---

As of writing this document (June 6, 2025), the API interface of SG4D100M on Vertex AI is still under development. The following content is prepared before the release date and may differ in the final implementation.

---

Getting Started with SG4D100M

Objective

This notebook demonstrates how to use the Vertex AI API to interact with SG4D100M for molecular embedding generation.

Setting up

First, let’s set up the required configuration:

• API_ENDPOINT: The URL endpoint for sending your requests
• ACCESS_TOKEN: The bearer token for authenticating your HTTP requests

In [1]:

import subprocess

# Replace these with your actual project details
PROJECT_NAME = "__YOUR_PROJECT_NAME_GOES_HERE__"
PROJECT_LOCATION = "__YOUR_PROJECT_LOCATION_GOES_HERE__"

# Construct the API endpoint URL
API_ENDPOINT = '/'.join([
    f"https://{PROJECT_LOCATION}-aiplatform.googleapis.com/v1",
    f"/projects/{PROJECT_NAME}",
    f"/locations/{PROJECT_LOCATION}/",
    "publishers/syntheticgestalt",
    "models/sg4d100m:rawPredict",
])

# Get the access token using gcloud CLI
result = subprocess.run(
    args=["gcloud", 'auth', 'print-access-token'],
    capture_output=True,
)
ACCESS_TOKEN = result.stdout.decode('utf-8').strip()

Running Online Inference

Model Input

Let’s create embeddings for the following compounds:

In [2]:

import polars as pl

# Example molecules (caffeine and a Theanine)
request_df = pl.DataFrame([
    {"smiles": "CN1C=NC2=C1C(=O)N(C(=O)N2C)C"},  # Caffeine
    {"smiles": "CCNC(=O)CC[C@H](N)C(=O)O"},      # Theanine
])
request_df

shape: (2, 1)

smiles
str
"CN1C=NC2=C1C(=O)N(C(=O)N2C)C"
"CCNC(=O)CC[C@H](N)C(=O)O"

API Invocation

We can run the predictions using HTTP POST requests:

In [8]:

import requests

response = requests.post(
    url=API_ENDPOINT,
    headers={
        "Authorization": f"Bearer {ACCESS_TOKEN}",
        "Content-Type": "application/json",
    },
    json={
        "sg4d100m_version": "sg4d100m-2025-06-06",
        "messages": request_df.to_dicts(),
    },
)

Model Output

The model returns molecular embeddings that can be converted to a DataFrame:

In [9]:

response_df = pl.from_dicts(response.json())
response_df

Next Steps

• Use the embeddings for similarity search
• Perform clustering analysis
• Predict molecular properties