Diffusion models have become one of the most powerful paradigms in generative machine learning, famous for their success in image synthesis and increasingly prominent in chemistry and drug discovery. The core idea is elegantly counterintuitive. During training, structured data such as a molecular graph is progressively corrupted by adding random noise until it becomes indistinguishable from static. A neural network is then taught to reverse that process, learning to strip away noise step by step until a coherent structure emerges from pure randomness. When applied to molecules, the reverse diffusion process must reconstruct both the atoms and the bonds connecting them, turning formless noise into valid, diverse, and novel chemical graphs. Every single reverse step depends on accurate denoising, which makes the denoising network the linchpin of the entire enterprise.

The difficulty lies in what that denoising network must accomplish simultaneously. At the local level, it has to recover chemical connectivity: which atoms are bonded to which neighbors, and with what bond types. At the global level, it must grasp the topology of the whole graph, the overall scaffold that determines the molecule’s shape and, ultimately, its biological behavior. Convolution-style message passing networks excel at capturing local neighborhoods, while spectral graph methods, which operate on the eigenstructure of the graph Laplacian, are naturally suited to describing global patterns. Most existing models commit to one perspective, leaving the complementary one underexplored, and that compromise shows up in the quality of generated molecules.

Longbin Sun and Xiujuan Lei of Shaanxi Normal University, together with Mei Ma of Qinghai Normal University, tackled this trade-off head-on. Their DFDM model combines two parallel branches inside the denoising network. The first is a spatial branch built on GINE, a graph isomorphism network variant widely used for learning from molecular graphs through neighbor message passing. The second is a Chebyshev spectral branch, which leverages Chebyshev polynomial approximations of graph convolution to capture information distributed across the entire molecular graph. Rather than forcing a fixed marriage between the two, DFDM fuses their outputs using weights that change from layer to layer and are guided by the signal-to-noise ratio at each step of the reverse diffusion process.

This signal-to-noise-ratio guidance is the conceptual heart of the method. Early in the reverse process, when the molecular graph is still mostly noise, the amount of genuine signal is tiny; late in the process, signal dominates. Intuitively, when almost nothing is known about the molecule, broad global information is more useful than fine local detail, whereas near the end of generation, when atomic identities and bonds are crystallizing, precise local connectivity becomes paramount. By tying the fusion weights to the measured noise level, DFDM lets the model decide dynamically which branch to trust at every stage and in every layer, rather than committing to a static balance that may be wrong for most of the generation trajectory.

The team evaluated DFDM on two of the standard benchmarks in generative molecular chemistry: QM9, a dataset of roughly one hundred thirty-four thousand small molecules with quantum chemical properties, and ZINC250k, a quarter-million subset of drug-like compounds drawn from the ZINC database. To guard against lucky runs, they repeated the full generation-and-evaluation pipeline three times with independent random seeds. Across these runs, DFDM achieved the lowest mean Fréchet ChemNet Distance among the compared methods, a metric that measures how closely the distribution of generated molecules resembles that of real molecules as judged by a pretrained chemical neural network. Lower values indicate generated chemistry that statistically looks like real chemistry, spanning both diversity and plausibility.

Just as striking was DFDM’s performance on validity. The model achieved the highest mean validity without correction, meaning the largest fraction of its generated outputs were chemically sound molecules exactly as produced, without any post-hoc repair or filtering tricks. In molecular generation, validity is a perennial headache; many powerful models produce graph structures that violate basic valence rules and must be salvaged by external cheminformatics tools. A model that natively generates valid molecules most of the time saves computation and, more importantly, indicates that its internal representation of chemistry is genuinely coherent rather than superficially plausible.

Ablation experiments, in which components of the model are systematically removed, provided insight into why the dynamic approach works. On both datasets, replacing the dynamic fusion with fixed weights, which let the spatial and spectral branches contribute equally regardless of noise level, degraded the overall balance of metrics. The dynamic version consistently offered a better compromise across validity, uniqueness, novelty, and distributional distance. Even more revealing was a noise-stratified analysis on QM9, which examined which branch dominated at different points along the reverse diffusion trajectory. The results showed a systematic transition: at high noise levels the spectral, global branch carried more weight, while near the final denoising stages emphasis shifted decisively toward the spatial, local branch. This mirrors exactly the intuition that coarse architecture must be settled before fine chemical details can be fixed.

