MorphoCLIP

Cross-modal contrastive learning aligning molecular graphs with Cell Painting morphological profiles for zero-shot mechanism-of-action retrieval.

Overview

MorphoCLIP trains a CLIP-style dual encoder that maps molecular structures and Cell Painting CellProfiler feature vectors into a shared embedding space. Once trained, you can query the model with any SMILES string to retrieve phenotypically similar compounds — without any MoA label at inference time.

Key novelty

• First CLIP-style alignment between molecular graph encodings and Cell Painting profiles
• Hard negative mining strategy using known off-target MoA pairs
• Zero-shot MoA retrieval benchmark on JUMP-CP × ChEMBL

Installation

cd morphoclip
pip install -r requirements.txt

Data

python scripts/download_data.py
python scripts/preprocess.py

Training

python scripts/train.py --config configs/default.yaml

Evaluation

python scripts/evaluate.py --checkpoint checkpoints/best_model.pt

Tests

pytest tests/ -v

Architecture

SMILES ──► GATv2 (4 layers) ──► mean pool ──► projection ──► L2-norm ──► 256-d embedding
                                                                                │
                                                                    cosine similarity + InfoNCE
                                                                                │
812-d CellProfiler ──► Residual MLP ──► projection ──► L2-norm ──► 256-d embedding