Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport
summary
The gist
Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport introduces Morph…
In short
Morph is a flexible-size generative model for geometric graphs that can simultaneously insert, delete, and substitute nodes. It uses Unbalanced Optimal Transport to align graphs of different sizes during training. This allows it to generate molecules with sizes outside the training data while respecting structural priors like scaffolds, making it highly steerable for de novo molecular design.
Key concepts
- Flexible-Size Generative Model
- This model is designed to handle geometric graphs (like molecules) that can change their size dynamically. Unlike fixed models, Morph can perform multiple insertions and deletions during generation, allowing it to create molecules of varying sizes seamlessly.
- Unbalanced Optimal Transport (UOT)
- UOT is a mathematical technique used during training to align geometric graphs that have different numbers of nodes. It defines target probability paths by minimizing costs related to movement, deletion, and insertion penalties, ensuring the model learns how to transition between different graph sizes effectively.
- Conditional Generator Matching (CGM)
- CGM is the training framework used to teach the model. It works by recovering a marginal generator from conditional generators along paths between paired molecules. This process requires structural alignment provided by UOT to define the correct training targets.
Terminology used across episodes
This episode discusses
- Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport · Paper Radio
- Navigating the Design Space of Equivariant Diffusion-Based Generative Models for De Novo 3D Molecule Generation
- SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching
- Edit Flows: Flow Matching with Edit Operations
- Generator Matching: Generative modeling with arbitrary Markov processes
- Discrete Flow Matching
- Generative Flows on Discrete State-Spaces: Enabling Multimodal Flows with Applications to Protein Co-Design
- Flow Matching for Generative Modeling
- OneFlow: Concurrent Mixed-Modal and Interleaved Generation with Edit Flows
- Levenshtein Transformer
- Insertion-Deletion Transformer
- Beyond In-Place Corruption: Insertion and Deletion In Denoising Probabilistic Models
- Any-Order Flexible Length Masked Diffusion
- Insertion Language Models: Sequence Generation with Arbitrary-Position Insertions
- DiGress: Discrete Denoising diffusion for graph generation
- Graph Diffusion that can Insert and Delete
- All-atom Diffusion Transformers: Unified generative modelling of molecules and materials
- Zatom-1: Towards a Multimodal Foundation Model for 3D Molecules and Materials
- TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality
- Prior-Guided Flow Matching for Target-Aware Molecule Design with Learnable Atom Number
- Trans-Dimensional Generative Modeling via Jump Diffusion Models
The paper
Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport · Read on arXiv
Malte Franke, Stefan P. Schmid, Žarko Ivkovic, Kjell Jorner, Andreas Krause
ETH Zurich · NCCR Catalysis
The success of generative molecular design depends on a model's steerability toward high-reward samples. Because many molecular properties are intrinsically linked to molecular size, accurately capturing the joint distribution of properties and the number of atoms is essential. However, current diffusion and flow-based models fix the number of atoms, which ultimately limits their ability to navigate this complex relationship. To address this, we introduce Morph, a flexible-size generative model for conditional and unconditional 3D molecular design based on geometric graphs. By dynamically adapting size, Morph can seamlessly integrate existing structural priors, like scaffolds, and significantly enhances property steering. We show that Morph matches current fixed-size state-of-the-art models while significantly increasing sampling flexibility. We demonstrate out-of-distribution generation in regimes where previous models fail, paving the way for enhanced generative modeling for molecular design.
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport".
Jane: Here's the gist of the paper. Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport introduces Morph, a flexible-size generative model for geometric graphs that can simultaneously perform multiple insertions, deletions and substitutions to address the limitation of fixed-size models in molecular design The success of generative molecular design hinges on a model’s steerability toward high-reward samples Morph matches current fixed-size state-of-the-art models while offering the benefit of unparalleled sampling flexibility.
Tom: Model Architecture and Dynamics Morph is a discrete-continuous flow-based model for geometric graphs that can dynamically change the number of nodes in a graph by performing multiple insertion and deletion operations The state space G is constructed as the disjoint union of these quotient spaces where g ∈ G = n]max n=1 A n × R n×3 × Sn(B) (SE(3) × Sn) (1) Morph's generative process involves a continuous evolution governed by a velocity field vt and discrete jumps via a Continuous-Time Markov Chain (CTMC) The combined generator of the coupled process is the sum of their flow and jump generators
Ltf: (g) = vt(g) · ∇xf(g)flow + Z G f(g′) − f(g)ut(g′ g).
