Protein Plant Synthesis

If you had to produce a therapeutic protein in a plant, which host and which DNA design should you try first?

A leakage-controlled surrogate ML pipeline that designs DNA for therapeutic proteins across four plant hosts, and explains its own predictions.

4 hostsArabidopsis, rice, maize and date palm, compared under protein-grouped evaluation

Year: 2025–2026
Status: Working pipeline + dashboard · simulated target, not wet-lab data · 26 tests
Stack: Python, scikit-learn, XGBoost, LightGBM, SHAP, pandas, Streamlit, Plotly, py3Dmol
Source: MoallaMelek/AI-Driven-Therapeutic-Protein-Plant-Synthesis

From UniProt to a ranked decision

Proteins are fetched from UniProt and quality-checked. Each one gets a naive back-translation and host-aware optimised DNA for four plant hosts, using host-biased codon tables. Structure descriptors come from AlphaFold, with a deterministic fallback so the app still runs offline.

Codon adaptation (CAI), GC compatibility, rare-codon ratios and protein features feed a set of regressors. The final screen turns predictions into an operations view: which host to try, at what estimated cost and time.

System architecture: data collection, DNA generation and feature engineering, proxy target construction

Making the evaluation credible

Grouped splits

The same protein appears once per host. Random splits would leak an accession into both train and test, so splits are grouped by protein (GroupShuffleSplit, GroupKFold).

Beat the trivial model first

Every model is compared against a dummy baseline that predicts the mean. The dashboard asks the question explicitly: is this better than a trivial prediction?

Explain, then trust

Permutation importance, local SHAP explanations and residual analysis are part of the product, not an afterthought.

Validation RMSE of each model against the dummy baseline — Model lab: every regressor against the dummy baseline.

Local SHAP explanation for a single prediction — Local SHAP: what pushed this prediction up or down.

Scatter of codon adaptation index against the proxy target, coloured by host — Dataset analysis: CAI against the proxy target by host.

Production ranking dashboard listing hosts by score, cost and time — Decision view: hosts ranked by efficiency, cost and time proxies.

One more section for engineers: architecture, hyper-parameters and design decisions.

Under the hood

Models: Dummy baseline, Linear, Ridge, Random Forest (350 trees, depth 14), XGBoost (350, depth 6, lr 0.05), optional LightGBM
Features: CAI, GC content, codon diversity, rare-codon ratio, protein length, hydrophobicity, structure-derived descriptors
Evaluation: Protein-grouped holdout, RMSE / MAE / R², ablations
Explainability: Permutation importance, SHAP, residual diagnostics
Product: Multi-tab Streamlit with Plotly charts and py3Dmol structure views
Tests: 26 unittest cases: training protocol safety, artifact loading, dashboard helpers, codon analysis

What it doesn't do (yet)

The expression target is a deterministic simulated proxy built from codon, host and structure signals. It is explicitly not wet-lab measured.
It ships as a CLI and Streamlit app; the REST service layer is planned, not built.

Read the code and the full results on GitHub