What are diffusion models in drug discovery?

Diffusion models in drug discovery are a type of generative AI that creates new molecules by reversing a process of gradual corruption, the same principle behind AI image generation. During training, the model learns how structure is destroyed by adding noise step by step, then learns to reverse it. Applied to fresh noise, the learned denoising process produces a brand-new molecule. For drug design, the most powerful versions work in three dimensions, generating a molecule's actual atomic structure and often generating it directly inside a target protein's binding pocket. This lets the model design for physical fit from the start. Their main advantage over earlier generative methods is more stable training and higher-quality, more diverse 3D output.

How does DiffDock work?

DiffDock is a diffusion model that addresses molecular docking, predicting how and where a given molecule binds to a target protein, rather than generating new molecules. It reframes docking as a generative task: instead of searching and scoring candidate poses the way traditional docking software does, it uses a diffusion process to generate likely binding poses directly, along with confidence estimates for each. On a commonly used docking benchmark, it demonstrated substantially improved performance over existing state-of-the-art methods while running relatively quickly. This matters because accurate pose prediction underpins structure-based design, and DiffDock showed that the generative diffusion approach could outperform established docking algorithms on the problem.

How do diffusion models compare to VAEs for molecules?

Variational autoencoders (VAEs) and generative adversarial networks (GANs) were earlier approaches to generative molecular design, and diffusion models have largely overtaken them for 3D structure-based work for practical reasons. Diffusion models train more stably than GANs, which are notoriously difficult to train and prone to mode collapse, and they tend to produce higher-quality, more diverse output than VAEs, which can generate blurry or less varied results. For generating 3D molecular structures conditioned on a protein pocket, diffusion models have set the standard. VAEs still have uses, particularly where a smooth, continuous latent space is valuable for interpolation and controlled exploration, so the choice depends on the task rather than one approach being universally better.

Articles

Diffusion models for drug design: Generating novel molecules in 3D

The diffusion model revolution behind AI image generation has a direct parallel in molecular design. Here is how it works and how far the generated molecules are from being drugs.

Written byTrevor J Henderson

| 7 min read

Computational chemist viewing a diffusion model denoising a cloud of atoms into a 3D drug molecule inside a protein binding pocket on screen

A diffusion model generates a molecule by starting from noise and progressively denoising it into a structured 3D shape, here directly inside a target binding pocket.

Flow (2026)

Highlights in this article

Key takeaways
What is a diffusion model?
How diffusion is applied to molecular generation
3D vs. 2D generation
Key models: DiffDock, DiffSBDD, TargetDiff
Benchmarking
From generated structure to chemistry

Few areas of computational chemistry are moving faster than diffusion models for drug design, which borrow the same class of artificial intelligence (AI) method that powers modern image generation and point it at molecules. The core idea is elegant: start from random noise and progressively refine it into a structured molecule, learning the reverse of a process that gradually destroys structure. What makes this genuinely useful for drug design is that these models can work directly in three dimensions, generating a molecule's actual 3D shape, and even generating it inside a target protein's binding pocket, rather than producing a flat chemical formula and inferring the shape afterward.

This article explains the technical principles for a scientifically literate reader and assesses what the models deliver in practice. It sits within DDN's coverage of generative AI in molecular design, complements the discussion of AlphaFold and protein structure prediction, and connects to the broader guide to AI across drug discovery.


Key takeaways Diffusion models generate molecules by starting from noise and progressively denoising it into a structured form, the same principle behind AI image generation, adapted to chemistry. Their main advantage over earlier generative approaches such as VAEs and GANs is more stable training and higher-quality, more diverse 3D output, particularly for structure-based design. Working in 3D lets diffusion models generate a molecule directly inside a target's binding pocket, designing for physical fit from the start rather than checking it afterward. Leading models illustrate the range: DiffDock reframes docking as generation, while DiffSBDD and TargetDiff generate novel molecules conditioned on a protein pocket. The persistent limitation is chemical realism: diffusion models often generate structures that neglect synthetic feasibility and drug-like properties, so newer methods build those constraints into generation.

Key takeaways

Diffusion models generate molecules by starting from noise and progressively denoising it into a structured form, the same principle behind AI image generation, adapted to chemistry.
Their main advantage over earlier generative approaches such as VAEs and GANs is more stable training and higher-quality, more diverse 3D output, particularly for structure-based design.
Working in 3D lets diffusion models generate a molecule directly inside a target's binding pocket, designing for physical fit from the start rather than checking it afterward.
Leading models illustrate the range: DiffDock reframes docking as generation, while DiffSBDD and TargetDiff generate novel molecules conditioned on a protein pocket.
The persistent limitation is chemical realism: diffusion models often generate structures that neglect synthetic feasibility and drug-like properties, so newer methods build those constraints into generation.