Jane: Training Signal Construction The training relies on the Conditional Generator Matching (CGM) framework, which exploits the fact that the marginal generator Lt can be recovered as a conditional expectation of conditional generators defined along paths between paired endpoints This requires establishing structural alignment between graphs of different sizes using Unbalanced Optimal Transport (UOT) to define target probability paths for training The UOT objective seeks a transport plan P that minimizes the total cost involving movement, deletion, and insertion penalties Generative Process Implementation The generation process involves defining three mutually exclusive edit sets based on the optimal assignment matrix Pˆ∗ obtained from the UOT matching The interpolation path is defined by scheduling continuous dynamics for surviving nodes, independent event times for insertion and deletion nodes, and discrete rates for atom-type and edge-type substitutions Algorithm 3 details the jump-flow integration where node deletions, substitutions, and insertions are modeled using bounded hazard distributions to ensure termination at t = 1.
Tom: Performance on Molecular Design Morph demonstrates competitive performance on de-novo design tasks on QM9 and GEOM-Drugs while offering more flexible generation trajectories than other models It shows superior steerability, successfully generating valid molecules with sizes outside the training distribution when conditioned on properties like logP or QED score.
Title and authors: Tom: So Jane, this paper is about this model called Morph which aims to fix a problem with current molecular design AI where you can only generate molecules of one size.
Jane: Right, so the core idea here is that you want a model that can handle different sizes during the generation process without getting stuck on just one size.
Tom: Exactly, and they introduce Morph as this flexible-size generative model for geometric graphs, which means it can do insertions, deletions, and substitutions all at once.
Lu: From my view, what's really interesting is how they build the state space G as this union of spaces where you have the number of atoms changing dynamically.
Meng: But that sounds computationally heavy; how does this dynamic resizing actually work in practice when you're trying to ship something to production?
Jane: Well, they describe it as a discrete-continuous flow-based model, meaning it has continuous movement for the atoms and discrete jumps for when things get inserted or deleted.
Tom: And they show how this is governed by a combined generator that’s the sum of two parts: a flow part and a jump part, which is pretty technical.
Lu: The jump part uses something called a Continuous-Time Markov Chain, which describes those discrete changes in the graph structure itself based on certain rates.
Meng: I'm worried about fitting these drift and jump rates vt and ut because they rely on learning the marginal distribution pt of the graph, which sounds like an intractable problem.
Jane: That's where they get creative by using Conditional Generator Matching, or CGM, which lets them recover that marginal distribution indirectly through conditional expectations along paths between paired endpoints.
Tom: And to define those target probability paths for training, they use something called Unbalanced Optimal Transport, or UOT, to align graphs of different sizes.
Lu: The UOT objective is pretty clever; it seeks a transport plan P that minimizes the total cost involving movement penalties along with deletion and insertion penalties across the size difference.
Meng: So instead of training on just one fixed-size graph, they are training on how to smoothly transition between graphs of varying sizes using this transport plan.
Jane: That addresses the limitation where molecules like scaffolds might need to be modified or decorated in ways that change their total atom count, which is a huge part of molecular design.
Title and authors: Tom: And what they show is that Morph matches current fixed-size state-of-the-art models on de novo design tasks like QM9 and GEOM Drugs while offering this extra flexibility.
Lu: Their main contribution is extending the model to learn the joint distribution of size and target properties, which is what was missing in prior diffusion and flow models that fix the number of atoms.
Meng: I wonder how much more complex training becomes when you have to account for all those combinatorics across different numbers of atoms, because that sounds like it will increase training expense significantly.
Jane: And on the positive side, they demonstrate out-of-distribution generation when you condition Morph on things like logP or QED score, which is a big deal because it means it can explore chemical space beyond what was seen in the training data.
Tom: It also excels at scaffold decoration, which is useful for lead optimization tasks where you want to keep the core structure but tweak substituents.
Lu: They show that by using this approach, Morph can seamlessly integrate existing structural priors, like scaffolds, by defining interpolation paths that preserve common motifs while allowing flexible edits at substituent sites.
Meng: That sounds promising for scaffold decoration because it avoids having to re-design the whole molecule when you're just making minor changes.
Jane: So when we look at the overall process described in Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport, they detail how you define three mutually exclusive edit sets based on that optimal assignment matrix P hat star from the UOT matching.
Tom: The generation process then involves scheduling continuous dynamics for surviving nodes, independent event times for insertion and deletion nodes, and discrete rates for atom-type and edge-type substitutions.
Lu: They use bounded hazard distributions to model those jump-flow integrations, which ensures that the process terminates at time one regardless of how many edits occur.
Meng: I'm still thinking about the runtime; since they need to integrate these flow and jump generators, will this process be slow enough for real-time use, or is it just theoretical?
Jane: The paper mentions that the combined generator governs the trans-dimensional process via a Kolmogorov forward equation, which means learning the dynamics requires fitting those drift and jump rates.
