Research map / Large language models / Agents, code and math
Domain-Agnostic Molecular Generation with Chemical Feedback
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 MolGen, a SELFIES-based pre-trained molecular language model with domain-agnostic prefix tuning and a rank-loss chemical feedback stage to generate valid, property-optimized molecules across synthetic and natural product domains.
Problem
Pre-trained molecular language models built on SMILES generate many syntactically or chemically invalid strings, focus narrowly on synthetic molecules, and produce "molecular hallucinations" (valid but non-functional molecules) when optimizing properties without external data.
Approach
Pre-train a BART-style seq2seq model on over 100M SELFIES (guaranteed valid), share domain-agnostic prefix vectors across synthetic and natural-product domains, then align sequence likelihoods with property rankings via a rank loss (chemical feedback) combined with label-smoothed cross-entropy.
Key results
- On MOSES synthetic molecules, MolGen reaches 1.000 validity, 0.0015 FCD and 1.000 novelty, beating all baselines in distribution learning.
- For penalized logP maximization, MolGen attains top-1 score 80.30 versus 44.99 for the best baseline (MARS) when output length is unconstrained.
- Constrained optimization at similarity 0.6 yields 12.08 mean p-logP improvement versus 4.98 for GraphAF and 3.78 for RetMol.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply chemical feedback during molecule optimization (ablation) | MolGen · Synthetic molecules | Property scores | Helped |
| Remove chemical feedback (ablation) | MolGen · Synthetic and natural product molecules | Property scores | Hurt |
| Add domain-agnostic molecular prefixes (ablation) | MolGen · Molecular substructure attention | Substructure Attention Level | Helped |
| Use MolGen instead of MolDQN | Synthetic molecule discovery, output length limit · Synthetic molecules | Penalized logP (1st): 11.80 -> 30.51 | Helped |
| Use MolGen instead of MARS | Synthetic molecule discovery, no output length limit · Synthetic molecules | Penalized logP (1st): 44.99 -> 80.30 | Helped |
| Use MolGen instead of LIMO | Molecular distribution learning · Natural product molecules | Frag: .9826 -> .9994 | Helped |
| Use MolGen instead of Chemformer | Molecular distribution learning · Synthetic molecules | FCD: .0061 -> .0015 | Helped |
| Use MolGen instead of LIMO | Binding affinity optimization · ESR1 | Binding affinity (K_D): 0.72 -> 0.13 | Helped |
| Use MolGen instead of LIMO | Binding affinity optimization · ACAA1 | Binding affinity (K_D): 37 -> 3.36 | Helped |
| Use MolGen instead of GraphAF | Constrained p-logP optimization · δ=0.6 | Mean p-logP improvement: 4.98 -> 12.08 | Helped |
| Use MolGen instead of RetMol | Constrained p-logP optimization · δ=0.4 | Mean p-logP improvement: 11.55 -> 12.35 | Helped |
Surprising findings
- LIMO, which also uses SELFIES, has near-zero Scaffold similarity (0.1073) on MOSES, and only 0.0005 on natural products.
Limitations
- Supports only single-target optimization; the chemical feedback paradigm may struggle with multiple objectives.
- Computationally intensive pre-training and fine-tuning.
- Transformer opacity limits interpretability; reaction prediction evaluated only briefly (71.4% accuracy).
Experimental setup
Datasets: ZINC-15 MOSES ZINC250K NPASS PubChem
Benchmarks: MOSES ZINC250K NPASS natural products penalized logP optimization QED optimization ESR1 and ACAA1 docking
Metrics: validity Frag Scaf SNN IntDiv FCD Novelty penalized logP QED binding affinity (K_D)
Baselines: JT-VAE GCPN MolDQN MARS GraphAF GraphDF LIMO Chemformer RetMol RT AAE LatentGAN
Builds on: BART Chemformer prefix-tuning SELFIES JT-VAE
Closest papers by meaning
- Omni-Mol: Multitask Molecular Model for Any-to-any Modalities (NeurIPS 2025 · similarity 0.89)
- Foundation Molecular Grammar: Multi-Modal Foundation Models Induce Interpretable Molecular Graph Languages (ICML 2025 · similarity 0.88)
- Mol-LLaMA: Towards General Understanding of Molecules in Large Molecular Language Model (NeurIPS 2025 · similarity 0.88)
- MolTailor: Tailoring Chemical Molecular Representation to Specific Tasks via Text Prompts (AAAI 2024 · similarity 0.87)
- BindGPT: A Scalable Framework for 3D Molecular Design via Language Modeling and Reinforcement Learning (AAAI 2025 · similarity 0.87)
- MolVision: Molecular Property Prediction with Vision Language Models (NeurIPS 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.