← Portfolyoya Dön

DEEPScreen2 (Undergraduate Researcher)

Karmaşık moleküler tanımlayıcılara gerek kalmadan, 2D molekül görüntüleri üzerinde CNN, Vision Transformer ve YOLO omurgalarıyla ilaç-hedef etkileşimi tahmini yapan; veri toplama, eğitim ve yorumlanabilirlik adımlarını tek bir uçtan uca pipeline'da birleştiren açık kaynaklı bir derin öğrenme çerçevesi.

PythonPytorchCNNYOLOChEMBL APIMoleculeNetTherapeutics Data Commons (TDC)Muon & AdamW optimizerVision Transformer (ViT)
Computational models for drug–target interaction (DTI) prediction aim to reduce the cost of experimental screening, but often require complex molecular representations and long development cycles. Here we present DEEPScreen2, a modular, open-source framework that performs DTI prediction directly from RDKit-generated 2D molecular images, expanded with augmentation, using convolutional neural network and vision transformer backbones. DEEPScreen2 streamlines data curation, image generation, training and inference into an end-to-end pipeline that can be completed within seven days for a new target on standard hardware. Prospective applications to de novo design and drug repurposing identified two experimentally confirmed active molecules and two repurposed drugs, respectively. DEEPScreen2 thus demonstrates that simple image-based models can match or exceed traditional descriptor-based approaches for high-accuracy DTI prediction and rapid hypothesis generation.