---
title: README
emoji: ⚡
colorFrom: red
colorTo: purple
sdk: static
pinned: false
---

# MetaboLLM

[![arXiv](https://img.shields.io/badge/arXiv-2608.06253-b31b1b.svg)](https://arxiv.org/abs/2608.06253)
[![GitHub](https://img.shields.io/badge/GitHub-Code-181717.svg)](https://github.com/dohyunku9/MetaboLLM)

**MetaboLLM** is a family of metabolomics-specialized large language models designed to integrate biochemical knowledge across heterogeneous resources and support metabolite-, pathway-, reaction-, and enzyme-centered reasoning and description generation.

MetaboLLM was developed through continual pretraining, supervised fine-tuning, and structured retrieval using harmonized biochemical knowledge from KEGG, HMDB, PubChem, and SMPDB.

<p align="center">
  <img src="https://huggingface.co/spaces/MetaboLLM/README/resolve/main/assets/Figure1.png" alt="Overview of the MetaboLLM framework" width="950">
</p>

## Highlights

- Integrated knowledge covering **237,243 metabolites**, **2,359 pathways**, **12,323 reactions**, and **5,993 enzymes**
- Released **four MetaboLLM model variants** across Qwen, Gemma, and Llama backbones
- Constructed a metabolomics benchmark containing **17 tasks**, **6,000 training examples**, and **4,200 test examples**
- Evaluated factual knowledge, class identification, biochemical relationships, and description generation
- Compared MetaboLLM against corresponding base models and five publicly available medical language models
- Evaluated transfer on the independently developed MetaBench benchmark

## Resources

| Resource | Description |
|---|---|
| [MetaboLLM-Qwen3-4B](https://huggingface.co/MetaboLLM/MetaboLLM-Qwen3-4B) | Primary MetaboLLM model based on Qwen3-4B |
| [MetaboLLM-Qwen3-8B](https://huggingface.co/MetaboLLM/MetaboLLM-Qwen3-8B) | MetaboLLM model based on Qwen3-8B |
| [MetaboLLM-Gemma-3-4B](https://huggingface.co/MetaboLLM/MetaboLLM-Gemma-3-4B) | MetaboLLM model based on Gemma-3-4B |
| [MetaboLLM-Llama-3.2-3B](https://huggingface.co/MetaboLLM/MetaboLLM-Llama-3.2-3B) | MetaboLLM model based on Llama-3.2-3B |
| [MetaboLLM-Benchmark](https://huggingface.co/datasets/MetaboLLM/MetaboLLM-Benchmark) | Training and evaluation benchmark across 17 metabolomics tasks |

## Model Family

| Model | Backbone | Release Format |
|---|---|---|
| MetaboLLM-Qwen3-4B | Qwen3-4B | PEFT LoRA adapter |
| MetaboLLM-Qwen3-8B | Qwen3-8B | PEFT LoRA adapter |
| MetaboLLM-Gemma-3-4B | Gemma-3-4B | PEFT LoRA adapter |
| MetaboLLM-Llama-3.2-3B | Llama-3.2-3B | PEFT LoRA adapter |

MetaboLLM-Qwen3-4B served as the primary model in the associated study. All released repositories contain adapter weights and require the corresponding base model.

## Integrated Biochemical Knowledge

MetaboLLM was developed from a unified resource integrating complementary information from:

- **KEGG** for metabolites, reactions, enzymes, and pathways
- **HMDB** for human metabolites, biological roles, chemical taxonomy, and compound descriptions
- **PubChem** for chemical structures, identifiers, molecular properties, and compound descriptions
- **SMPDB** for curated pathways and physiological descriptions

The harmonized resource contains:

| Entity type | Count |
|---|---:|
| Metabolites | 237,243 |
| Pathways | 2,359 |
| Reactions | 12,323 |
| Enzymes | 5,993 |

## MetaboLLM Benchmark

The benchmark contains 17 tasks organized into four complementary categories.

| Category | Tasks | Train | Test |
|---|---:|---:|---:|
| Knowledge recall | 5 | 1,000 | 1,000 |
| Class identification | 4 | 1,000 | 1,000 |
| Relation identification | 3 | 1,000 | 1,000 |
| Description generation | 5 | 3,000 | 1,200 |
| **Total** | **17** | **6,000** | **4,200** |

The benchmark evaluates:

- molecular identity and formula knowledge
- metabolite and reaction class recognition
- metabolite–pathway relationships
- metabolite–reaction relationships
- reaction–enzyme relationships
- metabolite, pathway, and enzyme description generation
- structure-rich and structure-poor metabolite description generation

## Benchmark Results

### Biochemical Knowledge and Structured Relationships

Mean accuracy across 12 multiple-choice and short-answer tasks.

| Category | Model | Mean Accuracy |
|---|---|---:|
| Medical LLM | MedGemma-1.5-4B | 49.0 |
| Medical LLM | FineMedLM-O1 | 57.9 |
| Medical LLM | II-Medical-8B-1706 | 63.0 |
| Medical LLM | Qwen2.5-Aloe-Beta-7B | 63.9 |
| Medical LLM | Meditron3-Qwen2.5-7B | 65.7 |
| Base model | Llama-3.2-3B | 51.5 |
| Base model | Gemma-3-4B | 52.2 |
| Base model | Qwen3-4B | 66.5 |
| Base model | Qwen3-8B | 66.8 |
| MetaboLLM | MetaboLLM-Llama-3.2-3B | 70.0 |
| MetaboLLM | MetaboLLM-Gemma-3-4B | 75.3 |
| MetaboLLM | MetaboLLM-Qwen3-8B | 75.9 |
| MetaboLLM | **MetaboLLM-Qwen3-4B** | **79.1** |

All four MetaboLLM variants outperformed their corresponding unadapted backbones and all evaluated medical language models. MetaboLLM-Qwen3-4B achieved the highest mean accuracy at 79.1%, compared with 65.7% for the strongest evaluated medical model and 66.5% for its corresponding base model.

