Team9_Toxicological_Profiles

Aim

Machine learning tool development for modeling concentration-activity profiles of estrogen-related receptors

Background

Estrogen related receptors (ERRs) are one of the first orphan nuclear receptors identified.[1] ERRs are required for high-energy production in response to the environmental and physiological challenges. They play an important role in the control of cellular energy, including mitochondrial biogenesis, gluconeogenesis, and oxidative phosphorylation[1, 2], with signaling implicated in metabolic disorders like type 2 diabetes[2], with EER-alpha (subfamily of EERs) identified as an adverse marker for breast cancer progression.[3]
Screening and identifying environmental compounds that perturb the EER signaling pathways could provide information for potential preventive measures in treating the mentioned metabolic diseases.[1] Predicting their activity continuously allows for a quantitative assessment of receptor activation or inhibition, reflecting real biological responses more accurately than binary (toxic/nontoxic) labels, additionally allowing for information on dose-response that is not present in binary toxicity prediction.

Contributors

• Marek Sokołowski
• Ritwik Ganguly
• Nilabja Bhattacharjee

Methods

Data Preprocessing and Graph Construction

Molecular data were obtained in SMILES format, each corresponding to a compound tested under varying concentrations for biological activity inhibition (e.g., from the Tox21 dataset). Each molecule was converted into a graph structure using RDKit, where atoms represent nodes and bonds represent edges. For each compound, a graph was constructed using the smiles2graph utility, and the resulting graphs were encapsulated in torch_geometric.data.Data objects.

Each node (atom) was featurized using the atom_features function, which encodes atomic number, degree, hybridization, aromaticity, and other physicochemical properties into a numerical feature vector. Edge indices were defined based on molecular bonds, but no explicit edge features were used.

To incorporate dose-dependent effects, the concentration value (log-transformed) was included as an additional scalar node-level feature, broadcast across all nodes in the graph. The final node feature dimension thus includes both structural atom-level information and experimental dosage context.

The dataset was split into training (80%), validation (10%), and test (10%) sets using stratified shuffling to preserve label distributions. The torch_geometric.loader.DataLoader was used to batch graphs with variable sizes efficiently for GNN processing. All features and targets were converted to float32 and normalized where necessary.

Graph Neural Network Architecture

We employed a Graph Convolutional Network (GCN) architecture implemented using PyTorch Geometric. The model consists of:

Three stacked GCNConv layers with hidden dimensions of 64, 32, and 16, each followed by ReLU activation.

A global_mean_pool operation aggregates node-level embeddings into a graph-level representation.

Two fully connected (linear) layers map the pooled graph embedding to the final scalar output, representing the predicted % inhibition.

The model was trained to perform regression, using the Mean Squared Error (MSE) loss between predicted and experimental inhibition values.

Training Procedure

The model was trained using the Adam optimizer with an initial learning rate of 1e-3 and a weight decay of 5e-4. A ReduceLROnPlateau scheduler was used to dynamically adjust the learning rate based on validation loss stagnation. Early stopping was employed with a patience of 20 epochs to prevent overfitting.

Each training epoch consisted of a forward pass, loss computation, backpropagation, and parameter update. Model performance was monitored on the validation set after each epoch. The best-performing model based on validation loss was checkpointed and subsequently used for final evaluation on the test set.

Results

Conclusions

Future directions

References

[1] Aubert G, Vega RB, Kelly DP. Perturbations in the gene regulatory pathways controlling mitochondrial energy production in the failing heart. In: Zhu H, Xia M, editors. Biochimica et Biophysica Acta (BBA) - Molecular Cell Research. Volume 1833, Issue 4. 1st ed. Elsevier; 2013. p. 840-847. https://doi.org/10.1016/j.bbamcr.2012.08.015.
[2] Audet-walsh, É., Giguére, V. The multiple universes of estrogen-related receptor α and γ in metabolic control and related diseases. Acta Pharmacol Sin 36, 51–61 (2015). https://doi.org/10.1038/aps.2014.121
[3] Huang R. A Quantitative High-Throughput Screening Data Analysis Pipeline for Activity Profiling. In: Zhu H, Xia M, editors. High-Throughput Screening Assays in Toxicology. Methods in Molecular Biology. 1473. 1 ed: Humana Press; 2016

Important Links

• link to the dataset
• introduction to the problem
• Simmilar solutions: https://paperswithcode.com/sota/drug-discovery-on-tox21

Dataset

The dataset has been created to allow for learning models based on EER toxicity in the following manner: Input: SMILES (string) and concentration (float). Output (4-dimensional vector): DATA corresponding to agonist, DATA corresponding to antagonist, DATA corresponding to viability, and DATA corresponding to autofluorescence at the provided concentration.