Tom: And they show that learning these targets relies on fitting vt and ut, but then they admit those targets rely on the intractable marginal distribution pt of the graph, which makes it inaccessible directly.
Title and authors: Lu: That's a key limitation they flag; even though the model is flexible, learning its exact dynamics is still theoretically difficult because you can't easily access that marginal distribution.
Meng: So, if we want to deploy this at scale, do we have to resort to a lot of integration steps just to get a good generation?
Jane: That's what they suggest; the CTMC nature implies that a high number of integration steps might be necessary for full generation.
Tom: Overall, this paper shows Morph is an excellent distribution learner that also enables flexible generation based on structure priors, which is validated on a scaffold decoration task.
Lu: It’s a strong step forward because it tackles the fundamental constraint of fixed-size models by making size adaptation part of the generative process itself.
Meng: For me, as someone focused on deployment, the main hurdle I see is scaling this up and managing those complex loss terms from the jump process.
Jane: So to wrap up this paper on Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport, it shows a flexible-size model that can simultaneously perform insertions, deletions and substitutions.
Tom: It achieves competitive performance to leading fixed-size generative models on de novo design tasks while being significantly more expressive.
Lu: By being able to generate out of distribution samples when conditioned on properties like logP or QED score, Morph is a promising building block for accelerating discoveries in the chemical sciences and beyond.
Meng: It’s definitely a powerful tool for steering design toward specific property targets, even if the training cost is high.
Jane: So that's what we saw here with Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport—a flexible-size three dee generative model for geometric graphs that can simultaneously perform multiple insertions, deletions and substitutions <ref:2606.07239#pg1,Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport>.
Tom: It’s a significant piece of work because it proves steerability is key, and Morph seems to be a strong candidate for the next generation of molecular design tools.
Lu: We'll be watching how they tackle the sample-efficiency trade-offs between this incredible steerability and needing many integration steps for full generation.
Meng: I'm going to keep an eye on those complexity costs because we need something that runs efficiently on real traffic, not just in simulation.
Jane: Alright everyone, that covers what we had on Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport for now.
The paper's summary: Tom: So, to wrap up what we just heard about Morph, this paper is basically saying that they built this model that can handle changing molecular sizes during generation—insertions, deletions, and substitutions all at once.
Jane: Right, so instead of a model that only makes molecules a certain size and then you have to manually re-design it if you need more atoms or fewer atoms, Morph is designed to adapt its structure while it's drawing the molecule.
Lu: What's really new here is the way they handle the training signal; they use this thing called Conditional Generator Matching, which lets them figure out how to train a model when you want different sizes.
Meng: But I’m still wondering about the training cost; if you have to account for every possible number of atoms and all those different ways things can change, how much more expensive does that make the training process?
Tom: Well, they used this technique called Unbalanced Optimal Transport to line up graphs of different sizes, which defines the target paths for their training data.
Jane: That means they are essentially teaching the model how to smoothly transition between a small molecule and a large one while it's learning.
Lu: And on the results side, they show that this approach allows Morph to generate molecules that don't even exist in their original training set when you condition them on properties like logP or QED score.
Meng: So it can explore chemical space outside what we initially saw, which is useful for finding new drug candidates that might have properties we haven't mapped yet.
Tom: Exactly, and they also showed it’s really good at scaffold decoration, meaning you can keep the main shape but swap out parts easily.
Jane: That flexibility with structural priors is a big deal because it means you don't have to completely rebuild the molecule when you just want to tweak a side chain or add a functional group.
Lu: The mechanism they use for generation involves scheduling continuous dynamics for the atoms that stay and discrete jumps for adding or removing nodes, all tied together by some combined generator.
Meng: That sounds like it's a bit more complex to run than just running one single flow process; what’s the practical cost of running this kind of jump-flow integration?
Tom: They use bounded hazard distributions to model those jumps, which makes sure the generation process finishes in a set amount of time, like at time one.
Jane: So while it's complicated to train and run, they show that Morph is competitive on de novo design tasks while being much more flexible than its fixed-size counterparts.
Lu: The limitation they mention is that learning the exact dynamics—those drift and jump rates—is hard because you can’t easily get access to the marginal distribution of the graph.
Meng: That tells me that if we want to deploy this at scale, we might have to run a lot of integration steps just to get a good result, which adds runtime.
Tom: So, even with its flexibility and OOD generation capabilities, Morph still has those scaling challenges and training hurdles that need addressing before it’s ready for heavy production use.
Jane: It's definitely a strong step forward because it tackles the fundamental constraint of fixed-size models by making size adaptation part of the generative process itself.
Lu: And their future work is focused on figuring out how to scale this up and balancing that incredible steerability against needing many integration steps for full generation.
Tom: So, we've seen that Morph is a powerful tool for steering design toward specific property targets, even if the training cost is high and deployment needs refinement.
