Designing Cell-Type-Specific Regulatory DNA with Guided Discrete Diffusion
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Designing Cell-Type-Specific Regulatory DNA with Guided Discrete Diffusion".
Marcus: Here's the gist of the paper. Designing regulatory DNA elements with precise cell-type-specific activity is broady relevant for cell engineering and gene therapy.
Ines: Here, we introduce DNA-CRAFT, a generative framework that integrates class-conditioned discrete diffusion with Monte Carlo tree search to design cell-type-specific and biologically faithful regulatory elements.
Marcus: By benchmarking DNA-CRAFT on regulatory sequence design tasks for human cell lines and immune cell types, we demonstrate that our model generates sequences with high predicted cell-type-specific activity and biological fidelity, achieving the best trade-offs compared to methods that use diffusion, autoregressive models, and gradient-based optimization.
Ines: Framework Overview DNA-CRAFT is a generative framework that integrates class-conditioned discrete diffusion with Monte Carlo tree search to design cell-type-specific and biologically faithful regulatory elements.
Title and authors: Ines: So we're looking at this paper called Designing Cell-Type-Specific Regulatory DNA with Guided Discrete Diffusion. It’s about building these sequences that work perfectly in one cell type and not another.
Marcus: Right. It tackles the problem of getting precise cell type specificity while still respecting how DNA normally functions, which is a big challenge for any gene therapy or even cell engineering project.
Yuki: From a population genetics standpoint, I'm interested in whether these models can actually capture the real biological diversity we see across different species and cell states.
Ines: The paper introduces DNA-CRAFT as this new framework that combines class-conditioned discrete diffusion with Monte Carlo tree search to handle both the specificity and the biological fidelity at once.
Marcus: Basically, they first train a diffusion model on a huge set of regulatory elements from ENCODE, which is over three point two million sequences from human and mouse genomes.
Yuki: So they are using that massive dataset to learn the basic building blocks of how DNA regulates itself across species.
Ines: That training involves a masked diffusion language model, or MDLM, which uses a bidirectional Mamba state space model as its backbone for denoising the data.
Marcus: Then they use something called discrete classifier-free guidance, or D-CFG, to condition that model so it learns the specific grammar for different regulatory classes like enhancers and promoters.
Yuki: That’s important because it means they aren't just guessing sequences; they are learning the natural patterns for those different element types.
Ines: And then they add this inference-time thing called conditional Monte Carlo tree guidance, or MCTG, which is their core innovation for cell type specificity.
Marcus: That MCTG is designed to maximize the difference in regulatory activity between your desired cell type and all the undesired ones you want to avoid.
Yuki: So it's like a guided search where the reward function explicitly pushes the sequence toward being active in one place while shutting down activity elsewhere.
Ines: They define this reward using something called a MinGap Score, which is basically maximizing the margin between desired and undesired cell types in terms of their mean activity.
Marcus: That’s where you see the practical application—they test it on generating enhancers for three human cell lines and sequences across eight immune cell types.
Yuki: The results they show are pretty strong, indicating that DNA-CRAFT finds a good balance between being very specific and still looking like real biological DNA.
Ines: They found that DNA-CRAFT achieves the best trade-off compared to other methods using diffusion or autoregressive models for this regulatory sequence design task.
Marcus: Specifically, they showed it hits higher predicted specificity scores than some diffusion alignment methods and fine-tuning approaches like DRAKES across all those cell lines.
Yuki: And in the immune cell study, they claimed DNA-CRAFT was the only method to get a positive MinGap score for sequences that decoupled chromatin accessibility between CD8 plus and CD4 plus T cells.
Ines: That decoupling is key because it means you can target one cell type without accidentally affecting another closely related type in the immune system.
Marcus: The ablation studies they did really supported this, showing that using both the MinGap score and class-conditioned sampling together improves both biological fidelity and predicted specificity of the resulting sequences.
Yuki: It’s interesting how they showed that just using standard sampling doesn't work consistently for high specificity, which supports their need for the tree guidance.
Ines: They also looked at what happens if you only use class conditioning, showing that adding the MinGap reward is what really boosts specificity and motif correlation compared to just being conditioned on a promoter class.
Marcus: It seems like this framework is robust because it handles both learning the grammar and then actively searching for the specific activity difference simultaneously, rather than optimizing one thing in isolation.
