Diffusion models have emerged as a promising class of generative models for molecular design because of their ability to learn complex chemical distributions and support conditional or guided generation. In this work, we investigate a latent diffusion pipeline for property-guided small-molecule generation. The proposed framework represents canonicalized SMILES strings using a frozen ChemBERTa encoder, trains a Transformer-based decoder to reconstruct SELFIES strings from molecular latent embeddings, and learns an unconditional denoising diffusion model in the ChemBERTa latent space. To enable property-directed generation, a differentiable latent property predictor is used to guide the diffusion sampling process. Generated SELFIES strings are converted to SMILES, validated using RDKit, and evaluated using molecular validity, uniqueness, novelty, and predicted property scores. We study guidance toward two drug-relevant binary endpoints: blood--brain barrier permeability and inhibition of human β-secretase 1. The results suggest that latent diffusion over pretrained molecular embeddings can generate chemically valid and novel molecules while enabling controllable steering toward desired molecular properties.
Figures & tables
Figure 1: Property-guided molecular generation framework. (a) Molecules are encoded into the ChemBERTa-2 latent space, optimized by classifier-guided reverse diffusion, decoded to molecular strings, and validated with RDKit. (b) The transformer decoder reconstructs SELFIES from perturbed latent embeddings. (c) The diffusion model predicts noise added to latent representations. (d) The property classifier scores generated molecules, while the guidance head steers sampling toward target properties.
Table 1: Datasets used for generative and property-model training, including their sources and data-splitting strategies.
Metric
Abbrev.
Direction
Definition
Validity
Validity
↑
Fraction of generated molecules that pass RDKit parsing, sanitization, and chemistry-consistency checks.
Uniqueness
Unique
↑
Fraction of valid generated molecules that are unique after canonicalization.
Novelty
Novel
↑
Fraction of valid generated molecules that are absent from the canonicalized training set.
Mean property score
Mean score
↑
Average predicted target-property probability assigned by the trained classifier gϕ to valid generated molecules.
Success rate
Success
↑
Fraction of valid generated molecules with predicted property score greater than 0.5.
Strict success rate
Strict
↑
Fraction of valid generated molecules with predicted property score greater than 0.8.
Table 2: Evaluation metrics for generated molecules. Arrows indicate the preferred direction.
Module
Epochs
Train loss
Val. loss
Train metric
Val. metric
Decoder
10
0.220
0.297
token acc. 0.951
token acc. 0.935
Diffusion
10
0.0789
0.0797
n/a
n/a
Table 3: Training metrics for the shared decoder and latent diffusion model used in both BBBP and BACE-1 guidance experiments. Decoder performance is reported using token accuracy, while diffusion performance is reported using the DDPM noise-prediction loss.
Model
BBBP AUC
BACE-1 AUC
ChemBERTa-2 [ 21 ]
0.701
0.809
MolPLA [ 29 ]
0.702
0.805
GROVER [ 30 ]
0.936
0.878
MolCLR [ 31 ]
0.736
0.890
GraphMVP [ 32 ]
0.724
0.812
MMSA [ 33 ]
0.745
0.824
Table 4: Contextual comparison of property-prediction ROC-AUC with representative molecular property predictors. Higher values are better.
Guidance
Validity
Unique
Novel
Success
Strict
Mean score
Opt.
Int. div.
Scaf. div.
g=0.0
1.000
0.999
1.000
0.892
0.668
0.811
0.117
0.899
0.876
g=0.5
1.000
0.998
1.000
0.915
0.707
0.830
0.136
0.900
0.864
g=1.0
1.000
0.998
1.000
0.902
0.708
0.826
0.132
0.901
0.839
g=2.0
1.000
1.000
1.000
0.934
0.706
0.838
0.144
0.897
0.868
g=3.0
1.000
1.000
1.000
0.947
0.785
0.864
0.169
0.897
0.851
g=5.0
1.000
0.999
1.000
0.947
0.800
0.874
0.180
0.896
0.849
Table 5: BBBP generation results across guidance strengths, obtained from generation using three random sampling seeds.
Table 6: BACE-1 generation results across guidance strengths, obtained from generation using three random sampling seeds.
Figure 3: Mean descriptor and distribution-matching trends across guidance strengths for BBBP and BACE-1. The figure reports average molecular weight, logP, TPSA, and QED, together with descriptor-space KL divergence and maximum mean discrepancy (MMD). Lower KL divergence and MMD indicate closer agreement with the reference distribution.
Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.
Goal-directed molecular generation requires satisfying heterogeneous constraints such as protein--ligand compatibility and multi-objective drug-like properties, yet existing methods often optimize these constraints in isolation, failing to reconcile conflicting objectives (e.g., affinity vs. safety), and struggle to navigate the non-differentiable chemical space without compromising structural validity. To address these challenges, we propose CAGenMol, a condition-aware discrete diffusion framework over molecular sequences that formulates molecular design as conditional denoising guided by heterogeneous structural and property signals. By coupling discrete diffusion with reinforcement learning, the model aligns the generation trajectory with non-differentiable objectives while preserving chemical validity and diversity. The non-autoregressive nature of diffusion language model further enables iterative refinement of molecular fragments at inference time. Experiments on structure-conditioned, property-conditioned, and dual-conditioned benchmarks demonstrate consistent improvements over state-of-the-art methods in binding affinity, drug-likeness, and success rate, highlighting the effectiveness of our framework.
Yanting Li, Zhuoyang Jiang, Enyan Dai +3
The Hong Kong University of Science and Technology (Guangzhou) · Jinan University, Guangzhou
We present SynLaD, a latent diffusion framework for small-molecule generation that unifies ligand-based drug design objectives (what to make) with synthetic accessibility (how to make it). Current models typically optimize one objective at the expense of the other, creating a bottleneck for discovering high-scoring and synthesizable molecules. SynLaD combines reaction-constrained generation with pharmacophore-conditioned 3D design by learning a latent space that decodes to both 3D structures and synthesis pathways. An encoder maps molecules to a latent representation used by two decoder heads: (i) a geometric head that reconstructs atom types and coordinates and (ii) an autoregressive synthesis head that outputs synthetic routes in a serialized, reaction-based notation. A diffusion transformer generates novel latents in the learned space, conditioned on pharmacophore profiles. Across analogue generation tasks for bioactive ligands, SynLaD outperforms existing baselines in synthesizable and diverse hit generation, demonstrating that a single model can produce shape-aligned molecules with feasible synthesis plans.
Miruna Cretu, John Bradshaw, Patricia Suriana +6
University of Cambridge, Cambridge, UK · Prescient Design (AI for Drug Discovery), Genentech, South San Francisco, USA · Work done during an internship at Prescient Design
Bridging molecular structures and natural language is essential for controllable design. Autoregressive models struggle with long-range dependencies, while standard diffusion processes apply uniform corruption across positions, which can distort structurally informative tokens. We present BiMol-Diff, a unified diffusion framework for the paired tasks of text-conditioned molecule generation and molecule captioning. Our key component is a token-aware noise schedule that assigns position-dependent corruption based on token recovery difficulty, preserving harder-to-recover substructures during the forward process. On ChEBI-20 and M3-20M, BiMol-Diff improves molecule reconstruction with a 15.4% relative gain in Exact Match and achieves strong captioning results, attaining best BLEU and BERTScore among compared baselines. These results indicate token-aware noising improves fidelity in molecular structure-language modelling.