ai-biotech
An early-stage AI for Biotechnology (AI4Bio) research project in computational biology. It builds a validated data pipeline that starts from protein sequences, uses AlphaFold structure predictions, and ends in a reproducible, residue-level feature dataset — the foundation for future AI-assisted biological research.
The stack is deliberately small and inspectable: Python, Biopython, pandas, NumPy, and matplotlib, running on a standard machine with no GPU required.
- Repository
- github.com/kelvinchandradjayapranata-hue/ai-biotech
- Status
- Stages 1–6 complete · Stage 6 review pending · Stage 7 (machine learning) not started
- Dataset
- 7 proteins · 1,098 residues × 29 columns
- Integrity tests
- 72 of 72 passing across four suites (20 + 12 + 16 + 24)
- Data source
- AlphaFold Database v6, plus an AlphaFold Server reference structure (ubiquitin, 76 aa)
- Language
- Python 3.12.10
Pipeline stages
| Stage | Output | Verification |
|---|---|---|
| 1 | Sequence analysis — ubiquitin, 76 aa | Reference input for Stages 2–3 |
| 2 | Structure analysis — ubiquitin (AlphaFold Server), pLDDT / pTM | Analysis artifacts in results/ |
| 3 | Residue-level features — 76 residues × 28 columns | 20/20 tests pass |
| 4 | Dataset manifest — 7 candidate proteins, 6 functional categories | 12/12 tests pass |
| 5 | Controlled download + validation — 14 files (7 CIF + 7 PAE JSON, AFDB v6) | 16/16 tests pass |
| 6 | Multi-protein dataset — 1,098 residues × 29 columns | 24/24 tests pass — review pending |
| 7 | Machine-learning experiments | Not started |
What this project is not
Stated plainly, because the distinction matters: this is currently a protein structure analysis and dataset engineering project. It is not a finished drug-discovery system. Specifically, none of the following exist yet, and none are claimed:
- No trained predictive model — no training, no train/test split, no evaluation metrics
- No molecular docking, molecular dynamics, or compound screening
- No wet-lab or experimental validation
- No comparison against experimental (PDB) structures — the dataset stores predictions only
- No biological hypothesis, label, or conclusion drawn from the predictions
Known limitations
- Small, narrow dataset. Seven single-chain proteins (five human, two non-human) form a pipeline-validation set — not a basis for biological generalisation. Residues within one protein are strongly correlated, so 1,098 rows are not 1,098 independent samples.
- Prediction-only ground truth. The dataset holds AlphaFold predictions with no attached experimental reference structures.
- Single-chain only. No complexes, no ligands, no post-translational modifications.
-
Two columns excluded by design. The Stage 3 schema includes
contact_prob_sumandn_high_conf_contacts, which depend on a contact-probability matrix that the public AlphaFold DB downloads do not provide. Both columns were excluded rather than imputed or approximated, and the exclusion is documented in the dataset schema.
Note on confidence scores. pLDDT, PAE, and pTM measure a model's internal confidence — not biological truth. A high-confidence prediction can still be wrong, and verifying a structure requires real experiments such as X-ray crystallography, NMR, or cryo-EM. Nothing in this project should be read as experimentally validated.
The repository's own code does not declare a licence yet. Structure predictions come from the AlphaFold Protein Structure Database (v6), released under CC BY 4.0, with attribution to EMBL-EBI, Google DeepMind, and UniProt.