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 target of the study is the gamma-secretase enzyme complex, a molecular machine embedded in cell membranes that performs the final cutting step in the production of amyloid-beta. The complex is built from four proteins: presenilin, nicastrin, APH-1, and PEN-2. When gamma-secretase cleaves the amyloid precursor protein, it can generate the longer and stickier Aβ42 peptide, which aggregates into the extracellular plaques that are a hallmark of Alzheimer’s pathology. Because mutations in the presenilin 1 gene are linked to familial forms of the disease and drive elevated Aβ42 output, the PSEN1 subunit has long been considered the prime site for therapeutic intervention.

Blocking gamma-secretase outright, however, has proven dangerous. The enzyme also processes Notch receptors, which govern cell differentiation and development, and complete inhibition has been associated with gastrointestinal toxicity and other serious side effects, a problem that contributed to the clinical failure of drugs such as semagacestat. The field has therefore shifted toward gamma-secretase modulators, compounds that selectively lower Aβ42 while leaving the processing of other substrates untouched. Even here, progress has been rocky: the modulator E2012 showed strong amyloid reduction but raised concerns about effects on cholesterol metabolism, and BMS-932481 was hampered by liver toxicity in early trials.

To guide their search for better modulators, the researchers focused on two reference compounds, BMS 299897 and ELN318463. Both bind at an allosteric pocket at the interface of transmembrane helices six and seven of the PSEN1 subunit, a region distinct from the catalytic aspartates but positioned to influence how the active site handles its substrate. ELN318463 is particularly notable because, in cell-based assays, it shows a seventy-five to one-hundred-twenty-fold preference for blocking amyloid-beta production over Notch signaling. Using the LigandScout software, the team generated individual pharmacophore maps for each drug, mapping out hydrogen bond donors and acceptors, hydrophobic regions, aromatic rings, and halogen bond donors, and then aligned the two maps to build a shared-feature pharmacophore model that captured the essential interaction points common to both inhibitors.

With this model in hand, the team screened the Comprehensive Marine Natural Products Database, a library of roughly 47,451 compounds sourced from algae, sponges, corals, and other marine organisms. After removing duplicates, 43,212 molecules were virtually screened against the shared pharmacophore over approximately forty-eight hours on a sixty-four-core processor. Six compounds emerged as top hits, with the best, CMNPD10454, achieving a pharmacophore fit score of about 110.4, indicating a near-perfect match with the key interaction features. Marine natural products are prized in drug discovery for their unusual chemical architectures, and many display antioxidant, anti-inflammatory, and neuroprotective activities, making them attractive starting points for new Alzheimer’s therapies.

The raw hits, however, were structurally complex and considered impractical as direct drug candidates. The team therefore turned to fragment-based de novo design. A key observation drove this step: both BMS 299897 and ELN318463 share a 4-chlorobenzenesulfonamide ring that plays a central role in hydrophobic interactions with the enzyme. Using the AlvaBuilder toolkit, which applies genetic algorithms to molecular design, the researchers generated fifty new synthetic modulators by keeping this ring fixed and grafting bioactive fragments from the marine hits onto it. The design constraints included Lipinski’s Rule of Five parameters, a synthetic accessibility score of five or below, limits on halogen count, and estimated aqueous solubility thresholds, all intended to ensure the resulting molecules were both effective in theory and chemically feasible to make.

The fifty designs were then filtered through absorption, distribution, metabolism, and excretion profiling with SwissADME, and only three molecules, numbered 6, 24, and 28, were predicted to cross the blood-brain barrier, a filter that has doomed many otherwise promising central nervous system drug candidates. Molecular docking against the human gamma-secretase crystal structure, using the Protein Data Bank entry 5A63, showed that all three bound more tightly than the control drug BMS 299897, which scored minus 8.9 kilocalories per mole. Molecule 6 achieved the best docking energy at minus 10.6 kilocalories per mole, driven by extensive hydrophobic contacts with residues including PHE411, VAL94, and ILE408, while Molecules 24 and 28 scored minus 9.7 and minus 9.6 respectively. Notably, none of the top compounds formed hydrogen bonds with the active site, suggesting that hydrophobic interactions dominate their binding mode.

Molecular dynamics simulations over one hundred nanoseconds, run with the Desmond software using the OPLS_2005 force field and physiological salt conditions, provided further evidence of stability. Molecule 24 produced the most stable protein-ligand complex, with the lowest root mean square deviation of 4.03 angstroms and the lowest residue fluctuation values, maintaining a compact radius of gyration and consistent solvent exposure throughout the simulation. It also formed a hydrogen bond with CYS4 and ionic interactions with LEU243 and LEU244, indicating robust and persistent binding. Toxicity profiling with the OECD QSAR Toolbox predicted no mutagenicity alerts and stable tautomeric forms for the lead compounds, along with lower bioaccumulation factors than the control drug, although renal toxicity alerts resembling a sulfasalazine-like profile warrant future experimental scrutiny.

An intriguing and unexpected finding emerged from the interaction analysis. While Molecule 6 retained strong contact with PSEN1, consistent with the pharmacophore model that guided its design, Molecules 24 and 28 preferentially anchored to the APH-1 subunit of the gamma-secretase complex instead. The precise role of APH-1 in modulating amyloid-beta generation remains uncertain, but prior structural studies suggest it participates in complex assembly and can influence the conformation of the catalytic subunit. The authors argue that these contacts may represent an alternative mechanism of modulation rather than a flaw in the design strategy, and they emphasize that future wet-laboratory experiments will be needed to determine whether APH-1 interactions contribute functionally to enzyme regulation or are merely incidental.

The study also addressed the practical question of how the lead compounds could actually be synthesized. Using the IBM RXN retrosynthesis platform, which combines template-based and template-free neural network approaches with Monte Carlo tree search, the team mapped a stepwise route for Molecule 24 starting from commercially available 4-chlorobenzene sulfonyl chloride, proceeding through click-chemistry and nucleophilic substitution steps to assemble the final structure. While all of these results remain computational predictions that require experimental validation in laboratory and animal models, the work demonstrates how marine chemical diversity, pharmacophore-guided screening, fragment-based design, and molecular simulation can be woven together to accelerate the hunt for safer Alzheimer’s therapies, offering a template for discovering next-generation gamma-secretase modulators that lower amyloid-beta without disrupting the essential cellular pathways that previous drug candidates damaged.

Subject of Research: Computational design of marine-derived gamma-secretase modulators to reduce amyloid-beta production in Alzheimer's disease

Article Title: In silico pharmacophore-guided modeling of marine-derived γ-secretase modulators for amyloid-beta reduction in Alzheimer's disease

Article References: In silico pharmacophore-guided modeling of marine-derived γ-secretase modulators for amyloid-beta reduction in Alzheimer's disease. (n.d.). https://doi.org/10.1016/j.heliyon.2026.e45453

Image Credits: AI Generated

DOI: 10.1016/j.heliyon.2026.e45453

Keywords: Alzheimer's disease, gamma-secretase modulators, marine natural products, pharmacophore modeling, molecular docking, molecular dynamics simulation, amyloid-beta, PSEN1, blood-brain barrier, drug discovery, virtual screening, de novo molecular design