Jane: Next time we talk about these models, I want to focus on how they handle those trade-offs between being super flexible and being fast enough for real-world use.
The paper's improvements: Tom: So, we’ve talked about how Morph lets you change molecule sizes during generation—insertions and deletions—and now we need to talk about what they suggest is the next step for this model.
Jane: They point out that a big part of the improvement is using Unbalanced Optimal Transport to align those graphs of different sizes before they even start interpolating.
Lu: Exactly, that UOT alignment is crucial because it creates that structural alignment needed to handle transitions between graphs with wildly different node counts efficiently.
Meng: What else do they suggest? I'm looking for something practical for deployment, something that makes this model more robust when we move off the benchmark data.
Tom: They also highlight the use of learnable atom number priors, which means the model can be guided to generate molecules with specific atom counts based on what it sees.
Jane: That’s really important because it gives us control over the output beyond just property scores; we can direct the generation towards specific structural goals.
Lu: They are also suggesting that they need to focus on scaling this up for real-world applications and figuring out the sample efficiency trade-offs between having super steerability and needing a lot of integration steps for a full molecule.
Meng: That sounds like the core problem we have with these flow models; how do we get a good result without running it for an impractical amount of time?
Tom: They’re also emphasizing that the model can perform scaffold decoration really well, which means keeping the core structure intact while making precise edits at substituent sites.
Jane: So, they are saying that Morph isn't just about generating molecules; it’s about giving us a flexible tool for lead optimization where we can modify structures without losing the essential core.
Lu: The implication is that this type of flexible modeling could become a really promising building block for accelerating discovery in chemistry because it handles the complexity of changing molecular size dynamically.
Meng: It sounds like they acknowledge the training expense is high, but they are betting that the gain in steerability and OOD generation justifies that cost for finding new chemical space.
Tom: So, to summarize this part: Morph is flexible because it can grow or shrink on the fly, and its improvements focus on making that control more explicit through better transport alignment and structural priors.
Jane: And the big question they leave us with is how to make this powerful tool fast enough for actual use without sacrificing all that flexibility.
Conclusion: Tom: So we've covered Morph’s ability to do insertions, deletions, and substitutions in one go for geometric graphs.
Jane: That's right, it’s this flexible model that adapts its structure during generation instead of being stuck at a fixed size.
Lu: The core idea is using Unbalanced Optimal Transport to align those different-sized graphs before the interpolation starts.
Meng: It means they are training the AI to learn how to smoothly move between any two molecules, regardless of their atom count.
Lalam: I’m confident in this; it’s a very solid way to model that kind of structural evolution in a generative setting.
Tom: And we saw that it can generate samples outside the training data when you condition it on things like logP or QED score, which is pretty significant for finding new compounds.
Jane: That means we aren't just generating what we’ve seen before; we can steer the AI towards regions of chemical space that are entirely new.
Lu: The implication is that this approach could be a strong building block for accelerating discovery because it handles the complexity of changing molecular size dynamically.
Meng: I still have my concerns about running it; if you have to do a lot of integration steps, what does that mean for deployment time and cost?
Tom: They acknowledge the training expense is higher due to all those combinatorics involved across different numbers of atoms.
Jane: But they are also showing how this flexible generation can be used effectively for scaffold decoration, keeping the main structure while tweaking side chains.
Lu: Their future work is focused on figuring out how to make this system scalable and balancing that flexibility against needing many integration steps for a full molecule.
Meng: So, we're looking at a model that's incredibly expressive but needs better optimization for real-world speed.
Tom: That’s the tension here; high steerability versus efficient computation on production hardware.
Jane: We’ll be watching how they tackle those trade-offs in their next set of work, because having a flexible tool is great, but it has to be usable too.
More episodes
- 2603.28595-Optimistic Actor-Critic with Parametric Policies for Linear Markov Decision Processes
- 2603.29765-Training-Free Dynamic Upcycling of Expert Language Models
- 2605.00778-Observable Performance Does Not Fully Reflect Adaptive System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
- 2605.06906-TraXion: Rethinking Pre-training Frameworks for Mobility and Beyond
- 2605.09165-Sparse Layers are Critical to Scaling Looped Language Models
- 2605.09439-Inverse Design for Conditional Distribution Matching
- 2605.10313-Signature Approach for Contextual Bandits with Nonlinear and Path-dependent Rewards: Regret Bound and Statistical Inference
- 2605.10878-Minimum Neural Weight Norm = Kolmogorov Complexity
- 2605.11617-MIST: Reliable Streaming Decision Trees for Online Class-Incremental Learning via McDiarmid Bound
- 2605.15085-From Data to Action: Accelerating Refinery Optimization with AI