Mol-LLM: Multimodal Generalist Molecular LLM

Mol-LLM is a multimodal generalist molecular large language model for chemistry that jointly uses molecular 1D sequences and 2D molecular graphs to solve a wide range of molecular tasks in a single unified framework. It introduces Molecular structure Preference Optimization (MolPO) to force the LLM to prefer correct molecular graphs over perturbed ones, resolving the “graph-bypass” issue common in prior multimodal molecular LLMs.

Model summary

• Backbone: Mistral-7B-Instruct-v0.3.
• Modalities:
  • Text (natural language instructions).
  • 1D molecular sequences (SELFIES; SMILES supported via translation).
  • 2D molecular graphs encoded by a hybrid GNN (GINE + TokenGT).
• Architecture: Mol-LLM uses a BLIP-2–style architecture where Q-Former (32 query tokens) projects graph embeddings into the LLM token space.

  • LLM:

    • Mistral-7B-Instruct-v0.3 as text backbone.
    • Extended tokenizer with SELFIES and numeric tokens, plus task tags for heterogeneous outputs (discrete labels, floats, descriptions).

  • Hybrid graph encoder:

    • GINE for local structural patterns.
    • TokenGT (transformer-based) for global structural dependencies and large graphs.
    • Both encoders produce graph and node (and edge) embeddings; concatenated embeddings are fed into the Q-Former.

  • Q-Former:

    • 5-layer SciBERT-style transformer with 32 learnable queries.
    • Cross-attends to graph embeddings and outputs fixed-length tokens appended after SELFIES tokens in the LLM input.
    • Selected over an MLP projector due to better alignment and graph-token efficiency.
• Tokenizer extensions: 3K SELFIES tokens, numeric tokens, and task tags for [SELFIES], [BOOLEAN], [FLOAT], [DESCRIPTION], and reaction-direction symbols.
• Training data: ~3.3M instruction-tuning examples over 27 tasks, with ~40K held-out test instances.

Mol-LLM is positioned as a state-of-the-art or comparable generalist molecular LLM on the most comprehensive benchmark suite evaluated so far, including out-of-distribution (OOD) settings.

Intended use

Mol-LLM is intended to solve molecular tasks via a single multitask model.

Supported task families:

• Reaction prediction:

  • Forward synthesis (product prediction, FS)
  • Retrosynthesis (reactant prediction, RS)
  • Reagent prediction (RP)

• Property prediction:

  • Regression: LogS, LogD, HOMO, LUMO, HOMO–LUMO gap
  • Classification: BACE, BBBP, ClinTox, HIV, SIDER

• Text–molecule tasks:

  • Description-guided molecule generation
  • Molecule captioning
  • IUPAC/SELFIES/formula translation as auxiliary tasks