Skip to content

LLMs-Pharmaceutical

LLM Applications for Cancer Drug Discovery and Oncology Clinical Trials

License Version Python DOI


Collection of Manuscripts 01-20 strengths, limitations, and results based on LLM-driven oncology research by ChemicalQDevice. 50+ Abstracts and Diagrams. Industry Guidance.


Overview

ChemicalQDevice develops AI applications for oncology clinical trials and pharmaceutical research. Anthropic, Google, OpenAI, and xAI large language models (LLMs) are utilized to generate machine learning training code, clinical trial simulations, drug discovery pipelines, and multi-model research workflows.

This repository contains 2,300+ files across three primary directories spanning LLM/AI code notebooks, published manuscripts, and R&D documentation.

Metric Value
Jupyter Notebooks 1,446
Python Scripts 35
Published Manuscripts 17+ (with DOIs)
R&D Documents 70
Research Directories 28
Repository Size 801 MB

Repository Structure

LLMs-Pharmaceutical/
├── Code/                          # 2,235 files — Notebooks, scripts, and research code
│   ├── Digital_Twin_PDAC/         #   10 notebooks — PDAC digital twin clinical trial proposals
│   ├── Drug Discovery/            #  290 files     — 14 subdirectories: LLM, RAG, agentic, multi-model
│   │   ├── Agentic-LLM/           #    4 notebooks — Agentic LLM drug discovery workflows
│   │   ├── Multi-LLM/             #   15 notebooks — Multi-LLM comparisons, meta-analyses, reports
│   │   ├── Quad-LLM/              #   14 notebooks — Quad-LLM evaluation pipelines
│   │   ├── ProtBert/              #    1 notebook  — Protein language model experiments
│   │   ├── ProtGPT2/              #    2 notebooks — Protein generation models
│   │   ├── Llama-3 RAG/           #    2 notebooks — Llama-3 retrieval-augmented generation
│   │   ├── Meta-Llama-3/          #    6 notebooks — Meta Llama-3 fine-tuning and inference
│   │   └── ...                    #   (LLM, LLM-RAG, LMM, XLLM, pLDM, iPhone 15 Pro)
│   ├── Hugging Face/              #   15 notebooks — Llama-3 fine-tuning, 32-pipeline benchmark
│   ├── LangChain/                 #    9 notebooks — Agent Supervisor, Code Assistant, RAPTOR
│   ├── Generative AI Live/        #    6 notebooks — Live coding: RAG, RAPTOR, Agent, fine-tuning
│   ├── Groq/                      #    1 notebook  — GroqCloud inference benchmarks (4 models)
│   ├── Open WebUI/                #              — Ollama/Docker deployment studies
│   ├── Python/                    #   35 scripts  — Python tutorial series (Classes → Logging)
│   ├── C++/                       #   47 files    — C++ tutorial source files
│   ├── Apple/                     #              — LLM Farm: MacBook Pro, iPad Pro, iPhone 15 Pro
│   ├── PennyLane/                 #  488 notebooks — QML algorithms and benchmarking
│   ├── Qiskit/                    #   18 notebooks — Quantum programming and device benchmarks
│   └── ...                        #   (QML and quantum parallel algorithm directories)
│
├── Manuscripts/                   # Published research — 17+ papers with DOIs
│   ├── README.md                  #   67 KB — Full abstracts, citations, and DOI badges
│   └── LICENSE.md                 #   CC BY 4.0
│
├── README.md                      # This file
├── LICENSE.md                     # Apache License 2.0
├── CHANGELOG.md                   # Version history
├── CONTRIBUTING.md                # Contribution guidelines
├── CODE_OF_CONDUCT.md             # Community standards
├── SECURITY.md                    # Security policy
├── SUPPORT.md                     # Getting help
├── CITATION.cff                   # Citation metadata
├── ruff.toml                      # Linter configuration
├── .gitignore                     # Git ignore rules
├── .github/                       # CI workflows and templates
├── Daraxonrasib Efficient LLM Trial Simulations.pdf
├── LaTeX LLM Oncology Table Prompting Guide.pdf
└── Scaling Oncology LLM Generated Code Jan 2026.pdf

Manuscripts

The Manuscripts/ directory contains the full abstracts, citations, and DOI links for all published work. Key publications are summarized below in reverse chronological order.