What is a diffusion model?

A diffusion model is a type of generative AI that learns to create data by reversing a process of gradual corruption. The idea is easiest to grasp through its now-familiar use in images: during training, the model takes a clear image and adds random noise to it step by step until nothing but noise remains, then learns to reverse that process, predicting how to remove noise at each step. Once trained, it can start from pure noise and denoise its way to a brand-new, coherent image. The same mathematics applies to molecules, with 3D atomic structure taking the place of pixels.

The core mechanics, in plain terms:

The forward process. Structure is progressively destroyed by adding noise in small steps, turning a well-formed molecule into a random cloud of atoms. This step is used only in training and is straightforward to define.
The reverse process. The model learns to undo the corruption one step at a time, and this learned denoising is the generative engine: applied to fresh noise, it produces a new structured molecule.
Equivariance. For molecules, the model must respect the physics of 3D space, so that rotating or translating a molecule does not change its chemistry. Diffusion models for molecules use equivariant neural networks that build this symmetry in, which is central to why they work well in 3D.
Conditioning. Generation can be guided by a condition, most importantly the structure of a target protein's binding pocket, so the model generates molecules tailored to a specific target rather than sampling chemical space at random.

The reason diffusion models displaced earlier approaches so quickly is largely practical: they train more stably and produce higher-quality, more diverse output than the generative methods that preceded them, a difference that turns out to matter a great deal for molecular generation.

How diffusion is applied to molecular generation

Translating the diffusion recipe from images to molecules requires deciding what exactly gets noised and denoised, and the answer shapes what a model can do. The most powerful approaches for drug design operate on the full 3D structure: the atoms, their types and their positions in space, and often the bonds between them.

What a molecular diffusion model generates:

Atom types and coordinates. The model jointly denoises which atoms are present and where they sit in 3D space, building a molecule as a set of positioned atoms rather than a text string, which keeps geometry central throughout generation.
Bonds and connectivity. Because atom positions alone do not fully specify a molecule, better models also generate the bonds explicitly, modeling the interdependence of atoms and bonds so the output is a chemically coherent structure rather than a plausible-looking point cloud.
Pocket-conditioned generation. For structure-based design, the target protein's binding pocket is supplied as a fixed condition, and the model generates a ligand within it, so the molecule is shaped to the site from the first denoising step.
Fragment and partial generation. By fixing part of a structure and denoising the rest, diffusion models support practical medicinal chemistry tasks such as fragment growing, scaffold hopping and linker design, not just whole-molecule generation from scratch.

That last capability matters more than it might seem. The ability to fix a known-good part of a molecule and regenerate the rest maps directly onto how medicinal chemists actually work, refining a lead rather than starting over, which is part of why 3D diffusion has drawn serious interest from drug design programs rather than remaining a purely academic technique.

3D vs. 2D generation

A central distinction in generative molecular design is whether a model works in two dimensions, generating a molecule as a graph or text string, or in three, generating its actual spatial structure. Diffusion models are not the only way to do 3D generation, but they are the approach that made high-quality 3D generation practical, and the difference has real consequences for drug design.

Dimension	What the model generates	Consequence for drug design
2D (graph/string)	Atoms and bonds as a topology, no coordinates	Fast and valid, but shape and fit must be inferred separately
3D (coordinates)	Atoms positioned in space, often with bonds	Designs for physical fit directly, at higher computational cost
3D, pocket-conditioned	A ligand generated inside the target's binding site	Accounts for the target's shape from the start of generation

The case for 3D is that binding is fundamentally a three-dimensional phenomenon: whether a molecule fits and binds its target depends on shape and spatial complementarity that a 2D representation does not capture. Generating directly in 3D, especially inside the pocket, lets the model reason about fit while it designs. The cost is computational expense and, as the benchmarking discussion below shows, a tendency to produce geometrically plausible molecules that are not always chemically sensible.

Binding is a three-dimensional phenomenon. A molecule that looks right on paper may not fit its target in space, which is the case for generating molecules in 3D, and also the source of diffusion models' hardest problem: 3D-plausible is not the same as chemically real.

