Research map / Probabilistic and scientific modeling / Molecules and proteins
Molecule Design by Latent Prompt Transformer
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 the Latent Prompt Transformer (LPT), a latent-variable model whose causal Transformer is prompted by a Unet-transformed Gaussian latent that also predicts the property, and optimizes molecules by gradual distribution shifting (SGDS) with software-in-the-loop scoring.
Problem
Molecule design as conditional generation/optimization of molecules with optimal values of a target chemical or biological property (binding affinity, QED, SA) computed by software, under limited queries and out-of-distribution property values.
Approach
A latent vector with a learnable Unet prior serves as a soft prompt for a causal Transformer string generator and also feeds an MLP property regressor; the model is trained by approximate maximum likelihood using short Langevin posterior sampling and then shifted toward higher property values via SGDS, replacing LSTM and latent EBM priors of LEBM.
Key results
- On single-objective ESR1 binding optimization, LPT-SGDS reaches top-1 K_D of 0.004 nM versus 0.03 nM for LEBM-SGDS and 0.72 nM for LIMO.
- On ACAA1, LPT-SGDS top-1 K_D is 0.037 nM versus 0.11 nM for LEBM-SGDS.
- In multi-objective optimization, LPT-SGDS top-1 ESR1 K_D is 0.04 nM versus 0.36 nM for LEBM-SGDS, with comparable QED and SA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LPT-SGDS instead of LEBM-SGDS | ESR1 single-objective molecule design · ESR1 | K_D (nmol/L), best three reported: 0.03 -> 0.004; 0.03 -> 0.005; 0.04 -> 0.014 | Helped |
| Use LPT-SGDS instead of LEBM-SGDS | ACAA1 single-objective molecule design · ACAA1 | K_D (nmol/L), best three reported: 0.11 -> 0.037; 0.11 -> 0.046; 0.12 -> 0.084 | Helped |
| Use LPT-SGDS instead of LEBM-SGDS | ESR1 multi-objective molecule design · ESR1 | K_D (nmol/L): 0.36 -> 0.04; 1.28 -> 0.05 | Helped |
| Use LPT-SGDS instead of LEBM-SGDS | ESR1 multi-objective molecule design · ESR1 | QED: 0.44 -> 0.58; 0.44 -> 0.46 | Mixed |
| Use LPT-SGDS instead of LEBM-SGDS | ESR1 multi-objective molecule design · ESR1 | SA: 3.99 -> 3.46; 3.86 -> 3.24 | Helped |
| Use LPT-SGDS instead of LEBM-SGDS | ACAA1 multi-objective molecule design · ACAA1 | K_D (nmol/L): 4.55 -> 0.18; 5.67 -> 0.21 | Helped |
| Use LPT-SGDS instead of LEBM-SGDS | ACAA1 multi-objective molecule design · ACAA1 | QED: 0.56 -> 0.50; 0.60 -> 0.61 | Mixed |
| Use LPT-SGDS instead of LEBM-SGDS | ACAA1 multi-objective molecule design · ACAA1 | SA: 4.07 -> 4.85; 4.58 -> 4.18 | Mixed |
Limitations
- Relies on black-box software (AutoDock-GPU, RDKit) as ground truth, so docking scores are proxies for real binding affinity
- Multi-objective results use heuristic thresholds (QED>0.4, SA<5.5) for selection
- Gains over LEBM are reported on docking-based metrics only, without wet-lab validation
Experimental setup
Datasets: ZINC
Benchmarks: ESR1 binding affinity optimization ACAA1 binding affinity optimization multi-objective ESR1/ACAA1 with QED and SA
Metrics: top-3 lowest K_D (nM) top-2 average K_D, QED, SA
Baselines: GCPN MolDQN MARS GraphDF LIMO LEBM-SGDS
Builds on: LEBM-SGDS SGDS Langevin dynamics SELFIES
Closest papers by meaning
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.88)
- Navigating Chemical Space with Latent Flows (NeurIPS 2024 · similarity 0.88)
- Learning Disentangled Equivariant Representation for Explicitly Controllable 3D Molecule Generation (AAAI 2025 · similarity 0.87)
- Manipulating 3D Molecules in a Fixed-Dimensional E(3)-Equivariant Latent Space (NeurIPS 2025 · similarity 0.87)
- Return of the Latent Space COWBOYS: Re-thinking the use of VAEs for Bayesian Optimisation of Structured Spaces (ICML 2025 · similarity 0.87)
- Chain-of-Generation: Progressive Latent Diffusion for Text-Guided Molecular Design (NeurIPS 2026 · 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.