NeuSin#

Index#

  1. Instantiate model class

  2. Define clock metadata

  3. Download clock dependencies

  4. Load features

  5. Load weights into base model

  6. Load reference values

  7. Load preprocess and postprocess objects

  8. Check all clock parameters

  9. Basic test

  10. Save torch model

  11. Clear directory

Let’s first import some packages:

[1]:
import os
import inspect
import shutil
import json
import torch
import pandas as pd
import pyaging as pya

Instantiate model class#

[2]:
def print_entire_class(cls):
    source = inspect.getsource(cls)
    print(source)

print_entire_class(pya.models.NeuSin)
class NeuSin(LinearReferenceClock):
    pass

[3]:
model = pya.models.NeuSin()

Define clock metadata#

[4]:
model.metadata["clock_name"] = "neusin"
model.metadata["data_type"] = "DNA methylation"  # Paper: The model is based on DNA methylation measurements.
model.metadata["species"] = "Homo sapiens"  # Paper: The study samples are Homo sapiens.
model.metadata["year"] = 2024
model.metadata["approved_by_author"] = "⌛"
model.metadata["citation"] = "Tong, Huige, et al. \"Cell-type specific epigenetic clocks to quantify biological age at cell-type resolution.\" Aging 16 (2024): 13452–13504."
model.metadata["doi"] = "https://doi.org/10.18632/aging.206184"
model.metadata["notes"] = "Neu-Sin is a neuron semi-intrinsic chronological-age clock: elastic-net regression was restricted to neuron age-DMCTs but fitted to methylation values not adjusted for brain cell fractions."
model.metadata["research_only"] = None
model.metadata["tissue"] = ["brain cortex"]  # Paper: For training the clocks, we used one of the largest collections of prefrontal cortex (PFC) samples profiled with Illumina 450k technology, encompassing 416 samples with a wide age range (18–97 years).
model.metadata["predicts"] = ["chronological age"]  # Paper: Finally, elastic net predictors of chronological age are trained from the Neu-DMCTs parameterized by a penalty parameter.
model.metadata["training_target"] = ["chronological age"]  # Paper: Having identified the significant neuron-DMCTs, we next trained Elastic Net regression (glmnet R package) models (alpha = 0.5) for age, each parameterized by a different penalty parameter (lambda) value.
model.metadata["unit"] = ["years"]  # Paper: For training the clocks, we used one of the largest collections of prefrontal cortex (PFC) samples profiled with Illumina 450k technology, encompassing 416 samples with a wide age range (18–97 years).
model.metadata["model_type"] = "elastic net regression"  # Paper: Having identified the significant neuron-DMCTs, we next trained Elastic Net regression (glmnet R package) models (alpha = 0.5) for age, each parameterized by a different penalty parameter (lambda) value.
model.metadata["platform"] = ["Illumina 450K"]  # Paper: For training the clocks, we used one of the largest collections of prefrontal cortex (PFC) samples profiled with Illumina 450k technology, encompassing 416 samples with a wide age range (18–97 years).
model.metadata["population"] = "adults"  # Paper: For training the clocks, we used one of the largest collections of prefrontal cortex (PFC) samples profiled with Illumina 450k technology, encompassing 416 samples with a wide age range (18–97 years).
model.metadata["journal"] = "Aging"
model.metadata["last_author"] = "Andrew E. Teschendorff"
model.metadata["n_features"] = 672  # Paper: The official Neu-SinCoef.rda object contains 673 rows: one (Intercept) row and 672 non-intercept CpG coefficient rows.
model.metadata["citations"] = 25
model.metadata["citations_date"] = "2026-07-05"

Download clock dependencies#

[5]:
supplementary_url = "https://raw.githubusercontent.com/Duzhaozhen/OmniAge/c10fbe8cb92957520fbff1d55ae1def0691252e5/OmniAgePy/src/omniage/data/CTS/Neu-Sin.csv"
supplementary_file_name = "coefficients.csv"
os.system(f"curl -sL -o {supplementary_file_name} {supplementary_url}")
[5]:
0

Load features#

[6]:
df = pd.read_csv('coefficients.csv')
if str(df.columns[0]).startswith('Unnamed'):
    df = df.iloc[:, 1:]
mask = df['probe'].astype(str).str.lower().isin(['intercept', '(intercept)'])
intercept_value = float(df.loc[mask, 'coef'].iloc[0]) if mask.any() else 0.0
coef_df = df.loc[~mask].reset_index(drop=True)
model.features = coef_df['probe'].tolist()