Yuki: So for someone just listening, it means we have a way to generate regulatory sequences that are biologically realistic but also highly tuned to a very narrow set of cell types.
Ines: Exactly. The whole point is that we can design these elements with high predicted activity while still staying within the rules of natural DNA structure.
Marcus: And this moves us past methods where you might get a sequence that looks right statistically but doesn't actually function well in the biological system you are trying to engineer.
Yuki: It opens up possibilities for designing targeted therapies where the regulatory switch only flips in the intended cells, which is a huge step for precision medicine.
Ines: So we have this DNA-CRAFT framework that uses diffusion and tree search to generate sequences that are both faithful to nature and highly cell-type specific.
Marcus: It’s a solid method because it addresses the dual challenge of fidelity and specificity with integrated guidance mechanisms.
Yuki: It’s a powerful tool for understanding how regulatory elements operate across the spectrum of cell types.
Ines: We'll keep an eye on how this framework evolves as we see more applications in designing novel therapeutic targets.
The paper's summary: Ines: So, to recap, this paper is about using a new AI framework called DNA-CRAFT to design DNA sequences that are tailored to work in just one type of cell, like an enhancer for a human liver cell but not a mouse cell or immune cell.
Marcus: Right. They’re taking these complex regulatory elements and trying to make them perfectly specific while still making sure they follow the basic rules of how DNA normally works in nature.
Yuki: What’s interesting is that they aren't just looking at one species, too; they trained this model on both human and mouse data from the ENCODE registry. That should give it a good handle on cross-species grammar.
Ines: Exactly. The training part uses a masked diffusion language model to learn all those natural regulatory patterns across those millions of sequences in both genomes. It’s learning the rules from scratch using real biological data.
Marcus: And they add this discrete classifier-free guidance, or D-CFG, to condition that model so it learns the specific grammar for different element types like promoters versus enhancers. It’s teaching it what a promoter looks like versus an enhancer looks like in a natural sense.
Yuki: I want to ask about how they handle the cell type part. How do you train something to understand which cell type you *want* it to be active in, and which one you *don't* want it active in?
Ines: That’s where the core innovation comes in: conditional Monte Carlo tree guidance, or MCTG. This is an inference-time algorithm that guides the search for the sequence. It’s designed to maximize the difference between what you want and what you don't want.
Marcus: So it’s not just optimizing for one thing, like making it active; it’s actively maximizing a gap in activity between two specific cell types. They use something called a MinGap Score for that reward calculation.
Yuki: That MinGap Score is basically measuring how much more active the desired cell type is compared to all the undesired ones you are trying to avoid. That sounds like it directly addresses the specificity problem we’ve been talking about in gene therapy applications.
Ines: It seems that this combination—the diffusion model learning the grammar and the tree search guided by that specific activity reward—gives them a really good balance. They found it beats other methods using diffusion or autoregressive models on both human cell lines and eight different immune cell types.
Marcus: And in those immune cell studies, they claimed this method was the only one that could successfully create sequences where the CD8 plus T cells were active while keeping them off for CD4 plus T cells, which is a big win statistically.
Yuki: So what this means for us in terms of biology is that we can actually design regulatory switches that are highly tuned to a specific cell's needs without completely breaking the biological rules of how those sequences function.
Ines: It means we can engineer things with much tighter control over when and where a gene turns on, which is huge for precision medicine in cell engineering and gene therapy.
Marcus: But they did flag one thing; they showed that just using the MinGap reward alone didn't work consistently without the class-conditioned sampling. You need both parts working together to get those high specificity scores.
Yuki: That’s important for caution. It shows that while the guidance is powerful, you still need that foundation of learning natural grammar from a massive dataset to make sure you don't just generate random noise that happens to score high on the specificity metric.
Ines: So the main point is this framework handles both fidelity—staying true to biology—and specificity—hitting those exact cell types—simultaneously, which is a much harder task than trying to do either one alone.
Marcus: It’s a solid method because it combines that learning aspect with the directed search. We should keep an eye on how they apply this to designing sequences for other complex diseases in the future.
The paper's improvements: Tom: So, we're looking at what the authors suggest as improvements to their DNA-CRAFT framework for designing those cell type specific regulatory sequences.
Ines: What they’re saying is that combining the class conditioned sampling with that MinGap specificity score reward really makes things better. They show that using both together improves both how biologically faithful the sequences are and how specific they are predicted to be.
