Controllable Sequence Editing for Biological and Clinical Trajectories

Imagine you are watching a movie about a person's life, or perhaps a time-lapse video of a flower blooming. Now, imagine you want to edit that movie. You want to say, "What if this person took a different job?" or "What if this flower was watered with a special nutrient instead of plain water?"

The problem with most current AI tools is that they are terrible editors. If you ask them to change the plot, they often rewrite the entire movie from the beginning, erasing the history you wanted to keep. Or, they might change the wrong things—like changing the character's eye color when you only wanted to change their job.

CLEF (Controllable Sequence Editing Framework) is a new AI tool that acts like a precision surgical editor for time-based stories, whether those stories are about cells growing, patients getting sick, or items being sold in a store.

Here is how it works, broken down into simple concepts:

1. The "Time Travel" Problem

Most AI models are like a car that can only drive forward. If you want to know what happens 10 years from now, they have to drive through every single second, minute, and hour in between. If they make a tiny mistake at year 1, that mistake gets bigger and bigger by year 10, ruining the prediction.

CLEF is different. It has a teleportation device. It can look at where a patient is today, apply a condition (like "start this new medication"), and instantly jump to the result 6 months from now, skipping the messy middle part. It doesn't have to guess every step in between; it calculates the destination directly.

2. The "Scalpel" vs. The "Sledgehammer"

Imagine a patient's health record as a giant dashboard with 50 different gauges (heart rate, blood sugar, white blood cells, etc.).

Old AI (The Sledgehammer): If you tell the AI, "Give the patient insulin," the old AI might think, "Okay, I'll change everything about this patient's future," including their height, their eye color, and their blood pressure, even though insulin only affects blood sugar. It lacks precision.
CLEF (The Scalpel): CLEF understands that insulin only changes blood sugar. It leaves the other 49 gauges exactly as they were. It edits the story only where the intervention happens, preserving the rest of the patient's unique history.

3. The Secret Sauce: "Temporal Concepts"

How does CLEF know exactly what to change and when? It learns something called Temporal Concepts.

Think of a recipe.

Old AI tries to memorize every single meal a chef has ever cooked.
CLEF learns the technique. It understands concepts like "The rate at which a cake rises" or "How fast a sauce thickens."

In CLEF's world, a "Temporal Concept" is a mathematical rule that describes how fast things change between two points in time.

If a patient's blood sugar is rising fast, the concept is "Steep Climb."
If they take medicine, the concept changes to "Gentle Descent."

Because CLEF learns these "techniques" (concepts) rather than just memorizing data, it can apply them to new situations it has never seen before. It's like teaching a chef a new cooking technique; they can now cook a dish you've never shown them, using the same technique.

4. Real-World Magic: The "What-If" Machine

The paper shows CLEF being used in two amazing ways:

Cellular Reprogramming: Imagine a cell is a clay ball. Scientists want to turn it into a heart cell. CLEF can simulate: "What if we add this specific chemical today?" and instantly show the clay morphing into a heart cell, while keeping the rest of the cell's structure intact.
Patient Health (Type 1 Diabetes): The researchers used CLEF on real patient data. They asked, "What if we could magically lower this patient's blood sugar to a healthy level?"
- CLEF generated a "counterfactual" timeline (an alternate reality).
- In this new timeline, the patient's blood sugar dropped, and interestingly, their white blood cell counts also dropped (because high sugar causes inflammation).
- This proves CLEF understands the hidden connections between different parts of the body, not just the numbers on the screen.

5. Why This Matters

Currently, if a doctor wants to know if a new drug will work, they have to wait years to see real-world results.

CLEF acts as a "Virtual Twin." It can build a digital version of a patient, run thousands of "What-If" scenarios in seconds, and tell the doctor, "If you give Drug A now, the patient will likely be healthy in 6 months. If you wait, they might get worse."

Summary

Think of CLEF as a Time-Traveling Editor for life's data.

It doesn't rewrite the past.
It doesn't break the future.
It lets you surgically insert a "What-If" into a timeline and see exactly how that specific change ripples forward, leaving everything else untouched.

It turns the complex, messy data of biology and medicine into a story we can edit, understand, and use to make better decisions for our health.

Here is a detailed technical summary of the paper "Controllable Sequence Editing for Biological and Clinical Trajectories" (CLEF), published as a conference paper at ICLR 2026.

1. Problem Definition

The paper addresses the challenge of conditional generation for longitudinal sequences (time-series data) in scientific and clinical domains. While existing generative models can produce modified trajectories based on a condition (e.g., a drug intervention), they suffer from two critical limitations:

Lack of Temporal Control: Most models cannot specify when a condition takes effect. They often assume the condition alters the entire sequence or requires iterative, step-by-step generation to reach a future state, leading to compounding errors.
Lack of Scope Control: Existing methods often assume a condition affects all variables in the sequence. In reality, interventions (like surgery or medication) typically affect only a specific subset of variables (e.g., blood glucose) while leaving others (e.g., heart rate) relatively unchanged, and they must preserve historical data prior to the intervention.

The authors define two specific tasks:

Immediate Sequence Editing: Forecasting the next time step immediately after a condition is applied.
Delayed Sequence Editing: Forecasting a trajectory at an arbitrary future time step ( $t_j$ ) after a condition is scheduled, skipping intermediate steps to avoid autoregressive error accumulation.

2. Methodology: The CLEF Framework

CLEF (ControLlable sequence Editing Framework) is a novel architecture designed to learn temporal concepts that encode how and when a condition alters future sequence evolution. It allows for targeted edits to specific time steps and variables while preserving the global integrity of the sequence.

