Dual branch feature fusion

dual-branch-feature-fusion

License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution. <!-- asb-license-banner -->

Summary

Fuse complementary molecular representations (fingerprint and graph-structured data) through parallel processing branches that independently embed each modality before combining them in a learned fusion layer. This dual-branch strategy exploits both the compactness of fixed-length fingerprints and the topological richness of molecular graphs to improve retention time prediction accuracy.

When to use

When you have molecular input data available in two or more distinct formats (e.g., RDKit-extracted fingerprints AND torch_geometric Graph objects representing molecular topology) and your prediction target (e.g., retention time, property value) may benefit from both fixed-feature and graph-structured context. Apply this skill when single-modality baseline models underperform and you suspect that structural information alone or feature vectors alone are insufficient.

When NOT to use

• Input is already a single unified feature table (e.g., pre-concatenated fingerprint + manually engineered descriptors). Use simple MLPs or linear regressors instead.
• Molecular graph data is incomplete or unavailable; fingerprints alone suffice. Single-branch models are simpler and faster.
• Training set is very small (< 100 examples). Multi-branch architectures risk overfitting; consider transfer learning or simpler baselines first.

Inputs

• Molecular fingerprint tensor: shape [batch_size, fingerprint_dim], dtype float32 (e.g., 1024-bit or 2048-bit RDKit fingerprints)
• Molecular graph batch object: torch_geometric Data object containing node features, edge indices, and batch assignment for multiple graphs
• Retention time labels (optional, for training): shape [batch_size], dtype float32

Outputs

• Retention time predictions: tensor of shape [batch_size, 1], dtype float32, containing scalar retention time values in the units of the training target
• Fused molecular embedding: tensor of shape [batch_size, embedding_dim], the learned joint representation before the regression head (useful for downstream analysis or transfer learning)

How to apply

Instantiate two independent processing branches: (1) a fingerprint branch that accepts RDKit-computed molecular fingerprints as dense tensors and projects them through fully connected layers; (2) a graph neural network branch that accepts torch_geometric Graph objects and applies graph convolution layers (e.g., GCN, GIN) to learn node-level and graph-level embeddings. After independent processing, concatenate or apply learned attention-based fusion on the two branch outputs to produce a joint embedding. Feed the fused representation into a regression head (fully connected layers + output neuron) to predict the scalar target. Verify correct forward-pass execution by running a sample batch through the complete model, confirming output shape matches [batch_size, 1] and contains numeric predictions without NaN or Inf values.

Related tools

• torch (Core deep learning framework for defining and training the dual-branch neural network modules)
• torch_geometric (Graph neural network library providing graph convolution layers and batch-wise graph handling for the graph branch)
• rdkit-pypi (Computes molecular fingerprints (input to the fingerprint branch) and generates initial node/edge features for graph construction)
• torch-scatter, torch-sparse, torch-cluster (Low-level PyTorch extensions required by torch_geometric for efficient sparse tensor operations and graph sampling)
• RT-Transformer (Reference implementation combining fingerprint and molecular graph inputs for retention time prediction) — https://github.com/01dadada/RT-Transformer

Examples

import torch; from rt_transformer import RTTransformer; model = RTTransformer(fingerprint_dim=1024, graph_hidden_dim=64, fusion_dim=128); fp_batch = torch.randn(8, 1024); graph_batch = create_graph_batch_from_smiles(smiles_list); preds = model(fp_batch, graph_batch); print(preds.shape)  # torch.Size([8, 1])

Evaluation signals

• Forward pass executes without RuntimeError or NaN values; output shape is exactly [batch_size, 1].
• Gradients flow through both branches and the fusion layer during backward pass (check torch.autograd.grad or .backward() without errors).
• Fused embedding dimensionality and composition reflect both fingerprint and graph branch contributions (inspect intermediate tensor shapes and concatenation).
• Model achieves lower mean absolute error or root mean squared error on hold-out test set compared to single-branch (fingerprint-only or graph-only) baselines.
• Ablation study: zeroing one branch's gradients or removing it should degrade performance, confirming both branches contribute predictive signal.

Limitations

• Architecture assumes fingerprints and graph representations are available for all input molecules; missing or malformed inputs require preprocessing or filtering.
• Different chromatographic conditions may produce different retention times for the same metabolite, limiting the model's cross-method transferability unless transfer learning is applied.
• Fusion layer design (concatenation vs. attention vs. bilinear pooling) is not automatically optimal; hyperparameter tuning and architecture search may be needed.
• Computational cost scales with graph size and batch size; very large molecular graphs or extremely large batches may cause memory overflow.

Evidence

• [other] The RT-Transformer model uses a dual-branch architecture that accepts fingerprint data from one branch and molecular graph data from another branch, with both inputs processed to produce a retention time prediction as output.: "The RT-Tranformer combine the fingerprint and the molecular graph data and predict retention time as the output."
• [other] Load molecular fingerprints from RDKit feature extraction and construct molecular graph representations using torch_geometric Graph objects.: "Load molecular fingerprints (from RDKit feature extraction) and construct molecular graph representations using torch_geometric Graph objects."
• [other] Initialize fingerprint processing branch and graph neural network branch with torch_geometric convolution layers.: "instantiate the fingerprint processing branch and the graph neural network branch (using torch_geometric convolution layers)"
• [other] Define a fusion layer that combines fingerprint and graph embeddings, then implement a regression head that outputs scalar retention time predictions.: "Define the fusion layer that combines fingerprint and graph embeddings. 4. Implement the regression head that outputs scalar retention time predictions."
• [other] Verify model runs forward pass without errors and produces output tensor of shape [batch_size, 1] with numeric predictions.: "confirm model runs forward pass without errors and produces output tensor of shape [batch_size, 1] with numeric predictions"
• [readme] Liquid chromatography retention times prediction can assist in metabolite identification in non-targeted metabolomics.: "Liquid chromatography retention times prediction can assist in metabolite identification, which is a critical task and challenge in non-targeted metabolomics."
• [readme] Different chromatographic conditions may result in different retention times for the same metabolite, limiting cross-method transferability.: "different chromatographic conditions may result in different retention times for the same metabolite"