Key models: DiffDock, DiffSBDD, TargetDiff

Three models are worth understanding because they represent the main ways diffusion is applied in structure-based drug design, and because they anchor much of the benchmarking literature. Each takes a distinct angle on the problem.

The landmark diffusion models for drug design:

DiffDock. Rather than generating a molecule, DiffDock reframes molecular docking as a generative task, using diffusion to produce ligand binding poses along with confidence estimates. It demonstrated substantially improved performance over existing state-of-the-art docking methods on a common benchmark while running relatively quickly, a notable result for the docking problem specifically.
DiffSBDD. A structure-based design model, DiffSBDD uses an equivariant diffusion approach to design molecules directly in the 3D environment of a protein's binding site, denoising sets of points into ligands conditioned on the pocket. It also supports inpainting-style partial generation for fragment and scaffold tasks.
TargetDiff. A 3D equivariant model, TargetDiff jointly generates atom coordinates and types conditioned on protein pocket atoms, and additionally supports binding-affinity prediction from its learned representations, performing strongly on a standard structure-based benchmark in both molecular geometry and binding metrics.

The distinction among them is instructive. DiffDock addresses where a known molecule binds; DiffSBDD and TargetDiff address what molecule to make for a given pocket. Both questions matter in a drug program, and diffusion has made credible progress on each, which is why these models are reference points rather than curiosities.

Benchmarking

Benchmarking is where enthusiasm meets evidence, and for diffusion models in drug design the picture is genuinely mixed. The models perform well on the metrics they were designed around, geometry and predicted binding, while independent evaluation has repeatedly exposed weaknesses in the properties that actually determine whether a generated molecule is a viable drug candidate.

What benchmarking reveals:

Strong on geometry and predicted binding. On standard structure-based benchmarks, leading diffusion models generate molecules with good 3D geometry and favorable predicted binding affinity, which is the result most often highlighted.
Weak on chemical realism. A recurring finding is that structure-based diffusion models often produce unrealistic 3D molecules that neglect structural feasibility and drug-like properties, generating shapes that score well on binding but would not make sensible drugs.
Benchmark dependence. Performance rankings shift depending on the benchmark and the metrics chosen, so comparisons across models must be read carefully, and headline claims of superiority often rest on a specific evaluation setup.
Property guidance helps. Newer models explicitly incorporate binding affinity and drug-like property guidance into training and sampling, and these generate molecules with more rational chemical structure, a direct response to the realism problem.

The honest summary from the evaluation literature is that many structure-based diffusion models produce unrealistic molecules unless structural feasibility and drug-likeness are explicitly built in, and that addressing this, through bond modeling and property guidance, is where the field is actively improving. Impressive binding scores on a benchmark are necessary but not sufficient evidence that a model is useful for real drug design.

From generated structure to chemistry

The final and most important question is what happens when a diffusion-generated molecule leaves the screen and meets a medicinal chemist. Here, the same reality that applies across generative design applies with particular force: a 2025 review of diffusion models in drug discovery identifies the core challenge for small molecules as bridging the gap from computational validity to practical chemical synthesizability. A molecule that a diffusion model places beautifully inside a pocket is only useful if a chemist can actually make it.

What has to happen after generation:

Synthesizability assessment. Generated molecules must be checked for whether they can realistically be synthesized, and a meaningful fraction of raw diffusion output fails this test, containing strained, reactive, or simply unmakeable substructures.
Drug-likeness filtering. Beyond synthesis, generated molecules are screened against drug-like property criteria, since a high predicted binding score means little if the molecule could never be developed into a viable therapeutic.
Experimental validation. A predicted binding pose is a hypothesis, not a fact, and confirming that a generated molecule actually binds as designed still requires experimental work, connecting back to the limits of prediction discussed for structure-based methods generally.
Medicinal chemistry judgment. Deciding which generated candidates are worth pursuing remains an expert task, with diffusion output serving as a source of well-shaped starting points rather than finished designs.

The balanced conclusion is that diffusion models are a genuine advance in generating novel 3D molecules, particularly for structure-based design, and that their trajectory, toward built-in chemical realism, is encouraging. But the gap between a geometrically impressive generated structure and a synthesizable, developable drug remains the decisive one, and closing it is where the practical value of these models will ultimately be judged. For how this fits the wider design landscape, DDN's generative AI in molecular design hub provides the surrounding context.


