HarmonyCell: Automating Single-Cell Perturbation Modeling

HarmonyCell integrates two core synergistic components:

• LLM-driven Semantic Unifier: This module prompts a frozen LLM with raw field descriptors to infer a canonical JSON mapping specification. It captures both direct field aliasing and dynamic logic expressions, enabling zero-shot adaptation to uncurated datasets without manual intervention, transforming disparate raw datasets into a strictly unified interface.
• Adaptive MCTS Engine with Hierarchical Action Space: To bridge the gap between known biology and novel perturbations, HarmonyCell employs an adaptive Monte Carlo Tree Search engine. It frames optimal statistical inductive bias as a structured search problem, navigating a three-level hierarchy (Modeling Paradigm, Architectural Backbone, Optimization Refinement) to dynamically synthesize architectures tailored to biological distribution shifts.

The system is meta-initialized via historical priors, warm-starting the search for similar tasks and ensuring ab initio exploration for novel contexts, optimizing for both prediction accuracy and computational efficiency.

HarmonyCell was rigorously evaluated across single-dataset and multi-dataset settings, encompassing diverse perturbation tasks and both semantic and distribution shifts. Key findings include:

• Semantic Heterogeneity Resilience: HarmonyCell achieved a 95% valid execution rate with 0% preprocessing errors on heterogeneous inputs, significantly outperforming general coding agents (0% success) by autonomously resolving semantic conflicts.
• Synergistic Data Scaling: Automated semantic unification enabled predictive gains, with models trained on HarmonyCell-harmonized datasets showing consistent performance improvements and significant positive transfer across domains.
• Statistical Generalization Efficiency: HarmonyCell consistently matched or exceeded expert-designed baselines in out-of-distribution tasks, effectively adapting to continuous covariate shifts (drug perturbation) and discrete combinatorial shifts (gene perturbation) by dynamically synthesizing optimal architectures.

The hierarchical MCTS search space ensures superior convergence speed and accuracy, avoiding local optima that trap simpler search methods.

Despite its effectiveness, HarmonyCell faces limitations inherent to search-based systems:

• The MCTS engine entails higher computational overhead compared to static baselines.
• The agent's 'creativity' is bounded by a pre-defined library of architectural primitives, limiting its ability for truly novel model design.
• The current framework focuses on unimodal data, leaving multi-modal integration and open-ended mathematical discovery as key directions for future research.