Biomedical algorithm project experience

Find verifiable relationships in complex biological data.

In these field projects I did biomedical cleaning, exploration, features, deep-learning models, and evaluation, covering research validation and industry R&D scenes.

Contact：Study ExchangeEmail：StudyExchange@163.com

Project experience

Work I have done

From raw data to reproducible experiments.

First understand the research question and evaluation metrics, then determine model complexity. Deep learning is used only when justified.

Data exploration and quality assessment

Assess missing data, bias, duplicates, label distributions, and potential leakage to determine which conclusions the data can support.

Biological feature engineering

Engineer features from sequences, structures, experimental conditions, and domain variables while preserving an interpretable, traceable data pipeline.

Deep learning modeling

Select suitable network architectures, loss functions, and sampling strategies for the task, and validate design choices through ablation studies.

Rigorous experimental evaluation

Use appropriate data splits, cross-validation, and baseline models to avoid optimistic results from a single random split.

Result interpretation and visualization

Go beyond a single score by analyzing errors, model boundaries, and the factors that drive outcomes.

Reproducible notes

Organize processing scripts, training code, evaluation, run notes, and key experiment logs.

Real project breakdown

Glycan–protein binding prediction project

The competition required predicting Fraction_Bound, a binding ratio from 0 to 1, from glycans or molecules, proteins, and experimental concentrations. The project iterated around three core challenges: generalization to unseen proteins, scarce strong-binding samples, and complex interactions.

TaskRegression prediction

Input molecules, proteins, and concentrations; output binding ratios.

Dataset size571,930

Total training samples recorded by the project.

Approaches exploredS1–S22

Traditional features, pretrained dual towers, multi-tower GNN, and ensembles.

Best submission0.013 / 0.007

S11 Public / Private MAE; lower is better.

Outline of the final approach

S11: Joint modeling of graph structure, protein semantics, and traditional chemical features

The best single model did not merely concatenate fields. It used separate encoders for each input type and explicitly learned glycan–protein interactions.

01Glycan graph encoding

Represent GlycanAA as a heterogeneous graph and extract topological features with RGCN, GIN, and RelGATv2 towers.

02Protein encoding

Use pretrained ESM-2 embeddings to preserve protein-sequence semantics and generalization to unseen proteins.

03Chemical feature augmentation

Add 1024-bit Morgan fingerprints to capture local structural signals the graph network may miss.

04Interaction and regression

Combine concentration features and predict Fraction_Bound through low-rank bilinear interactions, gating, and a deep MLP.

Key project insightStrong local OOF performance does not guarantee online generalization. The test set contained many proteins unseen during training, so the project reduced reliance on statistics from seen proteins, adopted ESM-2 sequence representations, and validated by protein groups to prevent data leakage.

Example use cases

Start with a clearly defined research question.

These are field scenes I practiced or unpacked; the method depends on real data and the research goal.

Binding prediction

Glycan–protein binding probability prediction

Combine glycan representations, protein sequences, and experimental context; address class imbalance; and compare encoding and network architectures.

Output: predictive model, validation metrics, key experiments, and error analysis

Medical data

Patient risk stratification or outcome prediction

Integrate clinical metrics, time series, or omics features with rigorous data-splitting and external-validation strategies.

Output: risk model, calibration results, important features, and applicability boundaries

Research analysis

Pattern discovery in experimental data

Apply dimensionality reduction, clustering, differential analysis, and visualization to high-dimensional experimental data to generate testable hypotheses.

Output: analysis report, charts, code, and candidate research directions

How the project was done · Project experience

Research questions require evidence to be defined before models.

Biomedical projects usually begin with the research hypothesis, experimental design, and data-generation process. Model scores are only part of the outcome; rigorous validation, reproducible conclusions, and understanding where a model fails matter more.

Materials prepared in the project Data quality and feasibility report Reproducible training and inference code Model metrics, ablations, and error analysis Charts and technical reports for papers or internal review

01
Research question and data review
Define prediction targets, experimental units, label sources, batch effects, ethics, and data permissions to determine whether the available data can answer the question.
02
Baseline and validation design
Establish interpretable baselines, group appropriately by patient, protein, molecule, or time, and agree on metrics and external validation in advance.
03
Modeling and error-analysis loop
Iterate features, models, and loss functions, examining errors by subgroup, target range, and difficult samples instead of relying only on an aggregate score.
04
Reproduction, reporting, and follow-up experiments
Organize code, environments, and experiment records; deliver reviewable reports; and translate model findings into testable hypotheses for the next experiment.

Public Kaggle competition record

1 / 2

Glycan–Protein Binding Prediction

Glycan–protein binding competition practice, showing complex biological data work, modeling, and iteration. Rank follows the public leaderboard.

View leaderboard