💻 computer science

TusoAI: Agentic Optimization for Scientific Methods

TusoAI is an agentic AI system that autonomously develops and optimizes domain-specific computational methods by integrating unstructured scientific knowledge into a tree representation, demonstrating superior performance over existing tools and uncovering novel biological insights in tasks ranging from single-cell RNA-seq denoising to genetic disease association discovery.

Original authors: Alistair Turcan, Kexin Huang, Lei Li, Martin Jinye Zhang

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alistair Turcan, Kexin Huang, Lei Li, Martin Jinye Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a scientist trying to solve a complex puzzle, like figuring out how a specific gene causes a disease. Usually, you have to build your own custom tool to solve it. This is like trying to build a specialized wrench from scratch every time you need to tighten a bolt. You have to read hundreds of manuals (scientific papers), guess how the bolt works, try building the wrench, realize it's too heavy, try again, and repeat this for years. It's slow, expensive, and frustrating.

TusoAI is a new "robot scientist" designed to automate this wrench-building process. Instead of just using existing tools, TusoAI learns how to invent and improve the tools themselves.

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

1. The "Knowledge Tree" (The Library of Wisdom)

Before TusoAI starts building, it doesn't just guess. It goes to a massive library of scientific papers and reads the most important ones.

The Metaphor: Imagine a chef who wants to invent a new recipe. Instead of just guessing, they read 100 cookbooks, summarize the best techniques for chopping, seasoning, and heating, and organize these tips into a structured "Tree of Wisdom."
How TusoAI does it: It creates a "Knowledge Tree." The branches are categories (like "how to handle noisy data" or "how to model gene interactions"), and the leaves are specific instructions. This ensures the robot doesn't just hallucinate; it builds on real, proven scientific knowledge.

2. The "Chef's Kitchen" (The Optimization Loop)

Once TusoAI has its tree of knowledge, it starts cooking. It doesn't just make one dish; it runs a high-speed competition.

The Metaphor: Imagine a kitchen where TusoAI is the head chef. It asks 5 different junior chefs (AI agents) to try a recipe.
- Instruction-Based Cooking (80% of the time): The head chef picks a specific tip from the "Knowledge Tree" (e.g., "Try adding a pinch of salt") and tells the junior chefs to apply it. If the dish tastes better, the chef remembers that tip is useful and uses it more often next time.
- Diagnostic Cooking (20% of the time): Sometimes, the dish tastes bad, but the chef doesn't know why. So, they ask the chefs to run a "diagnostic" (like checking the oven temperature or the freshness of the ingredients). Based on the data logs, the chef figures out what went wrong and fixes it.
The Result: Every time a junior chef makes a better version of the tool, it gets added to the pool. The worst ones are discarded, and the best ones get tweaked again. This happens over and over for 8 hours.

3. The "Evolutionary Race" (Finding the Best Solution)

TusoAI is designed to avoid getting stuck in a rut.

The Metaphor: If you only ask chefs to add a tiny bit more salt, you might never discover that adding lemon juice is the real secret. TusoAI forces its "chefs" to try very different approaches. It keeps a diverse group of solutions so it doesn't accidentally throw away a weird idea that turns out to be brilliant.
The "Bayesian" Twist: As the robot learns which "branches" of the Knowledge Tree lead to better results, it starts trusting those branches more. It's like a smart investor who puts more money into the stocks that are currently performing well, while still keeping a little bit of money in risky, new ideas just in case.

What Did It Actually Achieve?

The paper tested TusoAI on 11 different scientific challenges, mostly in genetics and biology.

Beating the Experts: In 6 different tasks involving single-cell data (like analyzing thousands of individual cells), TusoAI built tools that performed better than the best human experts and other AI systems.
New Discoveries: When applied to two specific genetics problems, TusoAI didn't just tweak existing methods; it found better ways to do the math.
- Case 1: It improved a method for finding which cells are linked to diseases, discovering 21% more connections than the original human-made method.
- Case 2: It improved a method for linking genetic variants to genes, finding new connections that previous methods missed, including a link between a specific gene and glucose levels that makes perfect biological sense.

The Bottom Line

TusoAI is not a tool that just uses science; it is a tool that does the engineering of science. It mimics the human cycle of reading, hypothesizing, building, testing, and fixing, but it does it at a speed and scale no human team could match. It takes a problem description, reads the relevant literature, and autonomously writes, tests, and refines the software code needed to solve that problem, often finding solutions that are better than what humans have designed so far.

Technical Summary: TUSOAI: Agentic Optimization for Scientific Methods

Problem Statement

Scientific discovery is frequently bottlenecked by the slow, manual development of computational tools required to analyze complex experimental data. While Large Language Models (LLMs) have demonstrated capabilities in synthesizing literature and generating code, existing LLM-based systems generally fall into two categories: (1) those that perform scientific analyses using existing tools and pipelines, or (2) those that develop machine learning methods for general applications without effectively integrating unstructured, domain-specific scientific knowledge. Consequently, there is a lack of systems capable of autonomously developing and optimizing new computational methods for specific scientific tasks by iteratively reviewing literature, testing assumptions against empirical data, and implementing insights into efficient software.