With data split:
train: 5640 samples,
test: 704 samples,
validation: 704 samples.

Columns Description

The dataset is based on the Tox21 Dataset merged with SMILES and IDs from Pubchem.

• SAMPLE_DATA_TYPE: distinguish channels:

  • cell_red: Autofluorescence of molecules
  • agonist(1-3):
  • antagonist(1-3):
  • viability(1-3): where 1-3 indicates measurement repetition.

• DATA(1-14): measurement of fluorescence in a specific experiment, it contains float values, positive and negative. None values indicate a lack of measurement

• CONC(1-14): concentration in $\mu$M at which DATA is measured, it consists of positive float values, with None values in case of a lack of measurement where each DATA is measured at a specific CONC indicated by id (ie, DATA1 is measured at CONC1, DATA2 is measured at CONC2, etc),

• canonical_smiles:

• similarity_order:

• iupac_name:

Information for tracking measurement or molecule source for data check: PUBCHEM_CID, TOX21_ID, SAMPLE_NAME, PUBCHEM_SID.

• CURVE_CLASS2: Tox21 group description of each measurement series (DATA/CONC 1-15) based on the Hill equation or set manually.
• PubChemFingerprint:
• similarity_order:

Alternative Solutions

Based on the Tox21 Dataset there is Therapeutics Data Commons (TDC) dataset, and Tox24 Chalange dataset. Tox21 benchmark:

• XGBoost:

Tox24:

• Keggle competitors
• XGBoost:
  • Features:
    1. Fingerprints: ErG, Mordred, Pubchem, MACCS, RDKit
    2. Fingerprints: 'fgr', 'datamol', 'ALogPSOEState', 'Mold2', 'SIRMSmix', 'MAP4', 'atombond', 'estate', 'JPlogP', 'ISIDAfragments'
  • algorithms:
    1. Cross-validation using XGBoost
    2. XGBoost

Best TDC models according to Papers With Code:

• MapLight
  • Features:
    1. fingerprints: extended-connectivity fingerprints (ECFP), Avalon, and extended reduced graph approach (ErG),
    2. 200 molecular properties, including: "number of rings, molecular weight"
  • algorithms:
    1. Parameter search via grid search CV,
    2. CatBoost,
    3. LightGBM
• XGBoost:
  • Features:
    1. Fingerprints: MACCS, ECFP, Mol2Vec,
    2. Descriptors: Pubchem (881 distinct structural features), Molecular Access System (Mordred), RDKit
  • algorithms:
    1. Parameter search via randomized grid search CV,
    2. Extreme Gradient Boosting (XGBoost)

Workflow:

flowchart TD;
    A[Dataset Creation<br>for EER Toxicity prediction<be>based on Tox21 assays: tox21-err-p1, tox21-spec-hek293-p1<be>and PubChem standarysation] -->|Data Cleaning| B@{shape: cyl, label: Mearged Dataset};
    B --> C@{shape: trap-b, label: Model Preselection};
    C --> D[Traditional Machine Learning];
    C --> E[Graph Neural Network];
    C --> F[Gradient boosted aproaches];
    D & E & F --> G@{ shape: diamond, label: Model Benchmarking};
    G --> H[Model training on the full dataset];
    H --> I[Based on the best model outcomes:<br>**Prediction of toxicophoric groups**<br>a functional groups with the potential of inducing toxicity];

Loading

Hackaton Progress

flowchart TD;
    subgraph Done
        A("Database creation<br>based on the simmilarity matrix") --> B(Creation of Descriptions for the Hackation Participants);
        A --> C(EDA)
    end

    subgraph in progress
     A --> D(Model creation);
     A --> E(Consideration of a knkown Hill curve class in data split)
    end

Loading