Why does this matter beyond the benchmark tables? De novo molecular design sits at the front end of drug discovery, where the search space of possible small molecules is astronomically large, estimated in some analyses to exceed ten to the sixtieth candidate structures. Generative models act as intelligent compasses in that vast space, proposing molecules that are chemically valid, structurally novel, and statistically similar to known active compounds. Diffusion-based generators have rapidly become favorites in this arena, alongside variational autoencoders, generative adversarial networks, and flow-based models, because they combine stable training with high sample quality. Improvements in how diffusion models represent and denoise molecular graphs therefore translate directly into better raw material for virtual screening and lead optimization pipelines.

The DFDM results also carry a broader lesson for graph generative modeling in general. The finding that a model should weight global spectral information when the data is noisy and local spatial information when the data is nearly clean suggests that the denoising problem itself has a multiscale structure, and that generative architectures should respect it. Rather than chasing ever larger networks, the authors’ approach shows that intelligent coordination of complementary inductive biases, each suited to a different scale and a different point in the generation process, can yield concrete, measurable gains. The work was supported by the National Natural Science Foundation of China, and the authors note that no new datasets were generated or analyzed beyond the benchmarks used. As diffusion models continue their migration into chemistry, dynamic fusion of local and global views may well become a standard ingredient in the generative toolkit, helping machines not merely dream up molecules, but dream them up correctly.

Subject of Research: A dynamic fusion diffusion model for generating two-dimensional molecular graphs by combining spatial and spectral graph neural network branches

Article Title: A dynamic fusion diffusion model for molecular generation

Article References: Sun, L., Lei, X., & Ma, M. (2026). A dynamic fusion diffusion model for molecular generation. Molecular Diversity. https://doi.org/10.1007/s11030-026-11718-9

Image Credits: AI Generated

DOI: 10.1007/s11030-026-11718-9

Keywords: molecular generation, diffusion model, graph neural network, drug discovery, denoising network, QM9, ZINC250k, de novo molecular design, Fréchet ChemNet Distance, spectral graph convolution, machine learning, generative AI

The work, led by Ashwin Srinivasan of BITS Pilani in Goa, together with Tirtharaj Dash, A. Baskar, Michael Bain, Sanjay Kumar Dey, and Mainak Banerjee, addresses a class of hybrid neurosymbolic models that remains relatively under-explored. Rather than letting a neural generator run free and filtering its output afterward, SNG inverts the relationship. A symbolic learner first examines logical specifications of what counts as feasible data, sometimes from as few as a single example. That specification then constrains the conditional information supplied to a neural generator, which is required to reject any candidate instance that violates the symbolic description. The result is a pair consisting of a symbolic description of feasible instances and a set of newly generated instances that provably satisfy it.

Technically, the framework rests on a formal semantics built from partially ordered sets. The researchers construct what they describe as base and fibre posets, combined into an overall partial order that characterizes the space of hybrid systems. The symbolic component in their implementation draws on a restricted form of Inductive Logic Programming, a family of techniques dating back decades that learns logical rules from examples and background knowledge. The neural component is a large language model prompted to produce candidate molecules as SMILES strings, the standard text encoding of molecular structure. Each candidate is checked against the learned logical specification before it is accepted, ensuring that every generated molecule conforms to the constraints the symbolic learner has inferred.

To score candidate hypotheses, the system employs a Bayesian heuristic originally developed in the ILP literature, which balances the prior probability of a hypothesis against how well it explains positive and negative examples. The search over hypotheses proceeds through nested hyper-rectangles defined over molecular factors such as predicted binding affinity, molecular weight, and synthetic accessibility score. Sampling within these rectangles uses a Latin Hypercube approach, and each successive experiment narrows the constraints, producing hypotheses that are logically entailed by their predecessors. The authors prove formal properties about this procedure, including guarantees that the generated sets remain within the feasible extension of the learned hypothesis.