Marcus: That’s a key point because we saw in the earlier data that standard sampling didn't give you consistent high specificity scores, so they’re showing us that the tree search guidance is really necessary. It’s not enough just to condition on the class; you also need that reward function pushing it toward your target cell type.
Yuki: From a population perspective, if they are getting better fidelity and specificity together, it means we can design regulatory elements that are not just statistically active in a lab setting but actually behave as expected in the living system. That’s crucial for understanding how these elements shape cell fate across different species.
Ines: Exactly. They also showed that conditioning on a specific class, like the enhancer class, actually helps improve motif correlation and diversity compared to just doing an unconditional or even a promoter-conditional generation run. It means you can be more precise about which biological features you’re generating.
Marcus: And they pointed out that because they trained this thing on both human and mouse genomes, it should capture cross species regulatory grammar better than if it were only trained on one organism. That’s a big data win for generalization.
Yuki: That cross-species capability is what makes the whole thing interesting for evolution—it suggests the underlying logic of these regulatory elements is similar across mammals, even if the specific implementation differs slightly between human and mouse.
Ines: So to summarize, they suggest that this integrated approach—class conditioning plus a specificity reward guiding the tree search—is actually better than using either technique in isolation for achieving high quality regulatory design.
Marcus: It’s moving away from methods where you have to pick one thing, like just maximizing activity or just learning the grammar, and instead using both to get that sweet spot between function and precision.
Yuki: It opens up a new way of thinking about how we manipulate these elements in cell engineering; it’s not just about making something work, it’s about making it work in the exact right context.
Conclusion: Tom: So we're wrapping up this look at "Designing Cell-Type-Specific Regulatory DNA with Guided Discrete Diffusion." Basically, the paper shows how combining class conditioned discrete diffusion with Monte Carlo tree search lets us design regulatory DNA that is both specific and faithful to biology.
Ines: Right. The main finding is that this combined method achieves a really good trade off between having high predicted cell type specificity and maintaining biological fidelity compared to other sequence design methods. It’s a solid way to get the right sequence without just generating random noise.
Marcus: I think what stands out most for the cohort folks is how they handled the batch effect implicitly by training on such a massive dataset from both human and mouse genomes, which should help their results generalize better across different cell types.
Yuki: I agree with Marcus; having that cross species data really supports the idea that these regulatory elements have some fundamental logic shared across mammals, even if the specific binding sites change. That’s important for understanding how we can predict function in novel contexts.
Ines: It means we're getting sequences that are not only specific to one cell type but also structurally sound according to natural grammar, which is exactly what you need for any real gene therapy application.
Marcus: And I still have that caveat about the ablation study; they made it clear that you absolutely need both the MinGap reward function and the class conditioned sampling working together for those highest specificity scores. You can't just rely on one part of the system.
Yuki: So, while this framework is powerful, it doesn't solve everything; you still have to manage your training data quality and ensure you’re using both components correctly to get the best results from the DNA-CRAFT model.
Ines: Exactly. It’s a sophisticated tool for generating regulatory sequences that are tailored for precision medicine in cell engineering, but it definitely leaves room for refinement in how we condition those models further.
Marcus: We should keep an eye on how they use this framework to tackle more complex biological systems, maybe moving beyond simple enhancers into whole promoter designs.
Yuki: I think the next step is seeing if they can apply this same logic to predict how these designed elements will interact with other regulatory networks in a cell.
Medical University of Vienna · CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences
q-bio.GN
Submitted: 2026-04-22
Updated: 2026-11-11
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Designing regulatory DNA elements with precise cell-type-specific activity is broady relevant for cell engineering and gene therapy.
Key concepts
- Discrete Diffusion Model
- This is a generative AI model trained on millions of natural DNA regulatory elements using a Mamba state space model. It learns the underlying rules or 'grammar' of how DNA sequences function as enhancers or promoters, allowing it to create new, realistic sequences.
- Class-Conditioned Sampling
- The model is conditioned on specific functional classes (like 'Enhancer' or 'Promoter'). This ensures that when designing a sequence, the generation process adheres to the characteristic regulatory grammar associated with that specific DNA element type, improving biological relevance.
- Conditional Monte Carlo Tree Guidance (MCTG)
- This is an inference-time algorithm used during sequence design. It guides the search for new sequences by using a reward function that explicitly maximizes the difference in activity between the target cell type and undesired cell types, directly driving cell-type specificity.
