Biomedical AI Resource Track Baseline
A small, fully transparent text-classification baseline that assigns public biomedical-AI resource descriptions to one of four atlas tracks:
- Computational Biology
- Human Biobank
- AI Agents & Agentic AI
- Human Data Integration & AI
This is an original curation-support model, not a duplicated foundation model and not a scientific or clinical predictor.
Intended use
The model can help route newly discovered Hugging Face resources into the appropriate section of the RNA, Human Genetics & AI Atlas. It is useful as a reproducible baseline for testing future improvements to atlas curation.
It must not be used for biological inference, target prioritization, clinical decisions, candidate screening, or assessment of scientific quality.
Architecture
- Multinomial Naive Bayes
- Lowercase word unigrams and adjacent word bigrams
- Features retained when observed at least twice
- Laplace smoothing, $\alpha=1$
- 165 retained features
- JSON weights; no proprietary runtime or binary serialization
- Python standard library only
Training data
The model was trained on all 20 descriptions in Shicheng/rna-human-genetics-ai-atlas, snapshot reviewed on 2026-11-01.
The exact SHA-256 fingerprint of resources.csv is stored in model.json. Resource titles and evidence summaries are the only predictive text. Likes, downloads, owners, and review questions are not used as model inputs.
Evaluation
Evaluation used leave-one-out cross-validation over all 20 records.
| Metric | Result |
|---|---|
| Accuracy | 0.8000 |
| Macro F1 | 0.7804 |
| AI Agents & Agentic AI F1 | 0.7500 |
| Computational Biology F1 | 0.8000 |
| Human Biobank F1 | 0.5714 |
| Human Data Integration & AI F1 | 1.0000 |
Full predictions, probabilities, per-class metrics, and the confusion matrix are available in evaluation.json.
Run inference
python predict.py "Agent that integrates genetics, clinical trials, and target-evidence provenance"
The output is JSON containing the predicted track and normalized class probabilities.
Reproduce training and evaluation
Place this repository beside a checkout of the atlas dataset using the same directory layout, then run:
python train.py
train.py regenerates both model.json and evaluation.json using only the Python standard library.
Limitations
- Twenty examples are far too few for a general-purpose classifier.
- The labels reflect one curator's practical taxonomy rather than a universal ontology.
- Resource descriptions often span multiple tracks.
- Probabilities are not calibrated confidence estimates.
- Human Biobank performance is weakest and should be reviewed manually.
- The evaluation reuses the same small curation snapshot and does not establish external generalization.
Relationship to the atlas
Maintainer
Created by Shicheng Guo as a public, nonconfidential demonstration of transparent biomedical-AI resource curation.
Dataset used to train Shicheng/biomedical-ai-resource-track-baseline
Evaluation results
- Leave-one-out accuracy on RNA, Human Genetics & AI Atlasself-reported0.800
- Leave-one-out macro F1 on RNA, Human Genetics & AI Atlasself-reported0.780