Design de novo biomolecules
with fine‑tuned language models

What it does

A single modality registry drives everything — add a biomolecule type in a few lines and the entire train → generate → evaluate → optimize flow just works.

Quickstart

Install, run the CPU smoke test to validate your setup, then fine-tune ProGen2 on your GPU.

# 1. Environment (Python 3.12)
python3.12 -m venv .venv && source .venv/bin/activate

# 2. PyTorch (CPU; for GPU use the cu128 index) + the package
pip install torch
pip install -e ".[chem]"          # adds RDKit + SELFIES

# 3. Offline smoke tests (no network / GPU needed)
denovo-mol pipeline -c configs/mol_flow_smoke.yaml
python scripts/make_tiny_local_model.py
denovo pipeline -c configs/smoke_local.yaml

# 4. Real fine-tune (downloads ProGen2 from Hugging Face)
denovo train    -c configs/progen2_protein.yaml
denovo generate -c configs/progen2_protein.yaml -n 500 -o generated/prot.txt
denovo evaluate -c configs/progen2_protein.yaml -i generated/prot.txt

# 1. Environment (Python 3.12)
python3.12 -m venv .venv && source .venv/bin/activate

# 2. PyTorch (Apple Silicon uses the MPS backend automatically)
pip install torch
pip install -e ".[chem]"

# 3. Offline smoke tests
denovo-mol pipeline -c configs/mol_flow_smoke.yaml
python scripts/closed_loop_demo.py

# 4. Real fine-tune (set train.fp16:false on Mac — use MPS defaults)
denovo train    -c configs/progen2_protein.yaml

# 1. Environment (Python 3.12, PowerShell)
py -3.12 -m venv .venv; .\.venv\Scripts\Activate.ps1

# 2. PyTorch (CPU; GPU = cu128 index + Python 3.12) + the package
pip install torch
pip install -e ".[chem]"

# 3. Offline smoke tests
denovo-mol pipeline -c configs\mol_flow_smoke.yaml
python scripts\make_tiny_local_model.py
denovo pipeline -c configs\smoke_local.yaml

# 4. Real fine-tune (downloads ProGen2 from Hugging Face)
denovo train    -c configs\progen2_protein.yaml

📖 Full platform-by-platform setup, GPU tips and troubleshooting live in RUN.md.

Choosing a model

On a 6 GB GPU, the recommended de novo model is ProGen2-small. The heavyweight foundation models (Evo 2, ESM-3) are best used via NVIDIA NIM / cloud.

Full feasibility analysis and NVIDIA NIM notes: docs/MODELS.md.

Benchmarks & results

Every generation run is scored on the four standard de novo metrics. The figures below are produced by scripts/benchmark.py and scripts/make_figures.py.

Measured results (GPT2-ZINC 87M, RDKit): de novo — 100% validity / 100% uniqueness / 100% novelty / 0.85 diversity on 1,000 molecules; property conditioning — mean QED 0.74 → 0.90; scaffold constraint — benzene containment steered from 57% to 100%; NVIDIA NIM (MolMIM) — aspirin (QED ≈ 0.55) optimized to QED 0.91 in the cloud. The loss and Bayesian-optimization charts remain illustrative until those runs.

Bayesian optimization

Sampling and training hyperparameters are tuned with Optuna's TPE sampler — a Bayesian method that finds strong settings in far fewer trials than grid or random search.

# Optimize decoding hyperparameters against validity + novelty + diversity
denovo optimize -c configs/progen2_protein.yaml \
    --mode sampling -m outputs/progen2_small --trials 25 \
    -o docs/results/bo_study.json

# Render the graphs below from the study
python scripts/make_figures.py --study docs/results/bo_study.json

Citation

If this project supports your research, please cite it.

@software{sanjay_denovo,
  author  = {Dr. Sanjay Anbu},
  title   = {De-Novo-LLM: Fine-tuning language models for de novo biomolecule generation},
  year    = {2026},
  url     = {https://github.com/sanjaydoc/De-Novo-LLM}
}