- MinGap Score
- This reward function measures how well a designed sequence separates the desired and undesired cell types. It calculates a margin based on predicted regulatory activity, aiming to maximize this gap to ensure high specificity in designing sequences for different immune cells.
Terminology
Summary
Designing regulatory DNA elements with precise cell-type-specific activity is broady relevant for cell engineering and gene therapy. p. 0 Published at the MLGenX workshop, ICLR 2026
CONDITIONAL MONTE CARLO TREE DIFFUSION FOR
DESIGNING CELL-TYPE-SPECIFIC AND BIOLOGICALLY
FAITHFUL REGULATORY DNA
Animesh Awasthi1,2
, Raphael Bednarsky1,2
, Moritz Schaefer1,2 & Christoph Bock1,2∗
1Medical University of Vienna, Institute of Artificial Intelligence,
Center for Medical Data Science, Vienna, Austria
2CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences,
Vienna, Austria
ABSTRACT
Designing regulatory DNA elements with precise cell-type-specific activity is
broadly relevant for cell engineering and gene therapy. Deep generative models
can generate functional gene-regulatory elements, but existing methods struggle to achieve high specificity against undesired cell types while adhering to the
genome’s natural regulatory grammar. Deep generative models
can generate functional gene-regulatory elements, but existing methods struggle to achieve high specificity against undesired cell types while adhering to the
genome’s natural regulatory grammar. Here, we introduce DNA-CRAFT, a generative framework that integrates class-conditioned discrete diffusion with Monte
Carlo tree search to design cell-type-specific and biologically faithful regulatory
elements. Deep generative models
can generate functional gene-regulatory elements, but existing methods struggle to achieve high specificity against undesired cell types while adhering to the
genome’s natural regulatory grammar. Here, we introduce DNA-CRAFT, a generative framework that integrates class-conditioned discrete diffusion with Monte
Carlo tree search to design cell-type-specific and biologically faithful regulatory
elements. Third, we employ conditional Monte Carlo
tree guidance, an inference-time alignment algorithm designed to maximize the
differential regulatory activity between desired and undesired cell types. By benchmarking DNA-CRAFT on regulatory sequence design tasks for human cell lines
and immune cell types, we demonstrate that our model generates sequences with
high predicted cell-type-specific activity and biological fidelity, achieving the best
trade-offs compared to methods that use diffusion, autoregressive models, and
gradient-based optimization. Third, we employ conditional Monte Carlo
tree guidance, an inference-time alignment algorithm designed to maximize the
differential regulatory activity between desired and undesired cell types. By benchmarking DNA-CRAFT on regulatory sequence design tasks for human cell lines
and immune cell types, we demonstrate that our model generates sequences with
high predicted cell-type-specific activity and biological fidelity, achieving the best
trade-offs compared to methods that use diffusion, autoregressive models, and
gradient-based optimization.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Framework Overview
DNA-CRAFT is a generative framework that integrates class-conditioned discrete diffusion with Monte Carlo tree search to design cell-type-specific and biologically faithful regulatory elements. The framework involves three main steps:
-
Training a discrete diffusion model on the ENCODE registry of 3. p. 0
2 million candidate regulatory elements. p. 0Where the paper says it
Here, we introduce DNA-CRAFT, a generative framework that integrates class-conditioned discrete diffusion with Monte Carlo tree search to design cell-type-specific and biologically faithful regulatory elements. We first train a discrete diffusion model on the ENCODE registry of 3.
p. 0Where the paper says it
We first train a discrete diffusion model on the ENCODE registry of 3. 2 million candidate regulatory elements.
Where the paper says it
We first train a discrete diffusion model on the ENCODE registry of 3. 2 million candidate regulatory elements.
-
Conditioning the model to learn class-specific regulatory grammars of naturally occurring DNA sequences, including enhancers and promoters. p. 0
Where the paper says it
2 million candidate regulatory elements. Second, we condition the model to learn class-specific regulatory grammars of naturally occurring DNA sequences, including enhancers and promoters.
-
Employing conditional Monte Carlo tree guidance, an inference-time alignment algorithm designed to maximize the differential regulatory activity between desired and undesired cell types. p. 0
Where the paper says it
Second, we condition the model to learn class-specific regulatory grammars of naturally occurring DNA sequences, including enhancers and promoters. Third, we employ conditional Monte Carlo tree guidance, an inference-time alignment algorithm designed to maximize the differential regulatory activity between desired and undesired cell types.