2025–2026 LLM and Clinical Trial Studies

Date Title DOI
Dec 2025 Code Generation Competition: 16 Proprietary vs. Open-Source LLMs & Iterative Learning Based on FDA FAERS 10.5281/zenodo.18029100
Nov 2025 AI Peer Review Acceleration of LLM-Generated Glioblastoma Clinical Trial Patient Matching ML 10.5281/zenodo.17774560
Nov 2025 LLM-Generated Glioblastoma Drug Synergy Machine Learning 10.5281/zenodo.17614396
Oct 2025 End-to-End Oncology Clinical Trial LLM Efficiency For Industry Adoption 10.5281/zenodo.17451709
Sep 2025 Accelerating FDA Compliance via AI Digital Twin Pancreatic Cancer Simulation 10.5281/zenodo.17239510
Aug 2025 QSP Metastatic Pancreatic Cancer AI Clinical Trial Simulation 10.5281/zenodo.17001137
Jul 2025 ChatGPT 100,000 Patient 24-Month In Silico Phase III Clinical Trial 10.5281/zenodo.16415815
Jun 2025 End-to-End PDAC Digital Twin Clinical Trial Proposals 10.5281/zenodo.15735068
May 2025 10 Year Glioblastoma Clinical Trial Meta-Analyses by Autonomous AI 10.5281/zenodo.15549831
Apr 2025 AI Revolution Toward the Cure of Lung Adenocarcinoma 10.5281/zenodo.15278152
Mar 2025 Autonomous LLM Agent and Scalable Reasoning LLM for Cancer Drug Cost Solutions 10.5281/zenodo.15072843
Feb 2025 Cost Containment of Global mAb Drugs and Cancer Clinical Trials via LLM Reasoning 10.5281/zenodo.14968404
Jan 2025 Clinical Decision Support Based on Bevacizumab Cancer Trials 10.5281/zenodo.14968162

2024 AI and Chemistry Studies

Date Title DOI
Dec 2024 Cancer vs. Conversational Artificial Intelligence 10.1101/2024.12.28.630597
Nov 2024 mAb Bioprocess Engineering In-Context Table Forecasts 10.26434/chemrxiv-2024-jzbj0
Oct 2024 Monoclonal Antibody Bioprocess Engineering Using Conversational AI 10.26434/chemrxiv-2024-3m7m1
Oct 2024 Paclitaxel Biosynthesis AI Breakthrough 10.26434/chemrxiv-2024-pqjd3

Full abstracts and citation details are available in Manuscripts/README.md.

Research Areas

Digital Twin PDAC — Clinical Trial Simulation

Code/Digital_Twin_PDAC/
├── Notebooks/
│   ├── Python_01_RP01.ipynb    ─── Research proposal report 1
│   ├── Python_02_RP02.ipynb    ─── Research proposal report 2
│   ├── Python_03–05_RP*.ipynb  ─── Continued proposal reports
│   ├── Python_06_RP06.ipynb    ─── Extended proposal analysis (5.4 MB)
│   ├── Python_07_Pro5Viz.ipynb ─── 5-proposal visualizations
│   ├── Python_08_Pro5Tab.ipynb ─── 5-proposal tabular data
│   ├── Python_09_ProOps4.ipynb ─── Opus 4 Extended proposal
│   └── Python_10_ProSon4.ipynb ─── Sonnet 4 Extended proposal
└── Images/
    ├── Proposals/
    └── Reports/

10 notebooks implementing a multi-AI pipeline for pancreatic ductal adenocarcinoma (PDAC) digital twin clinical trial proposals. Seven AI model combinations (o3re, g25p, son4, grk3, o3pr, ops4, o3ch) generated 40 meta-analyses, 6 reports, and 5 trial proposals scored up to 9.60/10.

Drug Discovery — Multi-Model LLM Pipelines

Code/Drug Discovery/
├── Agentic-LLM/           # 4 notebooks  — Agentic LLM workflows
├── Multi-LLM/             # 15 notebooks — Multi-model comparisons and meta-analyses
│   ├── Images/
│   ├── Notebooks/
│   ├── Reports/
│   └── Meta-Analyses/
├── Quad-LLM/              # 14 notebooks — Four-model evaluation pipelines
│   ├── Images/
│   └── Notebooks/
├── ProtBert/              # 1 notebook   — Protein language model
├── ProtGPT2/              # 2 notebooks  — Protein generation
├── Meta-Llama-3/          # 6 notebooks  — Llama-3 fine-tuning
├── Llama-3 RAG/           # 2 notebooks  — Retrieval-augmented generation
├── LLM/                   # LLM approach documentation
├── LLM-RAG/               # Drug synthesis RAG
├── LMM/                   # Large multimodal models
├── XLLM/                  # Precision medicine
├── pLDM/                  # 2 notebooks  — Protein latent diffusion
└── iPhone 15 Pro/         # On-device inference

290 files across 14 subdirectories covering LLM, agentic, RAG, multi-model, and protein language model approaches to drug discovery. Key studies include Groq inference benchmarks (~1000 tokens/sec), Hugging Face 32-pipeline evaluations, and LangChain agent workflows.

