🧬 biology

SynthPert: Enhancing LLM Biological Reasoning via Synthetic Reasoning Traces for Cellular Perturbation Prediction

SynthPert is a novel method that enhances large language models' ability to predict cellular responses to genetic perturbations by fine-tuning them on synthetic reasoning traces generated by frontier models, achieving state-of-the-art performance and strong cross-cell-type generalization.

Original authors: Lawrence Phillips, Marc Boubnovski Martell, Aditya Misra, Josefa Lia Stoisser, Cesar A. Prada-Medina, Rory Donovan-Maiye, Kaspar Märtens

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lawrence Phillips, Marc Boubnovski Martell, Aditya Misra, Josefa Lia Stoisser, Cesar A. Prada-Medina, Rory Donovan-Maiye, Kaspar Märtens

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

The "Smart Student" Approach to Biology: How to Teach AI to Predict the Future of Cells

Imagine you are trying to teach a student how to predict the weather.

The Old Way (The "Memorization" Method):
Most scientists have been trying to teach AI by showing it millions of past weather reports: "On Tuesday, it was cloudy and then it rained." The AI gets very good at memorizing these patterns. But if a weird, unprecedented storm comes along—one it has never seen before—the AI panics. It doesn't actually understand why clouds lead to rain; it just knows that "Cloudy + Tuesday = Rain."

The New Way (The "SynthPert" Method):
The researchers behind this paper decided to stop giving the AI just the "weather reports" and instead started giving it "The Scientist’s Notebook."

Instead of just saying, "This gene went up," they used a super-smart "Teacher AI" (like a world-class professor) to write out the reasoning behind the result: "The gene went up because the perturbation blocked a specific protein, which caused a chain reaction in the cell's energy center."

By training a smaller, faster AI on these "Reasoning Traces" (the step-by-step logic), they didn't just teach it what happened; they taught it how to think like a biologist.

The Three Big "Magic Tricks" of this Paper

1. The "Distillation Paradox" (The Small Student Beats the Professor)

Usually, a small student can't beat a world-class professor. But in this experiment, the researchers took a smaller AI and trained it specifically on the best explanations written by the "Professor AI."

Because the small AI wasn't distracted by the professor's mistakes or "fluff," it became a specialist. It actually ended up outperforming the professor at predicting cellular changes. It’s like a student who studies only the most perfect, distilled notes from a genius and ends up being more efficient at the exam than the genius themselves!

2. The "Universal Translator" (Generalization)

One of the biggest problems in biology is that a cell in your liver behaves differently than a cell in your eye. Most AIs trained on liver cells are useless when you ask them about eye cells.

However, because SynthPert learned the logic of biology (the "why") rather than just the data of the liver (the "what"), it was able to jump from liver cells to eye cells and still be incredibly accurate. It learned the "laws of physics" for cells, which work everywhere.

3. "Quality Over Quantity" (The 2% Rule)

Imagine if you could learn an entire semester of biology by reading only 2% of the textbook—but those pages were the most perfectly written, logical explanations ever created.

That is what these researchers did. They used a tiny fraction of the data, but because it was "high-quality reasoning data," the AI learned faster and better than if it had read the whole messy textbook.

Why does this matter for the real world?

Right now, discovering new medicines is like trying to find a needle in a haystack by hand. You have to test thousands of chemicals on real cells in a lab, which takes years and billions of dollars.

SynthPert is building a "Virtual Laboratory." If we can teach an AI to accurately "reason" through how a cell will react to a new drug, we can run millions of experiments in a computer in seconds. We can predict the side effects and the benefits before a single drop of medicine is ever made in a real lab.

In short: We aren't just teaching AI to recognize patterns; we are teaching it to understand the "language" of life.

Technical Summary: SynthPert – Enhancing Biological Reasoning in LLMs via Synthetic Reasoning Traces

Problem Statement

Predicting how cells respond to genetic perturbations (e.g., whether a gene's expression goes up, down, or stays the same) is a cornerstone of systems biology and drug discovery. Current approaches face two primary hurdles:

Generalization Gap: Classical deep learning models (like GEARS or scGPT) often struggle to generalize to unseen biological contexts or new cell types.
Reasoning Limitations: While Large Language Models (LLMs) possess vast biological knowledge, they are typically used as "black-box" predictors or simple embedding generators. Existing LLM frameworks for perturbation biology often simplify the task into binary classifications (e.g., "is it differentially expressed?" followed by "is it up or down?"), which does not reflect real-world experimental workflows where the outcome is unknown.

Methodology: SynthPert

The authors propose SynthPert, a method that shifts the focus from training on raw experimental data to training on synthetic reasoning traces (Chain-of-Thought).

Task Formulation: Instead of decomposing the problem, SynthPert performs a direct three-class prediction: (cell type, perturbation, gene) $\rightarrow$ {up, down, not differentially expressed}.
Synthetic Data Generation (Distillation):
- The researchers use a frontier model (OpenAI o4-mini) to generate mechanistic explanations for observed experimental outcomes.
- They employ a "Critic" model to grade these explanations on a 5-point scale. Only "excellent" explanations are retained for training.
- This process creates a high-quality dataset of "reasoning traces" that link biological inputs to outcomes through causal logic.
Supervised Fine-Tuning (SFT):
- A smaller, open-source model (DeepSeek-R1 8B) is fine-tuned using LoRA (Low-Rank Adaptation) on these high-quality synthetic traces.
- This allows the smaller model to learn the logic of biological regulation rather than just memorizing statistical correlations in the data.

Key Contributions

Synthetic Reasoning Distillation: A novel paradigm where LLMs are enhanced by fine-tuning on generated reasoning chains rather than raw data tuples.
Direct Three-Class Prediction: A more practical and biologically relevant task formulation compared to existing decomposed binary approaches.
Data Efficiency: Demonstrating that high-quality reasoning traces allow for superior performance even when using only a tiny fraction (2%) of the available training data.

Results

The model was evaluated on the PerturbQA benchmark across multiple cell lines (K562, RPE1, HepG2, Jurkat).

State-of-the-Art Performance: SynthPert achieved 78% AUROC on differential expression tasks, significantly outperforming existing methods like SUMMER, GEARS, and scGPT.
The "Distillation Paradox": Remarkably, SynthPert (an 8B parameter model) surpassed the performance of the frontier teacher model (o4-mini) in direct three-class prediction (89% accuracy vs. 52% accuracy).
Cross-Cell-Type Generalization: In a rigorous "zero-shot" test where the model was evaluated on an entirely unseen cell type (RPE1), SynthPert maintained 87% accuracy. This is a massive improvement over the 40% accuracy achieved by models trained directly on experimental data tuples.
Precision-Recall Trade-off: SynthPert proved to be more "calibrated" than the teacher model. While the teacher model tended to over-predict changes (high recall, low precision), SynthPert learned to be more selective, providing more reliable (high precision) predictions.

Significance

This work challenges the conventional wisdom that biological AI requires massive, diverse experimental datasets to achieve generalization. Instead, it suggests that reasoning structure is more important than factual volume.

By distilling the logic of biology from frontier models into smaller, specialized models, researchers can create highly efficient, interpretable, and generalizable tools for in silico biological experimentation. This provides a blueprint for applying LLM reasoning to other complex scientific domains where experimental data is sparse but foundational knowledge is abundant.