDDI DeepDDI DrugBank TWOSIDES FreeSolv CompSol CombiSolv MNSol Abraham Chromophore

Metrics: AUC-ROC Accuracy MAE RMSE MSE R2 PR-AUC

Baselines: ChemT5 MolCA MolT5 MolTC Galactica

Builds on: MolTC Uni-Mol LoRA ReactionT5

Closest papers by meaning

• FlexMol: A Flexible Toolkit for Benchmarking Molecular Relational Learning (NeurIPS 2024 · similarity 0.88)
• Improving Large Molecular Language Model via Relation-aware Multimodal Collaboration (AAAI 2026 · similarity 0.88)
• 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.88)
• Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.88)
• KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.87)
• LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.87)