How the training signal is built
The training process involves several specialized components to build a robust generative model.
** Masked Diffusion Language Model (MDLM):**
The MDLM is used to learn the genome’s regulatory grammar and is trained on the ENCODE registry of regulatory elements, comprising over 3.2 million DNA sequences from the human and mouse genomes. p. 2 We train this model on the ENCODE registry of regulatory elements, comprising
over 3. 2 million DNA sequences from the human and mouse genomes (Figure 1a). We train this model on the ENCODE registry of regulatory elements, comprising
over 3. 2 million DNA sequences from the human and mouse genomes (Figure 1a). We use the
bidirectional Mamba Gu et al. (2022) state space model backbone for the denoising network xθ. We use the
bidirectional Mamba Gu et al. (2022) state space model backbone for the denoising network xθ.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Discrete Classifier-Free Guidance (D-CFG):
D-CFG constructs a class-conditioned MDLM by training on natural regulatory DNA elements and paired regulatory classes. The model is trained alongside an unconditional one to learn the conditional denoising network. p. 3 2 CLASSIFIER-FREE GUIDANCE FOR CLASS-CONDITIONED DIFFUSION MODELS
D-CFG constructs a class-conditioned MDLM by training on natural regulatory DNA elements and
paired regulatory classes. This is achieved by training a conditional denoising diffusion network
pθ(z
1:L
s
|y, z
1:L
t
) alongside an unconditional one pθ(z
1:L
s
|z
1:L
t
), where y ∈ {1, .Where the paper says it
Class Consolidation:
The fine-grained ENCODE V4 classifications are consolidated into five broad functional classes for DNA-CRAFT conditioning.
Where the gains come from
The core innovation lies in the conditional Monte Carlo tree guidance (MCTG) which adapts to regulatory element design. This adaptation introduces two key innovations for cell-type-specific regulatory sequence design:
(i) Constraining the tree search to the desired regulatory element class using conditional diffusion sampling because different classes of regulatory elements implement characteristic regulatory grammars.
(ii) Guiding the tree search using a reward that explicitly maximizes the differential activity between desired and undesired cell types to achieve high cell type specificity. p. 1 (2024), we constrain the tree search to the desired class using conditional
diffusion sampling. (ii) To achieve high cell type specificity, we guide the tree search using a reward
that explicitly maximizes the differential activity between desired and undesired cell types. (2024), we constrain the tree search to the desired class using conditional
diffusion sampling. (ii) To achieve high cell type specificity, we guide the tree search using a reward
that explicitly maximizes the differential activity between desired and undesired cell types.Where the paper says it
Where the paper says it
The reward calculation uses the MinGap Score to maximize the margin between desired and undesired cell types, defined as g(xi) = 1/C+ X c∈C+ sc(xi) Mean Activity (Desired) − max c∈C− sc(xi) Undesired Cell Types. p. 4 To enhance cell type specificity, we maximize the margin between desired
and undesired cell types rather than simply selecting dominant sequences. We employ the MinGap
Score Gosai et al.Where the paper says it
Evaluation and Results
DNA-CRAFT was benchmarked on two tasks: generating cell type specific enhancers for three human cell lines and generating differentially chromatin accessible sequences across eight immune cell types. The results show that DNA-CRAFT effectively navigates these complex design spaces, achieving the best trade-off between predicted specificity and biological faithfulness compared to state-of-the-art methods. p. 2 Published at the MLGenX workshop, ICLR 2026
We benchmark DNA-CRAFT on two broadly relevant tasks of regulatory DNA design: (i) Generating
cell type specific enhancers for three human cell lines and (ii) generating differentially chromatinaccessible sequences across eight immune cell types. Our results show that DNA-CRAFT effectively
navigates these complex design spaces, achieving the best trade-off between predicted specificity and
biological faithfulness compared to state-of-the-art methods that use diffusion, autoregressive models,
and gradient-based optimization. Published at the MLGenX workshop, ICLR 2026
We benchmark DNA-CRAFT on two broadly relevant tasks of regulatory DNA design: (i) Generating
cell type specific enhancers for three human cell lines and (ii) generating differentially chromatinaccessible sequences across eight immune cell types. Our results show that DNA-CRAFT effectively
navigates these complex design spaces, achieving the best trade-off between predicted specificity and
biological faithfulness compared to state-of-the-art methods that use diffusion, autoregressive models,
and gradient-based optimization.Where the paper says it
Where the paper says it
Cell Line Specific Enhancer Design:
In Table 2, DNA-CRAFT achieves higher predicted specificity scores compared to other diffusion-based alignment methods (SMC, TDS, CG) and fine-tuning approaches (DRAKES) across all cell lines.