What this means for drug design teams Diffusion models are worth taking seriously for structure-based design, where generating molecules directly inside a target pocket in 3D is a real capability, not a demo. Use DiffDock-style models for pose prediction and DiffSBDD- or TargetDiff-style models for pocket-conditioned generation, but read benchmark claims critically: strong geometry and predicted binding do not guarantee chemically realistic molecules. Prioritize models and workflows that build synthesizability and drug-like property guidance into generation rather than filtering afterward, since raw diffusion output often includes molecules that cannot be made or developed. Treat generated structures as well-shaped starting points for medicinal chemistry, confirm binding experimentally, and keep expert judgment in the loop on what is worth pursuing. For the broader design picture, the generative AI in molecular design guide and the AI in drug discovery overview map how 3D generation connects to the rest of the pipeline.

What this means for drug design teams

Diffusion models are worth taking seriously for structure-based design, where generating molecules directly inside a target pocket in 3D is a real capability, not a demo. Use DiffDock-style models for pose prediction and DiffSBDD- or TargetDiff-style models for pocket-conditioned generation, but read benchmark claims critically: strong geometry and predicted binding do not guarantee chemically realistic molecules. Prioritize models and workflows that build synthesizability and drug-like property guidance into generation rather than filtering afterward, since raw diffusion output often includes molecules that cannot be made or developed. Treat generated structures as well-shaped starting points for medicinal chemistry, confirm binding experimentally, and keep expert judgment in the loop on what is worth pursuing. For the broader design picture, the generative AI in molecular design guide and the AI in drug discovery overview map how 3D generation connects to the rest of the pipeline.

This article was produced in accordance with Drug Discovery News’ Editorial Policies.

Frequently Asked Questions (FAQs)

What are diffusion models in drug discovery?
Diffusion models in drug discovery are a type of generative AI that creates new molecules by reversing a process of gradual corruption, the same principle behind AI image generation. During training, the model learns how structure is destroyed by adding noise step by step, then learns to reverse it. Applied to fresh noise, the learned denoising process produces a brand-new molecule. For drug design, the most powerful versions work in three dimensions, generating a molecule's actual atomic structure and often generating it directly inside a target protein's binding pocket. This lets the model design for physical fit from the start. Their main advantage over earlier generative methods is more stable training and higher-quality, more diverse 3D output.
How does DiffDock work?
DiffDock is a diffusion model that addresses molecular docking, predicting how and where a given molecule binds to a target protein, rather than generating new molecules. It reframes docking as a generative task: instead of searching and scoring candidate poses the way traditional docking software does, it uses a diffusion process to generate likely binding poses directly, along with confidence estimates for each. On a commonly used docking benchmark, it demonstrated substantially improved performance over existing state-of-the-art methods while running relatively quickly. This matters because accurate pose prediction underpins structure-based design, and DiffDock showed that the generative diffusion approach could outperform established docking algorithms on the problem.
What is 3D molecular generation?
3D molecular generation is the creation of new molecules as full three-dimensional structures, atoms positioned in space and usually the bonds between them, rather than as flat two-dimensional graphs or text strings. It matters for drug design because binding is fundamentally three-dimensional: whether a molecule fits and binds its target depends on shape and spatial complementarity that a 2D representation does not capture. Diffusion models made high-quality 3D generation practical, and the most useful versions generate a molecule directly inside a target's binding pocket, so the design accounts for the target's shape from the first step. The trade-off is higher computational cost and a tendency to generate geometrically plausible molecules that are not always chemically realistic.
How do diffusion models compare to VAEs for molecules?
Variational autoencoders (VAEs) and generative adversarial networks (GANs) were earlier approaches to generative molecular design, and diffusion models have largely overtaken them for 3D structure-based work for practical reasons. Diffusion models train more stably than GANs, which are notoriously difficult to train and prone to mode collapse, and they tend to produce higher-quality, more diverse output than VAEs, which can generate blurry or less varied results. For generating 3D molecular structures conditioned on a protein pocket, diffusion models have set the standard. VAEs still have uses, particularly where a smooth, continuous latent space is valuable for interpolation and controlled exploration, so the choice depends on the task rather than one approach being universally better.

Add Drug Discovery News as a preferred Google source to see more of our trusted coverage.

About the Author

Trevor J Henderson
Trevor Henderson is the Creative Services Director for the Laboratory Products Group at LabX Media Group. With over two decades of experience, he specializes in scientific and technical writing, editing, and content creation. His academic background includes training in human biology, physical anthropology, and community health. Since 2013, he has been developing content to engage and inform scientists and laboratorians.
View Full Profile