Load weights into base model#

[7]:
weights = torch.tensor(coef_df['coef'].tolist()).unsqueeze(0).float()
intercept = torch.tensor([intercept_value]).float()
[8]:
base_model = pya.models.LinearModel(input_dim=len(model.features))

base_model.linear.weight.data = weights.float()
base_model.linear.bias.data = intercept.float()

model.base_model = base_model

Load reference values#

[9]:
model.reference_values = None

Load preprocess and postprocess objects#

[10]:
model.preprocess_name = None
model.preprocess_dependencies = None
[11]:
model.postprocess_name = None
model.postprocess_dependencies = None

Check all clock parameters#

[12]:
pya.utils.print_model_details(model)

%==================================== Model Details ====================================%
Model Attributes:

training: True
metadata: {'approved_by_author': '⌛',
 'citation': 'Tong, Huige, et al. "Cell-type-specific and '
             'cell-type-independent DNA methylation clocks." Aging 16 (2024).',
 'clock_name': 'neusin',
 'data_type': 'methylation',
 'doi': 'https://doi.org/10.18632/aging.206184',
 'notes': None,
 'research_only': None,
 'species': 'Homo sapiens',
 'version': None,
 'year': 2024}
reference_values: None
preprocess_name: None
preprocess_dependencies: None
postprocess_name: None
postprocess_dependencies: None
features: ['cg10626816', 'cg13571388', 'cg17826530', 'cg06711298', 'cg08193650', 'cg10442729', 'cg17343483', 'cg20361600', 'cg21185289', 'cg16369288', 'cg22702772', 'cg26856080', 'cg06385118', 'cg18171715', 'cg18586891', 'cg05477834', 'cg15690342', 'cg18635552', 'cg18745317', 'cg25108022', 'cg00537387', 'cg08692175', 'cg25739875', 'cg18048071', 'cg19421584', 'cg22493372', 'cg02324367', 'cg05492433', 'cg16241714', 'cg26003909']... [Total elements: 672]
base_model_features: None

%==================================== Model Details ====================================%
Model Structure:

base_model: LinearModel(
  (linear): Linear(in_features=672, out_features=1, bias=True)
)

%==================================== Model Details ====================================%
Model Parameters and Weights:

base_model.linear.weight: [0.6255767941474915, 0.00018867039761971682, 1.4650660753250122, -4.989912986755371, 2.8106305599212646, 3.3365237712860107, -8.365673065185547, -7.561515808105469, -2.2176127433776855, 2.238844394683838, 5.353684425354004, 18.882240295410156, -1.537793755531311, 18.469179153442383, 0.3016071319580078, -0.4679134786128998, 0.7296689748764038, -9.179381370544434, 0.04645940288901329, 7.833886623382568, -2.897672414779663, -21.346012115478516, 9.050196647644043, 0.5711203813552856, 0.47357064485549927, 0.11948469281196594, 0.0023688615765422583, -1.3019375801086426, 57.94570541381836, -2.4218344688415527]... [Tensor of shape torch.Size([1, 672])]
base_model.linear.bias: tensor([20.1315])

%==================================== Model Details ====================================%

Basic test#

[13]:
torch.manual_seed(42)
input = torch.randn(10, len(model.features), dtype=float)
model.eval()
model.to(float)
pred = model(input)
pred
[13]:
tensor([[ -26.3359],
        [ 213.3675],
        [-104.1039],
        [-339.0749],
        [ -66.5274],
        [ 113.9317],
        [ 141.6260],
        [ -65.5180],
        [  -2.5413],
        [-400.5478]], dtype=torch.float64, grad_fn=<AddmmBackward0>)

Save torch model#

[14]:
torch.save(model, f"../weights/{model.metadata['clock_name']}.pt")

Clear directory#

[15]:
# Function to remove a folder and all its contents
def remove_folder(path):
    try:
        shutil.rmtree(path)
        print(f"Deleted folder: {path}")
    except Exception as e:
        print(f"Error deleting folder {path}: {e}")

# Get a list of all files and folders in the current directory
all_items = os.listdir('.')

# Loop through the items
for item in all_items:
    # Check if it's a file and does not end with .ipynb
    if os.path.isfile(item) and not item.endswith('.ipynb'):
        os.remove(item)
        print(f"Deleted file: {item}")
    # Check if it's a folder
    elif os.path.isdir(item):
        remove_folder(item)
Deleted file: coefficients.csv