Research map / Probabilistic and scientific modeling / Molecules and proteins

MolSpectra: Pre-training 3D Molecular Representation with Multi-modal Energy Spectra

ICLR 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine-learned interatomic potentials · 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.

Pre-trains a 3D molecular encoder (TorchMD-Net) with denoising plus contrastive alignment to a multi-spectrum Transformer (SpecFormer) trained by masked patch reconstruction over UV-Vis, IR, and Raman spectra from QM9S.

Problem

Existing 3D molecular pre-training via denoising models only classical potential energy surfaces and ignores quantized energy level structure visible in experimental spectra, limiting the learned molecular representations.

Approach

Introduce multi-modal energy spectra into 3D pre-training: a single-stream Transformer (SpecFormer) encodes concatenated patches of UV-Vis, IR and Raman spectra with masked patch reconstruction, and an InfoNCE contrastive loss aligns its outputs with the 3D encoder so that only the 3D encoder is needed downstream.

Key results

• On QM9, MolSpectra achieves the best MAE on 8 of 12 properties and beats Coord on 10 of 12 (e.g., HOMO 15.5 vs 17.7 meV, gap 26.8 vs 31.8 meV).
• On MD17 force prediction with 1k training samples, MolSpectra gets the lowest MAE on 6 of 8 molecules (e.g., aspirin 0.099 vs Coord 0.211 kcal/mol/Å).
• Adding MolSpectra to SliDe lowers QM9 MAE on most properties (e.g., HOMO 18.7 to 17.0 meV, alpha 0.050 to 0.043).

Measured effects reported in the paper

Surprising findings

• UV-Vis spectrum contributes least and IR contributes most among the three spectral modalities in the ablation.
• On MD17 the MolSpectra variant is worse than Coord on naphthalene (0.085 vs 0.053) and uracil (0.095 vs 0.074).

Limitations

• Spectral data is limited to about 134k QM9 molecules, restricting scaling studies.
• SpecFormer only handles one-dimensional spectra, not 2D NMR or correlation spectra.
• Built on TorchMD-Net backbone; not tested with newer equivariant architectures such as MACE or Allegro.

Experimental setup

Datasets: PCQM4Mv2 QM9S QM9 MD17

Benchmarks: QM9 MD17

Metrics: MAE

Baselines: SchNet EGNN DimeNet++ PaiNN SphereNet TorchMD-Net Transformer-M SE(3)-DDM 3D-EMGP Coord SliDe

Builds on: Coordinate denoising (Coord) Fractional denoising (Frad) SliDe TorchMD-Net InfoNCE

Closest papers by meaning

• Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.87)
• Mol-AE: Auto-Encoder Based Molecular Representation Learning With 3D Cloze Test Objective (ICML 2024 · similarity 0.87)
• 3D Denoisers Are Good 2D Teachers: Molecular Pretraining via Denoising and Cross-Modal Distillation (AAAI 2025 · similarity 0.87)
• Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.86)
• Multimodal Molecular Pretraining via Modality Blending (ICLR 2024 · similarity 0.85)
• GeoMFormer: A General Architecture for Geometric Molecular Representation Learning (ICML 2024 · similarity 0.85)