Controllable Sequence Editing for Biological and Clinical Trajectories.

Area of Science:

• Computational Biology
• Machine Learning
• Data Science

Background:

• Sequence models generate counterfactuals for "what if" reasoning but lack fine-grained control over edits.
• Existing methods assume global interventions or focus on univariate sequences, limiting applications requiring localized modifications.
• Precise, localized control is needed for interventions affecting specific time steps and subsets of variables.

Purpose of the Study:

• Introduce Clef, a controllable sequence editing model for counterfactual reasoning.
• Enable precise control over immediate and delayed effects of interventions in sequences.
• Facilitate selective editing of time steps while preserving unaffected sequence portions.

Main Methods:

• Clef learns temporal concepts to encode intervention timing and influence.
• The model selectively edits relevant time steps based on learned temporal concepts.
• Evaluation on cellular and patient trajectory datasets, including gene regulation and medical interventions.

Main Results:

• Clef improves immediate sequence editing by up to 36.01% in Mean Absolute Error (MAE) compared to baselines.
• Clef enables one-step generation of counterfactual sequences at any future time step, outperforming baselines by up to 65.71% in MAE.
• A case study demonstrated Clef's ability to identify clinical interventions for healthier patient trajectories in type 1 diabetes mellitus.

Conclusions:

• Clef offers a novel approach to controllable sequence editing for counterfactual reasoning.
• The model's temporal concept learning enables precise, localized interventions.
• Clef shows potential for advancing biological and clinical data analysis and personalized medicine.