With Daza Matrix, you can scientifically and efficiently design supplement formulas for any disease or health goal.

Daza Matrix can identify the most effective species and supplements from over 2 million global species and more than 100,000 types of supplements worldwide. It further builds a customized Formula Pool based on your selections from the identified species and supplements, and generates the final, complete formula from the Formula Pool.

The Daza Matrix has five core advantages:

1.  It is the world’s only AI-Powered Supplement Formula Design System fully based on PubMed medical literature.

2.  Global holistic perspective. When searching for effective ingredients targeting diseases or health goals, it covers all global species and global supplements with an extremely broad holistic vision, unlike the limited perspective of large language models.

3.  Elimination of hallucinations. By adopting the integrated solution of CNN + NLP + Domain-Specific NLP + LLM, it completely solves the three core defects of pure large language models: hallucinations, randomness and mediocre generic outputs.

4.  Revolutionary efficiency improvement. Compared with Standard Biomedical NLP and manual processing, the Daza Matrix increases the analysis efficiency of medical literature by 1,500 to 4,000 times.

5.  Easy to operate. The whole process of effective ingredient screening and formula design is intuitive, convenient and user-friendly. Both professional users and the general public can master it with ease.

Some readers may not know what PubMed is, so we will give a brief introduction first. PubMed is the world’s largest and most authoritative biomedical literature database. It is the most frequently visited platform by global doctors, nutritionists and health managers, serving as a direct access to the latest global medical research evidence and a professional academic knowledge base. PubMed only includes literature published in SCI journals, where SCI is the abbreviation of Science Citation Index. This means PubMed exclusively collects high-quality academic journals worldwide. There are more than 100,000 journals globally, among which only about 10,000 are SCI journals, accounting for roughly one-tenth of the total.

We will take designing a formula for gastric ulcer as an example to explain how to use the Daza Matrix.

Designing formulas with the Daza Matrix only requires five steps.

Design Gastric Ulcer Formulas with Daza Matrix

Step 1: Enter keywords of the disease or health goal.

Here we enter "gastric ulcer"

Step 2: Click to extract and analyze data for each batch. Semantic-level Word Frequency Value sorting will automatically identify species and supplements closely related to your target disease or health goal.

All relevant literature retrieved by keywords — whether tens of thousands, over one hundred thousand, or even hundreds of thousands of articles — can be fully obtained. The Daza Matrix divides all literature to be acquired into batches of 2,000 articles each. The operation is simple: click Extract for each batch in sequence.

After finishing the extraction of all batches, click Species for each batch in turn. This will extract all species closely related to gastric ulcer identified from the 2,000 literature entries in each batch, and sort them by Semantic-level Word Frequency Value.

The Semantic-level Word Frequency Value of a species is a weighted result of multiple key factors including species word frequency, species proportion, Journal Impact Factor and literature type. The ranking based on this semantic-level word frequency value truly reflects the tightness of the correlation between each species and gastric ulcer.

After completing species identification for all batches, click Supplements for each batch sequentially. This will extract all supplements closely associated with gastric ulcer identified from the 2,000 literature entries in each batch, and sort them by Semantic-level Word Frequency Value.

This correlation includes both positive correlations and negative correlations.

Step 3: Click Correlation Coloring to visually distinguish species and supplements with positive and negative correlations.

You can see every species is marked with green or reddish-brown in different shades.

Positive correlations (indicating effectiveness) are presented in three shades of green, while negative correlations (indicating adverse effects) are shown in three shades of reddish-brown. The deeper the color, the stronger the correlation intensity.

The combination of Semantic-level Word Frequency Value sorting and Correlation Coloring makes the most effective species for a specific disease or health goal stand out clearly at a glance.

To see the specific effects of a given species or supplement on a disease, simply click on the species or supplement, then click Keyword Literature in the pop-up menu. You will then be able to view PubMed literature related to this species or supplement and your target disease or health goal.

Step 4: Add your preferred species and supplements to the Formula Pool. Then further customize the pool by specifying options such as Mandatory Ingredients and Number of Ingredients.

To add your preferred species to the customized Formula Pool, click on the species and select Add to Formula Pool in the pop-up dialog box.

After adding all candidate species you intend to use to the Formula Pool, add your preferred supplements to the pool in the same way.

Your customized Formula Pool for gastric ulcer will display all added candidate species and supplements, with species marked in green and supplements marked in blue.

