Research map / Point clouds / Point cloud classification
nach0-pc: Multi-task Language Model with Molecular Point Cloud Encoder
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 nach0-pc, a T5-based encoder-decoder LM with a molecular point cloud encoder and SMILES+XYZ text output, pre-trained by reconstructing dropped or blurred BRICS fragments, and evaluated on six spatial molecular generation tasks.
Problem
Existing chemistry LMs work on SMILES/SELFIES strings that lack 3D spatial information, while text-based 3D formats (PDB, XYZ) are very long and lack explicit connectivity, so LMs struggle with spatial molecular generation.
Approach
Encode molecules as order-invariant point clouds with a point cloud encoder (summed token embeddings, distance-based relative attention biases via Scalar Sinusoidal Embeddings), output SMILES followed by XYZ coordinates in SMILES order, and pre-train by reconstructing BRICS fragments removed or blurred from point clouds.
Key results
- Single-task nach0-pc reaches 57.7% recall COV (avg) on GEOM-Drugs conformation generation, above nach0 (54.0%) and below Torsional Diffusion (72.7%).
- Multi-task nach0-pc achieves 96.2% novelty and 0.781 diversity on GEOM-Drugs distribution learning, the best diversity among compared models.
- On pocket-conditioned generation nach0-pc reaches 97.63% validity (multi-task) versus TargetDiff's 90.36%, but has lower diversity (0.32/0.40 vs 0.72).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use single-task rather than multi-task fine-tuning (ablation) | nach0-pc · GEOM-DRUGS conformer generation | Recall COV Avg: 50.1 -> 57.7 | Helped |
| Use single-task rather than multi-task fine-tuning (ablation) | nach0-pc · ZINC linker design | Validity: 81.6% -> 89.7% | Helped |
| Use single-task rather than multi-task fine-tuning (ablation) | nach0-pc · CrossDocked2020 pocket-conditioned generation | Validity: 91.78% -> 89.82% | Hurt |
| Use multi-task rather than single-task fine-tuning (ablation) | nach0-pc · GEOM-DRUGS spatial molecular distribution learning | Diversity: 0.734 -> 0.781 | Helped |
| Use point cloud input rather than text-only generation | nach0-pc vs nach0 · ZINC linker design | Validity: 55.2% -> 81.6% | Helped |
| Use point cloud input rather than text-only generation | nach0-pc vs nach0 · ZINC linker design | Uniqueness: 19.9% -> 27.6% | Helped |
| Use point cloud input rather than text-only generation | nach0-pc vs nach0 · GEOM-DRUGS conformer generation | Recall COV Avg: 54.0 -> 50.1 | Hurt |
| Use nach0-pc instead of Torsional Diffusion | conformation generation · GEOM-DRUGS | Recall COV Avg: 72.7 -> 57.7 | Hurt |
| Use nach0-pc instead of DiffDec | scaffold decoration · CrossDocked | QVina Score: -8.25 -> -8.123 | Hurt |
| Use nach0-pc instead of TargetDiff | pocket-conditioned generation · CrossDocked2020 | Vina Dock Avg: -7.80 -> -6.52 | Hurt |
Surprising findings
- Text-only OpenLLaMA slightly outperforms nach0-pc on some bond and ring distribution metrics, which the authors attribute to decoder-only parameter use.
- nach0-pc has a much lower diversity than diffusion baselines on pocket-conditioned generation despite high validity.
Limitations
- Inherits SMILES limitations: many SMILES per molecule and reliance on the nach0 SMILES-pretrained backbone.
- Quadratic attention cost limits scaling to long chemical sequences.
- Academic datasets cover only a small part of chemical space, with no testing on novel chemistry.
Experimental setup
Datasets: ZINC MOSES GEOM-Drugs CrossDocked2020
Benchmarks: GEOM-Drugs ZINC linker design CrossDocked scaffold decoration CrossDocked pocket-conditioned generation MOSES shape-conditioned
Metrics: validity uniqueness novelty diversity QED SA JS divergence COV AMR Vina score
Baselines: MolDiff nach0 OpenLLaMA GeoDiff Torsional Diffusion DiffLinker DiffDec TargetDiff
Builds on: nach0 T5 OpenLLaMA BRICS
Closest papers by meaning
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.89)
- 3D-MolT5: Leveraging Discrete Structural Information for Molecule-Text Modeling (ICLR 2025 · similarity 0.88)
- Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling (ICML 2025 · similarity 0.88)
- KnowMol: Advancing Molecular Large Language Models with Multi-Level Chemical Knowledge (NeurIPS 2025 · similarity 0.87)
- Towards 3D Molecule-Text Interpretation in Language Models (ICLR 2024 · similarity 0.87)
- NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation (ICLR 2025 · similarity 0.87)
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.