The application that gives the work its urgency is early-stage drug design, specifically the discovery of lead compounds, the validated starting points from which medicinal chemists develop actual drug candidates. The team evaluated their SNG implementation on benchmark protein targets whose biology is well understood, including JAK2 and DRD2, alongside dopamine beta-hydroxylase, or DBH, a comparatively poorly studied enzyme implicated in cardiovascular disease. Binding affinities of generated molecules were estimated computationally using the GNINA molecular docking program, which predicts how strongly a small molecule will lodge itself in the binding pocket of a protein target.

The results are striking. On the benchmark problems, where drug targets are well characterized, SNG performance was statistically comparable to state-of-the-art methods for molecular generation. More intriguing still were the exploratory experiments on DBH, a target for which little domain knowledge exists in the literature. Molecules generated by the system for this target exhibited predicted binding affinities on par with leading clinical candidates. Because the symbolic learner had to infer target-specific constraints from just five known inhibitors, the achievement suggests the framework can operate productively even in the data-sparse regimes that characterize most real drug discovery projects, where high-throughput screening data is expensive and incomplete.

The choice of neural generator turned out to matter in unexpected ways. The researchers compared OpenAI’s GPT-4o, Anthropic’s Claude 3.5 Sonnet, and a small 87-million-parameter GPT-2 model trained on roughly 480 million molecules from the ZINC chemical database. The general-purpose models performed well on heavily studied targets like JAK2 and DRD2, presumably because their vast training corpora include relevant biochemical literature. But the chemistry-specialized Molecule-GPT2 outperformed them on DBH, the lesser-studied target, where knowledge had to be extracted from the inhibitor molecules themselves rather than recalled from pretraining. Notably, the entire study cost only around ten US dollars in API fees per commercial model, underscoring that the framework is not wedded to any particular language model and could accommodate open-source alternatives.

Perhaps the most persuasive endorsement came from human experts. A structural biologist and a synthetic organic chemist reviewed the system’s outputs and found the symbolic specifications genuinely useful as preliminary filters for triaging candidates. Several of the generated molecules were judged viable for synthesis and wet-laboratory testing, a meaningful threshold for any computational design method. The experts’ assessments also highlighted a practical advantage of the neurosymbolic approach: because the system produces an explicit logical description of what makes a molecule feasible, chemists can inspect, critique, and refine the criteria rather than treating the generator as an inscrutable black box.

The authors are careful to situate their contribution within the broader neurosymbolic movement, sometimes called the third wave of neurosymbolic AI, which seeks to combine learning and reasoning rather than choosing between them. They acknowledge that the categorization of systems into purely symbolic or purely neural roles is necessarily a simplification, since induction typically requires deduction and complex reasoning involves trial-and-error learning. They also conjecture that the symbolic learner at the heart of SNG may yield more reliable hypotheses and reduce the number of experiments a laboratory robot would need to conduct in automated discovery loops. Extensions to wider classes of hybrid systems, including those where the neural component plays the primary role, are outlined in the paper’s appendices.

The implications extend well beyond one enzyme or one pharmaceutical project. By demonstrating that logical specifications learned from tiny datasets can steer a generative language model toward chemically meaningful and synthetically plausible outputs, the study offers a template for constrained generation in any scientific domain where correctness matters, from materials design to protein engineering. The researchers have released their code and data openly through a public repository, and the framework’s modest computational demands mean that laboratories without access to frontier-scale models could adopt the approach immediately. As generative AI moves from producing text to proposing tangible scientific artifacts, frameworks like SNG suggest that the path forward lies not in bigger models alone, but in models that know the rules they must obey.

Subject of Research: Hybrid neurosymbolic AI systems that integrate symbolic learning with neural generation for discovering lead molecules in early-stage drug design

Article Title: Symbolic Neural Generation with Applications to Lead Discovery in Drug Design

Article References: Srinivasan, A., Dash, T., Baskar, A., Bain, M., Dey, S. K., & Banerjee, M. (2026). Symbolic Neural Generation with Applications to Lead Discovery in Drug Design. Machine Learning, 115(9), Article 210. https://doi.org/10.1007/s10994-026-07136-5

Image Credits: AI Generated

DOI: 10.1007/s10994-026-07136-5

Keywords: neurosymbolic AI, drug discovery, large language models, inductive logic programming, molecular generation, lead discovery, machine learning, binding affinity, symbolic reasoning, drug design, GPT-4o, dopamine beta-hydroxylase