insilicomedicine/precious3-gpt-multi-modal

_{What's new?
🐍📓 Geroprotector identification in TCM Jupyter Notebook (13/8/2024)
💻 Mass screenings with P3GPT, Python library (29/11/2024)}

Precious3GPT

A multimodal multi-omics multi-species multi-tissue language model.

📃 Pre-print • 👾 Discord bot • 🧬 Validation digest
𝕏 @precious_gpt

• Developer: Insilico Medicine
• License: cc-by-nc-4.0
• Model size: 175.1 billion parameters (Core Model 89.4 million, Text modality 175 billion, Knowledge Graph modality 8.2 million)
• Domain: Biomedical
• Base architecture: MPT

Model summary

• Precious3GPT (P3GPT) is a unique language model that has been trained on 1.2MM omics data points, knowledge graphs, and biomedical texts (PubMed) to be used in drug discovery and aging research;

• P3GPT simulates biological processes on an omics level to return the transcriptomic, epigenetic, or proteomic signatures of a wide variety of perturbators;

• Various modes of execution allow users to replicate the workflows of chemical screenings, case-control observational studies, and other popular research settings;

• The context of P3GPT-simulated experiments can be defined with >60k biomedical entities, including 3 species, 569 tissues and cell lines, 635 health conditions, and 22k small molecules;

• You may work with P3GPT either by downloading model weights for a local deployment or by interacting with the Discord bot on the official Inisilico Medicine's server.

Model usage guide

Run model with an endpoint

Details

Step 1 - connect to endpoint

import requests

API_URL = "https://cu2s6lgb4jew3tht.us-east-1.aws.endpoints.huggingface.cloud"
headers = {
    "Accept" : "application/json",
    "Authorization": "Bearer hf_XXXX",
    "Content-Type": "application/json" 
}

def query(payload):
    response = requests.post(API_URL, headers=headers, json=payload)
    return response.json()

Step 2 - create input for endpoint

import json
with open('./generation-configs/meta2diff.json', 'r') as f:
    config_data = json.load(f)

# prepare request configuration
request_config = {"inputs": config_data, "mode": "meta2diff", "parameters": {
    "temperature": 0.8,
    "top_p": 0.2,
    "top_k": 3550,
    "n_next_tokens": 50,
    "random_seed": 137
}}

Actual request processed by Precisou3GPT

[BOS]<age_group2diff2age_group><disease2diff2disease><compound2diff2compound><tissue>lung </tissue><age_individ></age_individ><cell></cell><efo>EFO_0000768 </efo><datatype>expression </datatype><drug>curcumin </drug><dose></dose><time></time><case>70.0-80.0 80.0-90.0 </case><control></control><dataset_type></dataset_type><gender>m </gender><species>human </species>

Step 3 - send the request to endpoint

output = query(request_config)

Endpoint output structure

{
    "output": {
        "up": List, 
        "down": List
    },
    "mode": String, // Generation mode was selected
    "message": "Done!",  // or Error
    "input": String // Input prompt was passed

}

Note: If the mode was supposed to generate compounds, the output would contain compounds: List.

---

Run model locally

Details

Requirements: torch==2.0.1 einops==0.7.0 huggingface-hub==0.20.1 transformers==4.35.0

1. Download the repository https://huggingface.co/insilicomedicine/precious3-gpt-multi-modal

2. Inside the repository execute:

# init handler
from handler import EndpointHandler
precious3gpt_handler = EndpointHandler(path='./')

import json
with open('./generation-configs/meta2diff.json', 'r') as f:
    config_data = json.load(f)

# prepare request configuration
request_config = {"inputs": config_data, 
                  "mode": "meta2diff", 
                  "parameters": {
    "temperature": 0.8,
    "top_p": 0.2,
    "top_k": 3550,
    "n_next_tokens": 50,
    "random_seed": 137
}}

output = precious3gpt_handler(request_config)

---

Precious3GPT request configuration

Instruction (inputs.instruction in config)

Instructions define the experimental setting P3GPT will be simulating using the information provided in the prompt.