Immune Cell State Specific Sequences:
In Table 3, DNA-CRAFT is the only method that achieves a positive MinGap score, successfully decoupling chromatin accessibility in CD8+ / CD4+ T cells from closely related cell types.
Ablation Studies
The ablation study suggests that using both the MinGap specificity score and class-conditioned sampling, as incorporated in DNA-CRAFT, improves the biological fidelity and predicted cell-type specificity of the designed sequences.
MinGap Reward Ablation:
Standard sampling fails to consistently generate sequences with high specificity, highlighting the need for tree guidance. Replacing the MinGap set with the Pareto front yielded unspecific sequences since Pareto optimization retains sequences that excel in any one of the objectives, irrespective of their performance in other objectives. p. 1 they struggle to simultaneously optimize for two critical objectives: (i) achieving high activity in the
desired cell type(s) while minimizing activity in a potentially large number of undesired cell types,
and (ii) designing DNA sequences that closely resemble naturally occurring regulatory DNA, thereby
reducing safety concerns in the context of medical applications DaSilva et al. DNA sequence optimization methods Laarhoven & Aarts (1987); Angermueller et al.Where the paper says it
Regulatory Sequence Class Guidance:
Table 10 shows that the ”Enhancer” class conditioning indeed improved specificity, as well as motif correlation and diversity compared to unconditional and ”Promoter”-conditional generation.
Conclusion
DNA-CRAFT combines the generative fidelity of class-conditioned discrete diffusion with the directed exploration of specificity-driven tree search, providing a framework for designing regulatory elements that adhere to natural grammar while achieving high cell-type specificity. The study concludes that DNA-CRAFT effectively navigates the complex landscape of regulatory activity across multiple cell types, generating regulatory elements of the desired class that achieve highly cell-type-specific activity without sacrificing biological fidelity. p. 7 In
summary, these results demonstrate that DNACRAFT effectively navigates the complex landscape of regulatory activity across multiple cell
types, generating regulatory elements of the desired class that achieve highly cell-type-specific
activity without sacrificing biological fidelity. 4 CONCLUSION & DISCUSSION
We introduced DNA-CRAFT, a generative framework for designing regulatory DNA elements that
maximize predicted cell-type specificity while adhering to the natural regulatory grammars of the
genome.Where the paper says it
--- Page 1 ---
Published at the MLGenX workshop, ICLR 2026
CONDITIONAL MONTE CARLO TREE DIFFUSION FOR DESIGNING CELL-TYPE-SPECIFIC AND BIOLOGICALLY FAITHFUL REGULATORY DNA Animesh Awasthi1,2 Raphael Bednarsky1,2 Moritz Schaefer1,2 & Christoph Bock1,2∗ 1Medical University of Vienna, Institute of Artificial Intelligence, Center for Medical Data Science, Vienna, Austria 2CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences, Vienna.
