Quantum-AI Drug Discovery Framework
Budget: ₹100 – ₹400 INR
I am building a research-grade software framework that accelerates novel drug discovery by blending quantum AI techniques with state-of-the-art deep-learning workflows in PyTorch. The core idea is to create a virtual “multi-agent” laboratory where molecular data flows seamlessly—from raw files, through cleaning and integration, into quantum-enhanced models—so chemists can test hypotheses, explore compound libraries and predict activity or ADMET properties in minutes instead of months.
Technology focus
Python and PyTorch form the backbone for model development and experimentation, while Qiskit provides the quantum circuits I want to couple with classical neural nets for hybrid training loops. RDKit and DeepChem handle cheminformatics tasks, HuggingFace hosts transformer checkpoints, and PostgreSQL plus Spark ensure the data lake remains searchable and fast. A lightweight C++/Rust micro-service may be spun up for any performance-critical kernels, and a React / Node.js front-end will eventually surface results to scientists.
Data & pipeline needs
• Parse and normalise diverse chemical file formats (SMILES, SDF, PDB) into a unified PostgreSQL schema.
• Implement feature generation (fingerprints, graph embeddings) and on-the-fly augmentation.
• Orchestrate ETL in Spark so large vendor datasets can be mined, de-duplicated and versioned.
AI & quantum modelling
The first milestone centres on building a PyTorch-based model zoo (GNNs, diffusion models, transformer predictors) that can optionally swap in quantum layers written in Qiskit. Automatic hyper-parameter sweeps, experiment tracking and checkpointing are essential. Later milestones will introduce multi-agent reinforcement learners that propose novel scaffolds, rank candidates and feed discoveries back to the central datastore.
Deliverables for the initial phase
1. Reproducible environment (conda or Docker) with all dependencies pinned.
2. Data ingestion scripts plus a populated PostgreSQL schema using a provided sample set.
3. A working hybrid model: classical GNN + quantum variational circuit, trained on an activity dataset and benchmarked against a pure classical baseline.
4. Documentation and a short Jupyter notebook walkthrough that lets a scientist reproduce results on a GPU-equipped workstation.
Acceptance criteria
• End-to-end pipeline runs with one command and completes without manual intervention.
• Hybrid model achieves statistically significant improvement over baseline on the supplied test set.
• Code passes unit tests and follows PEP-8 where applicable.
If you have deep PyTorch experience and feel comfortable experimenting with Qiskit, this project should be both challenging and rewarding.
Technology focus
Python and PyTorch form the backbone for model development and experimentation, while Qiskit provides the quantum circuits I want to couple with classical neural nets for hybrid training loops. RDKit and DeepChem handle cheminformatics tasks, HuggingFace hosts transformer checkpoints, and PostgreSQL plus Spark ensure the data lake remains searchable and fast. A lightweight C++/Rust micro-service may be spun up for any performance-critical kernels, and a React / Node.js front-end will eventually surface results to scientists.
Data & pipeline needs
• Parse and normalise diverse chemical file formats (SMILES, SDF, PDB) into a unified PostgreSQL schema.
• Implement feature generation (fingerprints, graph embeddings) and on-the-fly augmentation.
• Orchestrate ETL in Spark so large vendor datasets can be mined, de-duplicated and versioned.
AI & quantum modelling
The first milestone centres on building a PyTorch-based model zoo (GNNs, diffusion models, transformer predictors) that can optionally swap in quantum layers written in Qiskit. Automatic hyper-parameter sweeps, experiment tracking and checkpointing are essential. Later milestones will introduce multi-agent reinforcement learners that propose novel scaffolds, rank candidates and feed discoveries back to the central datastore.
Deliverables for the initial phase
1. Reproducible environment (conda or Docker) with all dependencies pinned.
2. Data ingestion scripts plus a populated PostgreSQL schema using a provided sample set.
3. A working hybrid model: classical GNN + quantum variational circuit, trained on an activity dataset and benchmarked against a pure classical baseline.
4. Documentation and a short Jupyter notebook walkthrough that lets a scientist reproduce results on a GPU-equipped workstation.
Acceptance criteria
• End-to-end pipeline runs with one command and completes without manual intervention.
• Hybrid model achieves statistically significant improvement over baseline on the supplied test set.
• Code passes unit tests and follows PEP-8 where applicable.
If you have deep PyTorch experience and feel comfortable experimenting with Qiskit, this project should be both challenging and rewarding.