As you can see, the candidate species we have added to the Gastric Ulcer Formula Pool are: Aloe vera, Hericium erinaceus (Lion’s Mane Mushroom), Glycyrrhiza glabra (Licorice), Nigella sativa (Black Cumin), Dendrobium officinale, Zanthoxylum nitidum, Centella asiatica, and Moringa oleifera. The candidate supplements are: Curcumin, Quercetin, Zinc, Honey, Melatonin, Probiotics, Propolis, Hesperidin, Catechin, Luteolin, and Apigenin.

You can customize your gastric ulcer formula via four options in the Formula Pool: Mandatory Ingredients, Number of Ingredients, Target Species and Special Instructions.

Mandatory Ingredients: Used to specify ingredients that must be included in formula design. If an ingredient is highly effective and you require it to be indispensable in the formula, click the M next to the ingredient. In this gastric ulcer formula example, we set four Mandatory Ingredients: Aloe vera, Hericium erinaceus, curcumin, hesperidin.

Number of Ingredients: If no specific setting is made, 3 to 10 ingredients will be selected by default to form a formula. To adjust the quantity range, enter two comma-separated numbers as the lower and upper limits. For instance, entering 5,8 means the formula will adopt 5 to 8 ingredients.

Target Species: All formulas are designed by default for adult humans (or minors aged 12 and above). If a formula is for specific human groups (e.g., infants, young children, pregnant women) or other species (e.g., dogs, cats, parrots), fill in the relevant information here.

Special Instructions: You can add any additional important notes here. For example, if designing a gastric ulcer formula for diabetic patients, state this requirement in Special Instructions, so ingredients like honey that are unsuitable for diabetics will be excluded from the formula.

Step 5: The large language model will generate the ultimate complete formula based on your customized Formula Pool.

After completing all customizations in the Formula Pool, click the Super Formula Design button to enter the final stage of formula design. The large language model will generate the ultimate complete formula based on your customized Formula Pool.

Four types of large language models are currently supported. We will compare the final formulas generated by Doubao and DeepSeek-R1 respectively. Firstly, the large language model will conduct a secondary check to ensure all species and supplements selected in the Formula Pool are positively correlated with the target demand. Then it will elaborate the design ideas and logical rationale of the formula, explaining how multiple mechanisms work to treat the disease or achieve the health goal. After that, it will present the detailed formula based on the established rationale: selecting one to two ingredients for each mechanism, explaining their functions in the formula, as well as the daily dosage and administration method. All generated formulas fully comply with your customization settings in the Formula Pool. The final part is a summary of the entire formula design, including key notes and precautions.

If you are not satisfied with the formula generated this time, you may click Regenerate in the top-right corner to generate a new one.

Some people might say, "Large language models are extremely powerful these days—can’t I just have one design a formula for gastric ulcers for me?" Of course you can. You can use various large language models (including ChatGPT, Gemini, DeepSeek, Doubao, and others) and try different prompts to generate dietary supplement formulas targeting gastric ulcers.

However, you will eventually find three fatal flaws in formulas generated purely by large language models:

1.  Hallucinations

This is unavoidable. It is an inherent byproduct of the large language model’s training paradigm focused on language generation rather than factual verification. The Multi-Head Self-Attention Mechanism in its Transformer architecture further amplifies such hallucinations. For example, when using Doubao solely to design a formula for gastric ulcers, you will find that the generated formula often includes an ingredient called Slippery Elm Bark Powder (Slippery Elm) and claims it is derived from the bark of Ulmus rubra. However, you cannot find any literature linking Slippery Elm or Ulmus rubra to gastric ulcers or gastric acid on PubMed.

2.  Mediocrity

Formulas generated solely by large language models are very mediocre (and sometimes completely ineffective). This is because when designing formulas, large language models tend to use the most common and marginally useful (or sometimes completely useless) supplements, rather than the most effective but potentially less common ones. The reason is that large language models generate outputs based on probability distributions. They always produce the most statistically probable answers within a limited knowledge activation scope — even if trained on full PubMed literature, they cannot conduct a true global traversal of all relevant academic data.

In simple terms, you can understand it this way: when generating answers, large language models essentially provide the answer with the highest probability within a certain range. Firstly, this range is not global; secondly, the answer with the highest probability within this range does not necessarily mean it is the correct answer within this range.

This algorithmic mechanism often makes the formulas generated by large language models seem to actively avoid the most effective species and supplements.

3.  Randomness

Different formulas are generated each time. Randomness introduced to prevent overfitting and enhance generalization ability leads to inconsistent outputs for identical requirements, making the generated formulas unreliable and confusing for users.