1. disease2diff2disease - generate an omics signature characterizing a disease / determine the disease based on a given signature;
2. compound2diff2compound - generate an omics signature of a compound-induced perturbation / determine the compound given its omics signature;
3. age_group2diff2age_group - generate differential omics for age groups / determine age groups provided differential gene lists

Generation Modes (mode in config)

Generation modes are not part of the prompt processed to P3GPT but may affect the way P3GPT's response is presented or processed:

1. meta2diff: The compound2diff2compound instruction can be executed either way. This mode tells P3GPT to return differentially expressed genes and not compounds;
2. diff2compound: The reverse of the meta2diff mode. Make sure to fill in 'up' and 'down' in the prompt first!
3. meta2diff2compound: Runs meta2diff first and applies diff2compound to its output with one call.

See Precious3GPT_example.ipynb tutorial notebook to learn more about building P3GPT requests.

---

Other meta-data (inputs. in config)

P3GPT can only simulate the experiments featuring the biomedical entities and metadata values present in p3_entities_with_type.csv

If you aim to study a tissue, a compound, or something else using P3GPT, make sure to check that the names of the entities you are using match those in this file.

Examples

In the following examples, all possible configuration fields are specified. You can leave some meta-data fields in the inputs section empty string("") or empty list([]).

Example 1: generate a disease signature

Details

If you want to generate a signature given specific metadata you can use the following configuration. Note, up and down fields are empty lists as you want to generate them. Here, we ask the model to generate a signature for a male human within in the 70-90 years age group, in the "lung" tissue, with "EFO_0000768" (Idiopathic pulmonary fibrosis).

{
    "inputs": {
        "instruction": ["age_group2diff2age_group", "disease2diff2disease"], 
        "tissue": ["lung"],
        "age": "",
        "cell": "", 
        "efo": "EFO_0000768", 
        "datatype": "", "drug": "", "dose": "", "time": "", "case": ["70.0-80.0", "80.0-90.0"], "control": "", "dataset_type": "expression", "gender": "m", "species": "human", "up": [], "down": []
    }, 
    "mode": "meta2diff", 
    "parameters": {
        "temperature": 0.8, "top_p": 0.2, "top_k": 3550, "n_next_tokens": 50, "random_seed": 137
    }
}

See the corresponding P3GPT output:

{
  "output": {
    "up": [["PTGDR2", "CABYR", "MGAM", "TMED9", "SHOX2", "MAT1A", "MUC5AC", "GASK1B", "CYP1A2", "RP11-266K4.9", ...]], // generated list of up-regulated genes
    "down": [["MB", "OR10V1", "OR51H1", "GOLGA6L10", "OR6M1", "CDX4", "OR4C45", "SPRR2A", "SPDYE9", "GBX2", "ATP4B", ...]] // generated list of down-regulated genes
  },
  "mode": "meta2diff", // generation mode we specified
  "message": "Done!",
  "input": "[BOS]<age_group2diff2age_group><disease2diff2disease><tissue>lung </tissue><cell></cell><efo>EFO_0000768 </efo><datatype></datatype><drug></drug><dose></dose><time></time><case>70.0-80.0 80.0-90.0 </case><control></control><dataset_type>expression </dataset_type><gender>m </gender><species>human </species>", // actual input prompt for the model
  "random_seed": 137
}

Example 2: generate an aging signature

Details

Now, let's generate a signature for the whole blood of a healthy male human in the 70-90 years age group. Note, here we expect to generate the signatures for a healthy human, that's why we set efo to an empty string "".