Core Architecture

CLEF consists of four main components:

Sequence Encoder ( $F$ ): Extracts temporal features from historical sequence data ( $x_{\cdot, t_0:t_i}$ ). It can utilize any sequential encoder (e.g., Transformer, xLSTM, or pretrained foundation models like MOMENT). It also generates time positional embeddings and time-delta embeddings ( $\Delta t_{i,j}$ ).
Condition Adapter ( $H$ ): Maps condition tokens (e.g., medical codes, drug names) into a latent representation ( $h_s$ ). These tokens can come from frozen pretrained embedding models.
Concept Encoder ( $E$ ): The core innovation. It learns temporal concepts ( $c$ ) by combining the sequence features ( $h_x$ $h_{x}$ ), the time-delta embedding, and the condition embedding ( $h_s$ $h_{s}$ ).
- Mathematically: $c = \text{GELU}(\text{FFN}(h_x \odot h^s_{t_j}))$ , where $h^s_{t_j}$ is a time- and condition-specific embedding.
- These concepts represent the trajectory or rate of change of variables between time $t_i$ and $t_j$ under condition $s$ .
Concept Decoder ( $G$ ): Generates the future sequence ( $\hat{x}_{\cdot, t_j}$ $\overset{x}{^}_{\cdot, t_{j}}$ ) by applying the learned concept to the last observed state ( $x_{\cdot, t_i}$ $x_{\cdot, t_{i}}$ ) via element-wise multiplication:
- $\hat{x}_{\cdot, t_j} = c \odot x_{\cdot, t_i}$ .
- This multiplicative approach is less sensitive to variables with different units of measurement and allows for direct user intervention on the concept $c$ to simulate counterfactuals.

Connection to Counterfactual Prediction

CLEF is grounded in the potential outcomes framework. Under assumptions of consistency, positivity, and sequential ignorability, CLEF learns balanced representations that approximate the conditional mean function of counterfactual outcomes. The architecture isolates the causal effect of the treatment from spurious factors, enabling accurate estimation of counterfactual trajectories without explicit density estimation.

3. Key Contributions

Novel Framework (CLEF): A flexible model for instance-wise conditional generation that supports both immediate and delayed sequence editing with precise control over timing and scope.
Temporal Concepts: The introduction of learnable temporal concepts that capture trajectory patterns, enabling single-step generation to arbitrary future time points and preserving global sequence integrity.
Zero-Shot Counterfactual Generation: CLEF demonstrates the ability to generate counterfactual trajectories for unseen conditions (zero-shot) by leveraging the semantic space of condition embeddings.
New Benchmarks & Datasets: The authors introduce four new benchmark datasets for cellular reprogramming (WOT, WOT-CF) and patient immune dynamics (eICU, MIMIC-IV), alongside evaluations on established sales and tumor growth datasets.
Clinical Case Study: A demonstration of CLEF's utility in simulating Type 1 Diabetes trajectories, where direct intervention on learned concepts generates realistic "healthier" or "sicker" patient outcomes.

4. Experimental Results

CLEF was evaluated on 8 datasets (cellular, clinical, and sales) against 9 state-of-the-art baselines (including Transformers, xLSTM, MOMENT, CRN, and Causal Transformer).

Immediate Sequence Editing: CLEF improved immediate editing accuracy by an average of 16.28% (MAE) compared to non-CLEF counterparts. It successfully preserved unedited variables better than baselines.
Delayed Sequence Editing: CLEF outperformed non-CLEF models by 26.73% (MAE) on average. Crucially, it achieved this via single-step generation, avoiding the compounding errors seen in autoregressive baselines that require filling intermediate time steps.
Counterfactual Prediction: Under high time-varying confounding ( $\gamma \ge 3$ ), CLEF-enhanced models (CLEF-CRN, CLEF-CT) significantly outperformed baselines. Notably, CLEF achieved strong performance without explicit balancing loss functions, suggesting the learned temporal concepts inherently capture balanced representations.
Zero-Shot Generation: In the WOT-CF dataset (cellular trajectories), CLEF outperformed baselines in generating counterfactual trajectories for conditions not seen during training.
Generalization: CLEF models showed superior generalization to new patient distributions (tested via SPECTRA splits) compared to standard baselines, maintaining stability even as train/test similarity decreased.
Case Study (Type 1 Diabetes): By intervening on temporal concepts to halve glucose levels, CLEF generated trajectories that were significantly more similar to healthy patients ( $R^2$ improvement) than the original T1D trajectories. Conversely, doubling glucose levels shifted trajectories toward "sicker" states. The model also captured indirect physiological effects (e.g., reduced glucose leading to reduced White Blood Cell count).

5. Significance and Impact

Virtual Cells and Patients: CLEF bridges the gap between generative AI and scientific simulation, enabling the creation of "virtual cells" and "virtual patients" for large-scale in silico experimentation. This allows researchers to test hypotheses about drug timing and dosage without physical trials.
Interpretability and Control: Unlike black-box generative models, CLEF's use of temporal concepts allows for direct, interpretable intervention. Clinicians can simulate "what-if" scenarios by editing specific concept values (e.g., "what if we reduce glucose by 50%?").
Clinical Decision Support: The ability to generate realistic counterfactual trajectories shifted toward healthier outcomes offers a powerful tool for personalized medicine, potentially helping to identify optimal intervention strategies for complex diseases like Type 1 Diabetes.
Methodological Advancement: The paper challenges the necessity of complex balancing losses in counterfactual prediction, showing that architectural design (temporal concepts) can inherently learn robust, treatment-invariant representations.

In summary, CLEF represents a significant step forward in controllable time-series generation, offering precise, interpretable, and accurate tools for modeling complex biological and clinical trajectories under specific interventions.