Breaking Down CORE: A New Approach to Gene Expression Prediction

Understanding cellular mechanisms through perturbation experiments is a cornerstone of biological research. However, the high cost and scarcity of these experiments push scientists to predict gene expression responses for conditions they haven't directly observed. Enter the field of large language models (LLMs) acting as 'virtual cell' simulators. These models use mechanistic reasoning to infer differential gene expression, suggesting a shift from purely data-driven methods to more knowledge-driven paradigms.

The Shortcomings of Current Models

Despite the promise, there's a cavern between plausibility and prediction accuracy. Current LLM-based methods often produce explanations that seem biologically plausible but fail to accurately predict perturbation-specific effects. Often, these methods overestimate differential expression, revealing that they lean more on intrinsic gene response tendencies than actual perturbation reasoning. In some cases, they even underperform compared to a simple gene-frequency baseline.

The underlying issue? These methods evaluate perturbation-gene pairs in isolation and ignore how different perturbations affect the same gene. It's a flaw that limits their effectiveness in truly understanding gene behavior under varying conditions.

CORE: Shifting the Paradigm

Enter CORE, or Contrastive Organization of Relational Evidence. This innovative approach reframes prediction as a comparison task. Instead of evaluating evidence in isolation, CORE organizes it into positive and negative outcomes from related perturbations, utilizing a biomedical knowledge graph for evidence retrieval. The result? Improved calibration and a marked boost in perturbation-specific predictions across various settings.

For instance, when applied to drug-perturbation data, CORE-Reasoning enhances Qwen3.5-9B's aggregate metrics by up to 28.6%. On generic perturbation data, CORE-Voting impressively lifts the macro-per-gene AUROC from a chance level to an average of 0.703 across four cell lines. It's a testament to the power of contrastive evidence organization in achieving reliable LLM-based perturbation reasoning.

Why This Matters

Why should this development grab your attention? The AI-AI Venn diagram is getting thicker, and CORE's approach is a prime example of how AI can elevate the field of genomics. By enhancing our ability to predict gene expression responses, we can direct resources more efficiently towards promising experiments, potentially accelerating breakthroughs in medical research.

But there's a broader implication here. If CORE's method marks a shift in how we approach complex biological predictions, could this be the beginning of a new era where AI not only supports but enhances scientific hypotheses? Are we witnessing the dawn of agentic AI in biomedical research?

In the collision of AI and biology, CORE stands out not just as a tool but as a paradigm shift. It's not just a convergence. It's a bold step forward.