{
    "inputs": {
        "instruction": ["age_group2diff2age_group"],
        "tissue": ["whole blood"],
        "age": "",
        "cell": "",
        "efo": "",
        "datatype": "", "drug": "", "dose": "", "time": "", "case": "40.0-50.0", "control": "", "dataset_type": "expression", "gender": "m", "species": "human", "up": [],
        "down": []
    },
    "mode": "meta2diff",
    "parameters": {
        "temperature": 0.8,
        "top_p": 0.2,
        "top_k": 3550,
        "n_next_tokens": 50,
        "random_seed": 137
    }
}

P3GPT's output: json { "output": { "up": [["IER3", "APOC2", "EDNRB", "JAKMIP2", "BACE2", ... ]], "down": [["TBL1Y", "TDP1", "PLPP4", "CPEB1", "ITPR3", ... ]] }, "mode": "meta2diff", "message": "Done!", "input": "[BOS]<age_group2diff2age_group><tissue>whole blood </tissue><cell></cell><efo></efo><datatype></datatype><drug></drug><dose></dose><time></time><case>40.0-50.0 </case><control></control><dataset_type>expression </dataset_type><gender>m </gender><species>human </species>", "random_seed": 137 }

Multi-Modality

By default, all tasks with a signature in the input prompt are executed with multimodal features. For each gene in the up-/down- lists, P3GPT pulls the embeddings from the Knowledge Graph and Text neural modelity mappers. Then, the embeddings are averaged to obtain one embedding for each modality and each gene list (4 averaged embeddings in total).

Cite this model

Please, cite the following bioRxiv pre-print if you use P3GPT in your research papers or other published materials:

@article {Galkin2024.07.25.605062,
    author = {Galkin, Fedor and Naumov, Vladimir and Pushkov, Stefan and Sidorenko, Denis and Urban, Anatoly and Zagirova, Diana and Alawi, Khadija M and Aliper, Alex and Gumerov, Ruslan and Kalashnikov, Aleksand and Mukba, Sabina and Pogorelskaya, Aleksandra and Ren, Feng and Shneyderman, Anastasia and Tang, Qiuqiong and Xiao, Deyong and Tyshkovskiy, Alexander and Ying, Kejun and Gladyshev, Vadim N. and Zhavoronkov, Alex},
    title = {Precious3GPT: Multimodal Multi-Species Multi-Omics Multi-Tissue Transformer for Aging Research and Drug Discovery},
    elocation-id = {2024.07.25.605062},
    year = {2024},
    doi = {10.1101/2024.07.25.605062},
    publisher = {Cold Spring Harbor Laboratory},
    abstract = {We present a multimodal multi-species multi-omics multi-tissue transformer for aging research and drug discovery capable of performing multiple tasks such as age prediction across species, target discovery, tissue, sex, and disease sample classification, drug sensitivity prediction, replication of omics response and prediction of biological and phenotypic response to compound treatment. This model combines textual, tabular, and knowledge graph-derived representations of biological experiments to provide insights into molecular-level biological processes. We demonstrate that P3GPT has developed an intuition for the interactions between compounds, pathologies, and gene regulation in the context of multiple species and tissues. In these areas, it outperforms existing LLMs and we highlight its utility in diverse case studies. P3GPT is a general model that may be used as a target identification tool, aging clock, digital laboratory, and scientific assistant. The model is intended as a community resource available open source as well as via a Discord server.Competing Interest StatementThe authors are affiliated with Insilico Medicine, a commercial company developing and using generative artificial intelligence and other next-generation AI technologies and robotics for drug discovery, drug development, and aging research. Utilizing its generative AI platform and a range of deep aging clocks, Insilico Medicine has developed a portfolio of multiple therapeutic programs targeting fibrotic diseases, cancer, immunological diseases, and a range of age-related diseases.},
    URL = {https://www.biorxiv.org/content/early/2024/07/25/2024.07.25.605062},
    eprint = {https://www.biorxiv.org/content/early/2024/07/25/2024.07.25.605062.full.pdf},
    journal = {bioRxiv}
}