### Biochemical Description Generation

BERTScore-F1 on a 0–100 scale.

| Category | Model | Metabolite Description | Pathway Description | Enzyme Description | Structure-Rich Metabolite Description | Structure-Poor Metabolite Description |
|---|---|---:|---:|---:|---:|---:|
| Medical LLM | MedGemma-1.5-4B | 81.05 | 80.97 | 79.33 | 84.89 | 84.73 |
| Medical LLM | FineMedLM-O1 | 83.05 | 82.87 | 81.22 | 84.10 | 83.96 |
| Medical LLM | II-Medical-8B-1706 | 82.71 | 82.94 | 81.60 | 86.37 | 85.84 |
| Medical LLM | Qwen2.5-Aloe-Beta-7B | 83.34 | 83.39 | 81.68 | 85.47 | 84.52 |
| Medical LLM | Meditron3-Qwen2.5-7B | 83.69 | 83.14 | 81.99 | 84.47 | 83.76 |
| Base model | Llama-3.2-3B | 82.81 | 82.56 | 81.40 | 84.56 | 84.50 |
| Base model | Gemma-3-4B | 81.74 | 81.54 | 80.37 | 85.48 | 85.07 |
| Base model | Qwen3-4B | 82.39 | 81.85 | 80.68 | 86.48 | 85.80 |
| Base model | Qwen3-8B | 82.30 | 82.28 | 80.79 | 86.20 | 85.61 |
| MetaboLLM | MetaboLLM-Llama-3.2-3B | 90.93 | 86.74 | 83.78 | 87.13 | 87.65 |
| MetaboLLM | MetaboLLM-Gemma-3-4B | 87.39 | 83.83 | 80.69 | 87.03 | 87.74 |
| MetaboLLM | MetaboLLM-Qwen3-8B | 90.94 | 87.26 | 83.70 | 87.66 | 88.02 |
| MetaboLLM | **MetaboLLM-Qwen3-4B** | **91.71** | **88.04** | **84.12** | **87.79** | **88.06** |

MetaboLLM-Qwen3-4B achieved the highest BERTScore-F1 across all five description-generation tasks.

## External Benchmark Transfer

MetaboLLM was also evaluated on MetaBench, an independently developed public metabolomics benchmark.

- **MetaboLLM-Qwen3-8B** achieved the highest Knowledge MCQA accuracy at **56.42%**
- **MetaboLLM-Qwen3-4B** achieved the highest pathway-description scores, including **85.19 BERTScore-F1**, **25.83 ROUGE-L-F1**, and **17.61 BLEU-2**

These results demonstrate transfer beyond the internally constructed MetaboLLM benchmark.

## Quick Start

Install the required packages:

```bash
pip install -U transformers peft accelerate torch
```

Example using MetaboLLM-Qwen3-4B:

```python
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "MetaboLLM/MetaboLLM-Qwen3-4B"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype="auto",
    device_map="auto",
)

model = PeftModel.from_pretrained(
    base_model,
    adapter_id,
)

messages = [
    {
        "role": "user",
        "content": "What is the biochemical role of pyruvate?"
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=256,
        do_sample=False,
    )

response = tokenizer.decode(
    outputs[0][inputs.shape[-1]:],
    skip_special_tokens=True,
)

print(response)
```

Please check the corresponding model card for the exact base-model identifier and model-specific loading instructions.

## Loading the Benchmark

```python
from datasets import load_dataset

dataset = load_dataset(
    "MetaboLLM/MetaboLLM-Benchmark",
    "Knowledge_recall",
)

print(dataset["test"][0])
```

Available configurations:

- `Knowledge_recall`
- `Class_identification`
- `Relation_identification`
- `Description_generation`

## Intended Uses

MetaboLLM is intended for research involving:

- metabolomics-specific question answering
- biochemical knowledge recall
- metabolite and reaction class identification
- metabolite–pathway and reaction–enzyme relationship identification
- metabolite, pathway, and enzyme description generation
- evaluation of metabolomics-specialized language models

## Limitations

- MetaboLLM may generate incorrect or unsupported biochemical statements.
- Performance outside metabolomics and biochemical knowledge tasks has not been comprehensively evaluated.
- Outputs should be independently verified and should not be used for clinical decision-making.

## Licenses

Each MetaboLLM repository is distributed according to its model card and license notice.

- Use of each adapter remains subject to the license and terms of its corresponding base model.
- Benchmark components created by the MetaboLLM authors are distributed under the terms described in the benchmark repository.
- Third-party and source-derived biochemical content remains subject to the licenses and terms of the original providers.

Please review the license files in the relevant model and dataset repositories before use.

## Resource Maintainers

- Dohyun Ku
- Min Gu Kwak