OmniScreen-AI

OmniScreen-AI: A Multi-Modal High-Throughput Drug Screening & Dynamic Verification Platform for PD-L1

A full-modality (small molecules / macrocyclic peptides / nucleic acids) high-throughput screening and dynamic kinetics verification platform for the tumor immunotherapy target PD-L1.

Project Highlights

• Three-modality parallel funnel: chemical space (small molecules), sequence space (antibodies/peptides), RNA space (siRNA/Aptamer)
• Four-stage closed-loop validation: AI pre-screening → structural docking → MD kinetics → free energy calculation
• Notebook-driven: all logic lives in 3 workflow notebooks, modular blocks, one-click Colab reproduction
• Tiered compute: Colab CPU (pre-screen + docking) → RunPod GPU (MD + free energy)

Repository Structure

OmniScreen-AI/
├── notebooks/
│   ├── OmniScreen_SM_Workflow.ipynb          # Small-molecule full pipeline (ChEMBL real compound library)
│   ├── OmniScreen_PE_Workflow.ipynb   # Protein/peptide full pipeline
│   └── OmniScreen_NA_Workflow.ipynb   # Nucleic-acid drug full pipeline
├── data/
│   ├── receptor/          # PD-L1 receptor PDB (5N2F, 4ZQK, etc.)
│   ├── raw_libraries/     # Initial candidate libraries (SMILES / FASTA / mRNA)
│   └── screened_results/  # Stage-wise screening outputs
├── docker/                # Optional: standardized runtime environment
├── pyproject.toml
└── requirements.txt

Quick Start (Cursor + Colab)

1. Clone this repo and open it in Cursor
2. Install the Google Colab and Notebook MCP extensions
3. Configure GitHub Token (required for pushing data):
   • Create a PAT at GitHub Settings → Tokens (check repo scope)
   • Cursor + Colab extension: run the "Configure GitHub Token" cell in the notebook and paste the token manually (Colab Secrets only work in the web UI; the Cursor kernel cannot read them)
   • Colab web UI: add GITHUB_TOKEN in 🔑 Secrets; code will read it automatically
4. Open notebooks/OmniScreen_SM_Workflow.ipynb and select the Colab kernel
5. Run cells in module order (Module 0 → Module 6)

Compute tips: Modules 1–3 run fine on Colab free-tier CPU/GPU; Modules 4–5 (OpenMM / free energy) are best on GPU platforms like RunPod — see in-notebook notes.

Data Persistence Strategy (Colab + Local Auto-Sync)

Colab runs in the cloud and cannot write directly to local disk. We use "cloud compute → pack outputs → Agent writes back locally":

Colab runs modules
  → export_for_local_sync() base64-encodes data/ files into cell output
  → Cursor Agent parses sync markers
  → Writes automatically to local /Users/schmit/Documents/OmniScreen-AI/data/

GitHub push is optional backup (needs token), not the primary sync path.

File type	Local data/	GitHub
*.csv, *.smi, *.png, *.pdb	✅ Agent auto-write	Optional push
*.dcd, *.xtc (MD trajectories)	❌ Too large	External storage

Three Pipeline Overview

Notebook	Modality	Core toolchain	Docs
OmniScreen_SM_Workflow	Small molecules	RDKit, AutoDock Vina, OpenMM, MM/PBSA	SM modules
OmniScreen_PE_Workflow	Protein/peptide	ESM-2, HDOCKlite, AlphaFold 3	PE modules
OmniScreen_NA_Workflow	Nucleic acids	ViennaRNA, Bowtie2, AlphaFold 3	NA modules

Data Notes

• Screening CSVs, chart PNGs, small SMILES libraries, receptor PDB → auto-commit to GitHub
• MD trajectories (.dcd), oversized compound libraries → not in Git; see data/*/README.md

Limitations & False Negatives

The core goal of computer-aided drug design (AIDD) is not "100% recall" but to shrink the candidate space with affordable compute among millions of molecules. Discarding good molecules due to software errors, misconfiguration, or algorithmic approximations is called a false negative in industry. The OmniScreen-AI pipeline has the same trade-off, mainly at Modules 2–3.

Bottleneck 1: Ligand preparation failure (ligand_prep_failed)

Cause: rigid software assumptions and capability limits.

During ligand prep, tools infer protonation state, charge, and rotatable bonds at physiological pH (~7.4). Novel molecules with rare heterocycles or complex stereochemistry may lack precedents in classical topology tools (MGLTools, Open Babel, etc.), causing charge or parameter assignment to fail.

In this project, Module 3 uses RDKit 3D embedding + Open Babel → PDBQT; on failure the molecule never enters Vina scoring and is marked ligand_prep_failed in docking_scores.csv — promising drugs may be silently dropped here.

Bottleneck 2: Vina docking failure or low scores (vina_failed / low-score elimination)

Cause: rigid receptor "blind probing" and compute compromises.

AutoDock Vina uses rigid receptor docking for throughput: PD-L1 is a fixed "lock" and only the ligand rotates. Real proteins are dynamic:

• Induced fit: some actives bind when the pocket is still small; static docking fails or scores poorly (e.g. −4.0 kcal/mol); given time, the ligand can open the pocket and stabilize. Vina's single static sample misses this.
• Insufficient search: low default exhaustiveness may miss reasonable poses in complex pockets, causing failure or artificially low scores.

The current OmniScreen SM flow docks only the top 250 after pre-screen (SM_CONFIG["max_dock"]), increasing false-negative risk for lower-ranked molecules.

Common industry mitigations

Strategy	Idea	Relation to this project
Large funnel throughput	Don't over-tighten Module 2; let thousands dock; use scale against rule-based misses	Current library ~2,300; ~864 pass pre-screen; widen Lipinski or max_dock
Deep-learning docking in parallel	DiffDock, Uni-Mol, etc. avoid classical charge params; recover Vina failures	🔜 Planned as Module 3 supplement
Higher search precision	Raise Vina exhaustiveness from 8 to 16–32	Adjustable in Module 3; CPU time grows
Dynamic re-validation	MD / MM-PBSA for induced fit and binding stability	Modules 4–5 (RunPod GPU) purpose

Design philosophy

Better to quickly filter a huge pool than stall compute at step one — but survivors still need dynamic "appeal."

OmniScreen uses Colab funnel (Modules 0–3) + RunPod kinetics (Modules 4–5): speed and scale first, then validation of static docking uncertainty. Docking scores and pre-screen labels are ranking hints, not final druggability verdicts.

See SM modules — limitations & assumptions for module-level detail.

License

MIT — see LICENSE