Research map / Probabilistic and scientific modeling / Molecules and proteins
GxVAEs: Two Joint VAEs Generate Hit Molecules from Gene Expression Profiles
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 GxVAEs, two joint VAEs (ProfileVAE encodes gene expression profiles; MolVAE generates SMILES conditioned on those features) to generate hit-like molecules for target proteins and diseases.
Problem
De novo generation of hit-like molecules that ignores disease-related cellular context; prior gene-expression-based generators have low validity and weak ligand similarity.
Approach
Two joint VAEs: a ProfileVAE extracts latent features from LINCS/CREEDS gene expression profiles, which condition a bidirectional GRU MolVAE over variant SMILES to generate molecules; disease reversal profiles (multiplied by -1) are fed in for therapeutic generation.
Key results
- GxVAEs validity reached at least 78.0% (HDAC1) across ten target proteins, about three times TRIOMPHE's validity.
- Tanimoto coefficient for AKT1 was 0.85 for GxVAEs versus 0.42 (TRIOMPHE) and 0.32 (ExpressionGAN); SMAD3 reached 0.98.
- Hydrocortisone Tanimoto similarity of 1.0 for atopic dermatitis generated molecules, per the authors' case study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GxVAEs instead of TRIOMPHE to generate candidate molecules from target-protein gene expression profiles | GxVAEs · AKT1, AKT2, AURKB, CTSK, EGFR, HDAC1, MTOR, PIK3CA, SMAD3, TP53 | validity (%): 20.9 -> 88.0 (AKT1); 4.3 -> 91.0 (MTOR) | Helped |
| Use GxVAEs instead of the strongest baseline for target-protein ligand similarity | GxVAEs · AKT1, AKT2, AURKB, CTSK, EGFR, HDAC1, MTOR, PIK3CA, SMAD3, TP53 | Tanimoto coefficient: 0.85 vs 0.42 (AKT1); 0.52 vs 0.69 (MTOR) | Mixed |
Limitations
- Diversity may be constrained by the size of the latent space; a fixed latent vector could limit novelty of generated molecules.
- Therapeutic claims rely on Tanimoto similarity to approved drugs rather than experimental validation of bioactivity.
Experimental setup
Datasets: LINCS L1000 CREEDS
Benchmarks: LINCS L1000 target perturbation profiles (10 target proteins) CREEDS disease profiles (gastric cancer, atopic dermatitis, Alzheimer's disease)
Metrics: validity uniqueness novelty QED Tanimoto coefficient (ECFP4)
Baselines: ExpressionGAN TRIOMPHE DRAGONET
Builds on: TRIOMPHE ExpressionGAN variant SMILES teacher forcing
Closest papers by meaning
- Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces (ICML 2025 · similarity 0.81)
- How Molecules Impact Cells: Unlocking Contrastive PhenoMolecular Retrieval (NeurIPS 2024 · similarity 0.81)
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.81)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.80)
- Pretraining Generative Flow Networks with Inexpensive Rewards for Molecular Graph Generation (ICML 2025 · similarity 0.80)
- Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space (NeurIPS 2025 · similarity 0.80)
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.