Methodology: TUSOAI

TUSOAI is an agentic AI system designed to autonomously develop and optimize computational methods given a scientific task description and an evaluation function. It mimics a scientist's cycle of method development through three core steps, operating on a single function within a larger codebase to allow for flexible integration with existing scientific scaffolding.

1. Knowledge Gathering and Structuring

TUSOAI begins by retrieving and summarizing up to 10 key scientific papers relevant to the task. It then constructs a Knowledge Tree, a two-level structured representation of domain knowledge:

Categories: High-level optimization strategies (e.g., "noise modeling," "model architectures") are drafted and refined using the paper summaries. Each category is assigned a probability weight ( $\pi_c$ ) representing its estimated utility.
Instructions: Specific, actionable instructions are generated within each category (e.g., "model dropout rates using Poisson noise").
Diagnostic Category: A predefined category ( $I_{diag}$ ) guides the agent to log diagnostic information (e.g., distribution checks, assumption validation) to aid in model diagnosis.

2. Initialization

The system initializes a pool of candidate solutions ( $S$ ) by drafting and refining solution descriptions based on the task and literature. These descriptions are then implemented and debugged to form the initial solution pool.

3. Iterative Optimization

TUSOAI enters an optimization loop (default time budget: 8 hours) where it selects diverse top-performing solutions and improves them via two modes:

Instruction-Based Optimization (80% probability): The agent samples an instruction category based on the Bayesian-updated probabilities ( $\pi_c$ ). It selects a specific instruction and optimizes the solution using the instruction and recent feedback. If the new solution outperforms the previous one, the category's probability is increased (Bayesian update), reinforcing the utility of that strategy.
Diagnostic-Based Optimization (20% probability): The agent selects a diagnostic instruction, runs the model to collect logs (e.g., training curves, assumption violations), and uses this information to diagnose and improve the model.

This process integrates hierarchical planning with Bayesian updates to balance solution quality and diversity, and fine-grained generation that couples optimization with diagnostic feedback.

Key Contributions

TUSOAI Framework: The development of an AI agent specifically tailored for scientific method discovery, integrating structured domain knowledge via a knowledge tree.
Novel Optimization Mechanisms: A framework featuring (i) knowledge trees for structured domain representation, (ii) hierarchical planning with Bayesian updates to manage the trade-off between solution quality and diversity, and (iii) fine-grained generation that integrates model optimization with diagnostic feedback.
Comprehensive Benchmarking: Evaluation on 11 scientific tasks (6 single-cell analysis tasks and 5 scientific deep learning tasks), demonstrating consistent outperformance over state-of-the-art MLE agents (e.g., AIDE), scientific AI agents (e.g., Biomni), and expert-designed methods.
Scientific Impact: Application to two key open problems in genetics, where TUSOAI improved existing computational methods and uncovered new biological insights missed by previous approaches.

Experimental Results

The authors evaluated TUSOAI against baselines including AIDE, Biomni, ChatGPT-Agent, and top-performing expert methods across diverse domains.

Single-Cell Analysis: TUSOAI achieved an average rank of 1.2 across 6 tasks (e.g., denoising, batch integration, spatial variable gene identification), significantly outperforming the second-best method (AIDE, rank 3.0). Notably, in the denoising task, TUSOAI designed a novel Non-Negative Matrix Factorization (NMF) approach that modeled dropout rates and Poisson noise, outperforming existing NMF-based methods like ALRA.
Scientific Deep Learning: On 5 tasks (e.g., Spherical vision, ECG diagnostics, genetic prediction), TUSOAI achieved an average rank of 2.8, surpassing expert-designed models and NAS methods.
Code Diversity: TUSOAI generated code with substantially higher diversity than AIDE, avoiding local optima by exploring a wider range of techniques (e.g., various dimensionality reduction and scaling methods) rather than incremental variations of a single approach.
Efficiency: All methods constructed by TUSOAI were computationally efficient, with single-cell tasks running in <3 minutes and deep learning tasks in <8 minutes.

Case Studies in Genetics

The paper highlights two specific applications where TUSOAI revealed new biology:

Detecting Disease-Critical Cell Populations: Optimizing the scoring function of scDRS (a method linking GWAS to scRNA-seq). The TUSOAI-optimized version achieved >40% higher power in causal simulations and identified 21% more true cell type–disease associations in real data. It discovered 26 disease-associated T cell subpopulations, including regulatory T cells linked to primary biliary cirrhosis, which were missed by the original method.
Linking Genetic Variants to Genes: Optimizing pgBoost for linking variants to target genes using single-cell multiome data. The optimized model improved enrichment of gold-standard links by 13.8% (eQTLs) and 7.2% (ABC links), particularly for longer variant–gene distances. It uncovered 7 new variant–gene links, including a link between a glucose-related variant and the GCK gene, consistent with known biological roles.

Significance and Claims

The paper claims that TUSOAI represents a promising step toward automated scientific method development. By mimicking the iterative cycle of a scientist—integrating structured domain knowledge, performing domain-specific optimization, and utilizing diagnostic feedback—TUSOAI accelerates the creation of robust computational methods. The authors assert that this capability not only improves performance on standard benchmarks but also actively contributes to scientific discovery by uncovering biological mechanisms and associations that were previously missed by existing expert-designed methods. The system demonstrates that LLMs, when coupled with structured knowledge and iterative optimization, can move beyond simple data analysis to the actual design of scientific algorithms.