🧬 biology

Flexible Flows for Biological Sequence Design

This paper introduces Flexible Flows, a Discrete Flow Matching framework that enhances biological sequence design by incorporating structured couplings, latent edit-based rate parameterization for variable-length generation, and latent classifier-free guidance to achieve state-of-the-art performance in density estimation and conditional sequence generation.

Original authors: Yogesh Verma, Dani Korpela, Harri Lähdesmäki, Vikas Garg

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yogesh Verma, Dani Korpela, Harri Lähdesmäki, Vikas Garg

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to design a new recipe for a biological dish, like a protein or a strand of DNA. The ingredients are letters (A, C, G, T for DNA; 20 different amino acids for proteins), and the "recipe" is the specific order of these letters.

The problem is that the number of possible recipes is astronomically huge, but only a tiny fraction of them actually work. If you just start mixing ingredients randomly, you'll almost certainly end up with something inedible (or in biology's case, non-functional).

This paper introduces a new tool called FlexFlow to help design these biological recipes. Here is how it works, using simple analogies:

1. The Old Way: The Blindfolded Chef

Previous methods for designing these sequences were like a chef who is blindfolded and told to change one ingredient at a time. They might swap a tomato for a potato, but they have no idea if that swap makes sense in the context of the whole dish.

The Problem: These methods often make "uninformed" swaps. They treat every letter as if it's equally likely to be replaced by any other letter, ignoring the fact that in nature, some swaps are common (like swapping a lemon for a lime) while others are impossible (like swapping a lemon for a brick).
The Result: They struggle to create long, complex recipes or to control exactly what the final dish looks like.

2. The New Way: FlexFlow (The Smart Sous-Chef)

FlexFlow is like a smart sous-chef who knows the rules of the kitchen (biology) and helps the main chef navigate the process. It uses three main tricks:

Trick A: The "Biological Map" (Structured Coupling)

Instead of starting with a random pile of ingredients, FlexFlow starts with a "tilted" map.

The Analogy: Imagine you are trying to find a specific house in a giant city. A random search is hard. But if you have a map that highlights the neighborhoods where houses usually look like the one you want, you start your search there.
How it works: FlexFlow uses a "map" based on real biological data (like how amino acids naturally swap in evolution). It biases the starting point of the design process toward regions that are biologically plausible. It doesn't force the design, but it gently nudges it away from impossible combinations right from the start.

Trick B: The "Magic Eraser and Pen" (Edit-Based Flows)

Sometimes, to get a recipe right, you don't just swap an ingredient; you need to add a new one or delete a bad one.

The Analogy: Imagine writing a story. Old methods could only change a word to another word. FlexFlow is like having a magic pen that can insert a whole new sentence, delete a paragraph, or swap a word, all while keeping the story's main theme intact.
How it works: This allows FlexFlow to create sequences of variable lengths. It can grow a sequence by adding letters or shrink it by removing them, which is crucial because biological sequences aren't always the same size.

Trick C: The "Volume Knob" (Latent Guidance)

Sometimes you want to tell the chef, "Make it spicier" or "Make it sweeter."

The Analogy: Imagine you are tuning a radio. Old methods tried to adjust the volume by turning a separate knob for every single speaker (every single letter in the sequence), which often led to static and noise. FlexFlow turns a single "Master Volume" knob in a hidden control room (the "latent space").
How it works: By adjusting this single hidden knob, the entire sequence changes coherently. If you want the protein to bind to a specific target, you turn the knob, and the whole recipe shifts together to match that target, rather than changing letters randomly.

3. The "Budget Controller" (Temperature Scaling)

The paper also introduces a way to control how the chef changes the recipe.

The Analogy: Imagine you have a budget for changes. You can tell the chef, "For this recipe, you are only allowed to add ingredients, no deleting," or "You can only swap ingredients."
How it works: The system uses a "temperature" setting to decide whether to insert, delete, or substitute letters. This gives researchers fine-grained control over the generation process without needing to retrain the whole system.

What Did They Prove?

The authors tested FlexFlow on three main tasks:

DNA Design: Creating DNA sequences that act as "enhancers" (switches that turn genes on) or "promoters" (start buttons for genes). FlexFlow created better, more realistic DNA than previous methods.
Peptide Design: Creating short protein chains (peptides) that stick to specific immune system molecules (MHC). This is like designing a key that fits a specific lock. FlexFlow created keys that fit much better than those made by other tools.
Toy Problems: They tested it on simple math patterns (like spirals and checkerboards) to prove it understands complex shapes and structures.

The Bottom Line

FlexFlow is a new, smarter way to design biological sequences. Instead of guessing randomly, it uses a map of biological rules, allows for adding or removing parts of the sequence, and uses a single "master control" to steer the design toward specific goals. The result is a tool that creates more accurate and functional biological sequences than current state-of-the-art methods.

Technical Summary: Flexible Flows for Biological Sequence Design (FlexFlow)

Problem Statement

Designing functional biological sequences (DNA, RNA, proteins, peptides) requires navigating vast, discrete combinatorial spaces under strict evolutionary and biophysical constraints. While autoregressive models (e.g., ProtGPT2, ESM-2) have demonstrated viability, they suffer from compounding errors, slow sequential sampling, and limited global coherence, particularly for long or structurally complex sequences.

Alternative generative frameworks, such as Discrete Diffusion and Discrete Flow Matching (DFM), offer faster, non-autoregressive sampling. However, existing discrete methods for biological applications face three critical limitations:

Limited Flexibility: They struggle with variable-length sequence generation and lack fine-grained control over the generation process.
Uninformative Couplings: They rely on uniform or masked couplings that ignore the structured, domain-specific relationships between biological sequence elements (e.g., evolutionary substitution biases).
Incoherent Guidance: Steering generation toward specific functional properties often relies on independent, token-wise adjustments that fail to capture the global dependencies necessary for complex regulatory or binding motifs.