ABSTRACT Designing regulatory DNA elements with precise cell-type-specific activity is broadly relevant for cell engineering and gene therapy. p. 0 Published at the MLGenX workshop, ICLR 2026
CONDITIONAL MONTE CARLO TREE DIFFUSION FOR
DESIGNING CELL-TYPE-SPECIFIC AND BIOLOGICALLY
FAITHFUL REGULATORY DNA
Animesh Awasthi1,2
, Raphael Bednarsky1,2
, Moritz Schaefer1,2 & Christoph Bock1,2∗
1Medical University of Vienna, Institute of Artificial Intelligence,
Center for Medical Data Science, Vienna, Austria
2CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences,
Vienna, Austria
ABSTRACT
Designing regulatory DNA elements with precise cell-type-specific activity is
broadly relevant for cell engineering and gene therapy. Deep generative models
can generate functional gene-regulatory elements, but existing methods struggle to achieve high specificity against undesired cell types while adhering to the
genome’s natural regulatory grammar. Published at the MLGenX workshop, ICLR 2026
CONDITIONAL MONTE CARLO TREE DIFFUSION FOR
DESIGNING CELL-TYPE-SPECIFIC AND BIOLOGICALLY
FAITHFUL REGULATORY DNA
Animesh Awasthi1,2
, Raphael Bednarsky1,2
, Moritz Schaefer1,2 & Christoph Bock1,2∗
1Medical University of Vienna, Institute of Artificial Intelligence,
Center for Medical Data Science, Vienna, Austria
2CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences,
Vienna, Austria
ABSTRACT
Designing regulatory DNA elements with precise cell-type-specific activity is
broadly relevant for cell engineering and gene therapy. Deep generative models
can generate functional gene-regulatory elements, but existing methods struggle to achieve high specificity against undesired cell types while adhering to the
genome’s natural regulatory grammar. Here, we introduce DNA-CRAFT, a generative framework that integrates class-conditioned discrete diffusion with Monte
Carlo tree search to design cell-type-specific and biologically faithful regulatory
elements. We first train a discrete diffusion model on the ENCODE registry of
3. 2 million candidate regulatory elements. Second, we condition the model to
learn class-specific regulatory grammars of naturally occurring DNA sequences,
including enhancers and promoters. Second, we condition the model to
learn class-specific regulatory grammars of naturally occurring DNA sequences,
including enhancers and promoters. Third, we employ conditional Monte Carlo
tree guidance, an inference-time alignment algorithm designed to maximize the
differential regulatory activity between desired and undesired cell types.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. p. 1 Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (a) represents the class-conditioned
discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (c)
represents applications of DNA-CRAFT for the design of cell-type-specific regulatory sequences. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (c)
represents applications of DNA-CRAFT for the design of cell-type-specific regulatory sequences. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (c)
represents applications of DNA-CRAFT for the design of cell-type-specific regulatory sequences.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
Designing regulatory DNA elements with precise cell-type-specific activity is broady relevant for cell engineering and gene therapy. The single most informative sentence about the paper's subject and finding is Designing regulatory DNA elements with precise cell-type-specific activity is broady relevant for cell engineering and gene therapy.
The training process involves several specialized components to build a robust generative model.
The MDLM is used to learn the genome’s regulatory grammar and is trained on the ENCODE registry of regulatory elements, comprising over 3.2 million DNA sequences from the human and mouse genomes. The backbone neural network used for denoising is a bidirectional Mamba state space model.
D-CFG constructs a class-conditioned MDLM by training on natural regulatory DNA elements and paired regulatory classes.
The core innovation lies in the conditional Monte Carlo tree guidance (MCTG) which adapts to regulatory element design. This adaptation introduces two key innovations for cell-type-specific regulatory sequence design
(i) Constraining the tree search to the desired regulatory element class using conditional diffusion sampling because different classes of regulatory elements implement characteristic regulatory grammars.
(ii) Guiding the tree search using a reward that explicitly maximizes the differential activity between desired and undesired cell types to achieve high cell type specificity.
DNA-CRAFT was benchmarked on two tasks: generating cell type specific enhancers for three human cell lines and generating differentially chromatin accessible sequences across eight immune cell types. The results show that DNA-CRAFT effectively navigates these complex design spaces, achieving the best trade-off between predicted specificity and biological faithfulness compared to state-of-the-art methods.
The ablation study suggests that using both the MinGap specificity score and classconditioned sampling, as incorporated in DNA-CRAFT, improves the biological fidelity and predicted cell-type specificity of the designed sequences. The results show that standard sampling fails to consistently generate sequences with high specificity, highlighting the need for tree guidance. p. 1 (2025) to learn the regulatory
grammar of promoters, enhancers, and other naturally occurring DNA elements (Figure 1a). Third,
to generate DNA sequences with high predicted cell-type-specific activity, we adapt Monte Carlo
tree guidance Tang et al. Second, we observe distinct motif enrichment signatures for each class, confirming that the
model has learned valid conditional distributions rather than generic regulatory signals. Third, these
signatures align with biological knowledge; for instance, sequences conditioned on the Promoter
class strongly enrich for promoter binding factors such as Nuclear Factor Y (NF-Y) and ETS-family
proteins, whereas the Enhancer class is dominated by lineage-determining factors (e.Where the paper says it
Where the paper says it
--- Page 3 ---
Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (c) represents applications of DNA-CRAFT for the design of cell-type-specific regulatory sequences.
