Multi Task Learning Feature Fusion

---
name: multi-task-learning-feature-fusion
description: Use when when you have multi-branch deep learning architecture predicting
  related but distinct peptide properties (charge state, isotope count, retention
  time) from raw mass spectrum, and you want to leverage auxiliary task gradients
  to improve primary task learning.
license: CC-BY-4.0
metadata:
  edam_operation: http://edamontology.org/operation_3927
  edam_topics:
  - http://edamontology.org/topic_0121
  - http://edamontology.org/topic_3520
  tools:
  - PyTorch
  - IsoFusion
  techniques:
  - mass-spectrometry
  license_tier: open
  provenance_tier: literature
derived_from:
- doi: 10.26599/bdma.2024.9020059
  title: IsoFusion
evidence_spans:
- _No usage/docs found._
claims: []
provenance:
  collection: https://w3id.org/holobiomicslab/asb-skill/collection/metabolomics/v2
  assembled_by: scripts/collect_metabolomics_collection.py
  sources:
  - build: coll_isofusion_cq
    doi: 10.26599/bdma.2024.9020059
    title: IsoFusion
  dedup_kept_from: coll_isofusion_cq
schema_version: 0.2.0
attribution:
  generator: AgenticScienceBuilder
  original_doi: 10.26599/bdma.2024.9020059
  all_source_dois:
  - 10.26599/bdma.2024.9020059
  zenodo_doi: 10.5281/zenodo.20794027
  curators: []
  promoter: Louis-Félix Nothias
  sponsor: CNRS & Université Côte d'Azur
---

# multi-task-learning-feature-fusion

## Summary

Implement a FuseBlock module that integrates feature tensors from multiple prediction branches (isotope, charge, retention time) via dimension-wise fusion and learned weighted combination, enabling joint optimization across auxiliary tasks to improve main task performance in end-to-end peptide feature detection from mass spectrometry data.

## When to use

When you have multi-branch deep learning architecture predicting related but distinct peptide properties (charge state, isotope count, retention time) from raw mass spectrum, and you want to leverage auxiliary task gradients to improve primary task learning. Use this skill when individual branch outputs are high-dimensional feature tensors that would benefit from explicit cross-dimensional integration rather than independent prediction heads.

## When NOT to use

- Input branches produce independent scalar predictions rather than feature tensors — use simple concatenation or weighted averaging instead.
- Tasks are unrelated or negative transfer is suspected between auxiliary and main tasks — validate multi-task benefit empirically first.
- Model already includes explicit attention or gating mechanisms connecting branches — FuseBlock would be redundant.

## Inputs

- Feature tensor from isotope prediction branch
- Feature tensor from charge state prediction branch
- Feature tensor from retention time prediction branch

## Outputs

- Unified fused feature representation tensor
- Gradient signals propagated to each input branch

## How to apply

Define a FuseBlock module that accepts three separate feature tensors from isotope, charge, and retention-time prediction branches. Fuse dimensions by concatenating or applying learned weighted combinations across the three feature sources. Feed the combined representation through fully connected layers to produce a unified fused feature tensor. Route gradients from the fused representation back to each branch during backpropagation to propagate multi-task learning signals. Validate output tensor shapes match expected dimensions and confirm gradient flow through all branches during test backprop pass.

## Related tools

- **PyTorch** (Framework for implementing FuseBlock module architecture, gradient computation, and multi-task backpropagation)
- **IsoFusion** (Reference implementation of FuseBlock within end-to-end peptide feature detection model) — https://github.com/xfcui/IsoFusion

## Examples

```
# In PyTorch, after defining isotope_features, charge_features, retention_time_features tensors from three branches:
fused = torch.cat([isotope_features, charge_features, retention_time_features], dim=1)
fused_out = fc_layers(fused)
loss_main.backward()
loss_auxiliary.backward()
```

## Evaluation signals

- Output tensor shape is consistent with concatenation/fusion of three input feature dimensions (e.g., [batch_size, fused_feature_dim])
- Gradient flow backward through FuseBlock successfully updates weights in all three input branches without NaN or zero gradients
- Loss from main task (e.g., peptide identification) decreases when auxiliary task gradients are backpropagated through fused representation, compared to independent branch training
- Fused feature representation encodes information from all three dimensions, verifiable by ablation (withholding one input branch) causing performance degradation
- Multi-task learning improves main task metric relative to single-task baseline, confirming auxiliary tasks (charge, isotope, retention time) help primary task learning

## Limitations

- FuseBlock requires careful tuning of learned weights or fusion strategy (concatenation vs. attention); poor initialization can degrade multi-task signal propagation.
- Auxiliary tasks (charge, isotope count, retention time prediction) must be sufficiently correlated with main task to avoid negative transfer that reduces overall performance.
- Method assumes all three feature branches produce comparable-scale tensors; mismatched feature dimensions may require normalization or separate projection layers before fusion.

## Evidence

- [intro] FuseBlock module architecture and multi-task learning integration: "FuseBlock that integrates features from different dimensions. (c) Using the multi-task learning to predict charge, number of isotopes and retention time simultaneously, the auxiliary task can help"
- [other] Feature tensor inputs from three prediction branches: "Define the FuseBlock module architecture to accept separate feature tensors from isotope, charge, and retention-time prediction branches."
- [other] Fusion mechanism via learned combination and gradient flow: "Implement dimension-wise feature fusion using concatenation or learned weighted combination across the three feature dimensions. Apply a series of fully connected layers to produce a unified fused"
- [other] Backpropagation through fused representation to all branches: "Integrate gradient flow from the fused representation back to each feature dimension branch to enable multi-task learning signal propagation."
- [intro] End-to-end peptide feature detection context: "our model is an end-to-end model that can predict charge, number of isotopes and retention time directly from the mass spectrum"