Hugging Face, LangChain, Generative AI

Code/Hugging Face/              # 15 notebooks — 32-pipeline benchmark, Llama-3 fine-tuning
Code/LangChain/                 #  9 notebooks — Agent Supervisor, Code Assistant, RAPTOR
Code/Generative AI Live/        #  6 notebooks — Live demos: RAG, RAPTOR, Agent, fine-tuning
Code/Groq/                      #  1 notebook  — GroqCloud 4-model inference comparison
Code/Open WebUI/                #              — Ollama/Docker GenAI deployment

Quantum-Inspired Machine Learning (Historical)

Earlier work focused on quantum-inspired machine learning (QiML) using PennyLane and Qiskit frameworks. This represents the largest portion of notebooks by count (600+) and covers algorithm prototyping, parallel quantum architectures, tensor networks, and device benchmarking.

Code/PennyLane/                            # 488 notebooks
Code/Qiskit/                               #  18 notebooks
Code/All PennyLane QML Demos/              #  26 notebooks
Code/All Qiskit ML Demos/                  #  13 notebooks
Code/All Qiskit, PennyLane QML Nov 23/     #  54 notebooks
Code/Efficiency Metrics for Parallel QML Algorithms/
Code/Parallel Quantum Algorithms torch.nn Seq, Mod/
Code/PyTorch, Keras, PL Parallel Quantum Algorithms/
Code/QML Parameters for Breakthrough Parallel Algorithms/
Code/Quantum Parallel Architectures Progression/
Code/Tensor Network vs FC Controllability/
Code/Tensor Network vs FC Explainability/
Code/Tensor Network vs Fully Connected Layer/
Code/Tensor Networks for Generative AI/

Research Timeline

2023 ──── Quantum ML foundations (PennyLane, Qiskit, parallel algorithms)
  │
2024 ──── LLM pivot: drug discovery, spectrometric analysis, bioprocess engineering
  │         ├── Multi-LLM and Quad-LLM evaluation frameworks
  │         ├── Protein language models (ProtBert, ProtGPT2)
  │         ├── Hugging Face 32-pipeline benchmark
  │         └── LangChain agents and RAG pipelines
  │
2025 ──── Clinical trial scale-up and regulatory workflows
  │         ├── Digital Twin PDAC (7 AI models, 40 meta-analyses)
  │         ├── 100K patient in silico Phase III trials
  │         ├── QSP pancreatic cancer simulations (250 ODEs/patient)
  │         ├── End-to-end oncology trial LLM efficiency (52-73% faster, 24-50% less cost)
  │         ├── Glioblastoma drug synergy ML (accuracy 0.9804)
  │         ├── AI peer review workflows
  │         └── 16-LLM code generation competition on FDA FAERS
  │
2026 ──── Scaling and methodology
            ├── Scaling Oncology LLM Generated Code
            ├── Daraxonrasib Efficient LLM Trial Simulations
            └── LaTeX LLM Oncology Table Prompting Guide

Key Results

  • Trial time reduction: 52–73% faster with LLM-assisted workflows (Oct 2025)
  • Trial cost reduction: 24–50% less expensive; 99.6–99.9997% cost reduction vs. in-person trials
  • Per-patient cost: $3.6 (in silico) vs. $59,500 (traditional) — a 16,528x difference
  • Drug synergy prediction: 0.9804 accuracy, 0.9705 Macro-F1 on glioblastoma drug pairs (Nov 2025)
  • In silico scale: 100,000 patients, 24-month, 5-arm Phase III triplicate (Jul 2025)
  • AI experiment speed: Core experiments completed in 3.1–5.4 hours (May 2025)
  • Proposal scores: Up to 9.60/10 for PDAC digital twin trial proposals (Jun 2025)

License

Citation

@misc{kawchak2026llmspharmaceutical,
  author    = {Kawchak, Kevin},
  title     = {LLMs-Pharmaceutical: LLM Applications for Cancer Drug Discovery and Oncology Clinical Trials},
  year      = {2026},
  publisher = {GitHub},
  url       = {https://github.com/kevinkawchak/LLMs-Pharmaceutical},
  doi       = {10.5281/zenodo.13273141}
}

Author

Kevin Kawchak — CEO, ChemicalQDevice, San Diego, CA kevink@chemicalqdevice.com

About

Oncology Trial Innovation

Resources

Code of conduct

Contributing

Security policy

Stars

80 stars

Watchers

4 watching

Forks

Releases

Contributors

Languages