Methodology: FlexFlow

The authors propose FlexFlow, a Discrete Flow Matching framework designed to address these limitations through three core architectural innovations:

1. Structured Forward Process (Structured Coupling)

Instead of using a non-informative baseline for the forward noising process, FlexFlow imposes a structural inductive bias by "tilting" the source distribution.

Mechanism: The method replaces standard uniform couplings with a structured coupling $\pi(\tilde{x}_0, x_1, K_\gamma)$ . This is defined by a transition kernel $K_\gamma$ that encodes domain-specific substitution preferences (e.g., BLOSUM62 for proteins, JC69/HKY85 for DNA/RNA).
Effect: The source distribution is biased toward biologically plausible substitution neighborhoods rather than the full simplex. This steers the generative path toward preferred functional regions without modifying the flow objective or training procedure. The coupling generalizes across sequence taxonomies simply by swapping the kernel $K_\gamma$ .

2. Latent Edit-based Rate Parameterization

To support variable-length generation, the model parameterizes the reverse flow rates via edit operations (insertion, deletion, substitution) conditioned on a shared global latent variable.

Hierarchical Parameterization: Given a noisy sequence $x_t$ , a global latent context $r$ is inferred via an encoder $\Psi(x_t)$ . This latent $r$ conditions the categorical distribution of edit operations $p_\Phi(op_i | x_i^t, r)$ at each position.
Tractability: This hierarchical approach induces coupling between token-level operations through the shared latent $r$ while maintaining conditional independence across positions, ensuring the model remains tractable.
Test-Time Control (Dirichlet Priors): The operation probabilities are rescaled using position-dependent temperature parameters $\beta_i$ , modeled via a Dirichlet prior. This allows for post-hoc control over the relative rates of insertion, substitution, and deletion during inference (e.g., acting as a "budget controller" for edit operations) without retraining.

3. Latent Classifier-Free Guidance

The authors introduce a novel guidance mechanism that operates entirely within the continuous latent space, contrasting with standard rate-space guidance.

Mechanism: During training, the encoder $\Psi$ is trained to predict both a conditional latent $r_c$ (given condition $c$ ) and an unconditional latent $r_\emptyset$ . At test time, the guided latent is computed as $r_{LG} = r_\emptyset + w(r_c - r_\emptyset)$ .
Advantages: This approach ensures global structural consistency because the guidance signal propagates through the shared latent $r$ to all per-token operations simultaneously. It avoids the incoherence of token-wise guidance and preserves the non-negativity of rates by construction.
Theoretical Basis: The authors prove that this latent interpolation corresponds to a gradient ascent step on the log-likelihood of an implicit classifier in the latent space.

Key Contributions

Structured Coupling: Introduction of a biologically informed prior that steers the generative path toward functional neighborhoods, replacing uniform couplings.
Latent Edit-based Rates: A parameterization enabling variable-length generation with controllable edit operations via a shared global latent and Dirichlet-prior temperature scaling.
Latent Guidance: A guidance scheme operating in continuous latent space to ensure global coherence, presented as an alternative to rate-space classifier-free guidance.
New Benchmark: The introduction of a peptide-MHC II benchmark for generating and evaluating MHC-conditioned peptide sequences.
Empirical Validation: Demonstration of state-of-the-art performance across density estimation, unconditional/conditional DNA generation, and peptide generation.

Experimental Results

The authors evaluated FlexFlow on diverse tasks, comparing it against diffusion and flow baselines (e.g., Bit Diffusion, Dirichlet FM, Fisher FM, Gumbel-Softmax FM).

Toy Datasets: On 1D and 2D discrete patterns (Sine-wave, Spiral, Rings, Checkerboard), FlexFlow achieved the lowest Wasserstein distance ( $W_2$ ), accurately reconstructing topological features.
Unconditional DNA Design (Enhancers): Evaluated on fly brain and human melanoma enhancer datasets. FlexFlow achieved the lowest Fréchet Biological Distance (FBD) (3.4 for Melanoma, 4.2 for Fly Brain), outperforming Dirichlet FM and Fisher FM.
Conditional DNA Design (Promoters): Given a transcription profile, the model generated promoter sequences. FlexFlow with Latent Guidance achieved the lowest Mean Squared Error (MSE) of 0.022, outperforming Dirichlet FM (0.026) and DDSM (0.033).
Peptide-MHC II Generation: On the new peptide-MHC II benchmark, FlexFlow with Latent Guidance achieved a DeepMHCII discriminator score of 0.66 (vs. 0.21 for Dirichlet FM) and a Fréchet Protein Distance (FPD) of 0.84. The results indicate that Latent Guidance prioritizes binding plausibility, while Rate Guidance offers a different trade-off favoring distributional coverage.
Efficiency: Inference speed was comparable to baselines (approx. 4.0 seconds for 500-length DNA sequences over 100 steps).

Significance and Claims

The paper claims that FlexFlow addresses the key limitations of existing discrete generative methods by integrating biological domain knowledge directly into the flow matching framework. By utilizing structured couplings, the method biases the generation toward plausible biological regions without altering the training objective. The latent edit-based parameterization and latent guidance scheme provide a mechanism for flexible-length generation and coherent, fine-grained control that previous methods lacked.

The authors position FlexFlow as a robust framework for biological sequence design that outperforms strong diffusion and flow-matching baselines in both unconditional and conditional settings. They note that future work will focus on adapting distillation-based training objectives to enable high-quality sequence generation in fewer sampling steps. The work is presented as a preprint, with the authors acknowledging support from the Research Council of Finland and other foundations for related AI in biodesign projects.