--- Page 4 ---
Published at the MLGenX workshop, ICLR 2026
Figure 1: Overview of the DNA-CRAFT Framework. Panel (a) represents the class-conditioned discrete diffusion model trained on the ENCODE registry of naturally occurring regulatory elements. Panel (b) represents conditional Monte Carlo tree guidance with specificity rewards. Panel (c) represents applications of DNA-CRAFT for the design of cell-type-specific regulatory sequences.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Integration of class-conditioned discrete diffusion with Monte Carlo tree search allows for simultaneous optimization of cell-type specific activity and biological faithfulness, addressing previous limitations where methods either optimize for a single scalar reward or generate sequences that violate natural regulatory grammar.
-
The system can design regulatory DNA elements with high predicted cell-type specificity while adhering to the genome's natural regulatory grammar, achieving the
best trade-offs compared to methods that use diffusion, autoregressive models, and gradient-based optimization.
-
The framework enables the generation of sequences that maximize the margin between desired and undesired cell types using a MinGap Score reward function, leading to sequences with
high predicted specificity scores
across various cell lines. -
The model can be trained on large cross-species datasets from both human and mouse genomes, ensuring that generated regulatory elements capture
cross-species regulatory grammar
and maximize biological diversity. -
By leveraging the class-conditioned sampling, the system can generate sequences rich in biologically relevant motifs, where
sequences conditioned on the Promoter class strongly enrich for promoter binding factors such as Nuclear Factor Y (NF-Y) and ETS-family proteins.
Abstract
AI-driven design of DNA with cell-type-specific regulatory activity is broadly relevant for cell engineering and gene therapy. DNA elements such as promoters and enhancers have gene-regulatory functions that are defined by their DNA sequence and follow a characteristic regulatory grammar. Deep learning models can effectively predict DNA regulatory activity across cell types, but current generative methods often produce DNA sequences that deviate from the native regulatory grammar, increasing the risk of unexpected behavior and off-target effects in real-world applications. Here, we introduce DNA-CRAFT, a method for designing regulatory DNA with high predicted cell-type-specific activity while preserving naturalness. DNA-CRAFT formulates regulatory sequence design as a guided search over a generative model pre-trained on millions of naturally occurring regulatory elements from the human and mouse genomes. Our method enables targeted exploration of the DNA sequence space, efficient evaluation of candidate designs, and explicit suppression of unwanted activity. Across benchmarks spanning human cell lines and primary immune cells, DNA-CRAFT combines high predicted cell-type specificity with strong motif and sequence-composition adherence, achieving favorable results across the evaluated DNA design tasks.
Sources
- Structured Denoising Diffusion Models in Discrete State-Spaces
- A Continuous Time Framework for Discrete Denoising Models
- Ctrl-DNA: Controllable Cell-Type-Specific Regulatory DNA Design via Constrained RL
- Efficiently Modeling Long Sequences with Structured State Spaces
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-Based Decoding
- Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution
- Unlocking Guidance for Discrete State-Space Diffusion and Flow Models
- Particle Denoising Diffusion Sampler
- Simple and Effective Masked Diffusion Language Models
- Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling
- Simple Guidance Mechanisms for Discrete Diffusion Models
- Simplified and Generalized Masked Diffusion for Discrete Data
- AdaLead: A simple and robust adaptive greedy search algorithm for sequence design
- Dirichlet Flow Matching with Applications to DNA Sequence Design
- PepTune: De Novo Generation of Therapeutic Peptides with Multi-Objective-Guided Discrete Diffusion
- Understanding Reinforcement Learning-Based Fine-Tuning of Diffusion Models: A Tutorial and Review
- Inference-Time Alignment in Diffusion Models with Reward-Guided Generation: Tutorial and Review
- Fine-Tuning Discrete Diffusion Models via Reward Optimization with Applications to DNA and Protein Design
- Practical and Asymptotically Exact Conditional Sampling in Diffusion Models
Related papers
- SeekRBP: Leveraging Sequence-Structure Integration with Reinforcement Learning for Receptor-Binding Protein Identification
- Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
- AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
- Leveraging genomic deep learning models for the prediction of non-coding variant effects
- Adaptive Allele-Haplotype Incidence Indexing for Pangenome Queries: Representation, Reconstruction, and Benchmarking