ACADEMIC SOFTWARE · OPEN SCIENCE

Academic Software & Open Science

Open-source deep learning frameworks and curated datasets driving biomolecular discovery and computational biomedicine.

01 —— ACADEMIC SOFTWARE & DEEP LEARNING FRAMEWORKS

Academic Software & Open Science

Open-source deep learning frameworks and curated datasets driving biomolecular discovery and computational biomedicine.

2 OPEN PACKAGES · FAIR DATASETS

Open-Source Packages

SynProtX & GraphEGFR Python AI tools

100%

Open Science & Code

Public GitHub repos & Colab execution

FAIR

Zenodo Datasets

Citable DOIs for weights & multi-omics data

Top Tier

Peer-Reviewed Venues

Published in GigaScience & JCC

Synergy PredictionPillar 01 · Representation

GigaScience (2025)

SynProtX

Large-Scale Proteomics & Graph Neural Networks for Synergistic Drug Combinations

Figure: SynProtX-GATFP (GATConv + Fingerprints + Multi-Omics Expression Layers)

SynProtX is a deep learning model leveraging large-scale proteomics, molecular graphs, and fingerprints to enhance the prediction of synergistic effects in anti-cancer drug combinations.

Integrates large-scale clinical & cell line proteomics (ProCan-DepMapSanger) and cancer gene expression (DepMap CCLE)
Graph attention network architecture (SynProtX-GATFP) fusing molecular graphs and chemical fingerprints
Validated across tissue-specific (Breast, Lung, Ovary, Skin) and large-scale synergy benchmarks (ALMANAC, FRIEDMAN, ONEIL)
Open-source implementation with reproducible training pipelines and Zenodo-archived model weights

Cancer Drug CombinationDeep LearningDrug DiscoveryGraph Neural NetworksMulti-omicsProteomics

GitHub Zenodo Paper

Bioactivity PredictionPillar 01 · Representation

Journal of Computational Chemistry (2024)

GraphEGFR

Molecular Graph Attention & Hybrid Fingerprints for EGFR Bioactivity Prediction

Figure: Attentive FP (Graph Attention) + CNN (Convolutional Layers) + DNN (Fully Connected Layers)

GraphEGFR is a deep learning model specifically designed to enhance molecular representation for the prediction of inhibitor bioactivity (pIC50) against wild-type HER1, HER2, HER4, and mutant HER1 proteins.

Hybrid multi-representation: Attentive FP graph attention, CNN for non-hashed fingerprints, and DNN for hashed fingerprints
Predicts bioactivity (pIC50) across wild-type HER1, HER2, HER4, and clinically vital HER1 mutants (T790M, L858R, delE746-A750)
Multi-task and transfer learning architecture specifically optimized for small-molecule bioactivity under data scarcity
Google Colab ready, complete with pre-trained checkpoints and curated benchmark datasets on Zenodo

EGFR KinaseGraph Attention MechanismMulti-task LearningTransfer LearningBioactivity Prediction (pIC50)

GitHub Zenodo Paper

02 —— OPEN SCIENCE & REPRODUCIBILITY COMMITMENT

Transparency, Reproducibility & FAIR Principles

Ensuring that all deep learning models, data representations, and curated benchmarks are accessible, verifiable, and freely available for the global research community.

OPEN SOURCE · FAIR DATA

Open-Source Architecture

Our codebases are publicly hosted on GitHub under open licenses (MIT/BSD), featuring modular Python architectures built on PyTorch, PyG (PyTorch Geometric), and RDKit.

FAIR Data & Zenodo DOIs

All pre-trained weights, hyperparameter checkpoints, and curated benchmark datasets are permanently archived with citable Zenodo DOIs to guarantee long-term data preservation.

Reproducible Pipelines

Step-by-step setup guides, deterministic seeding routines, and Google Colab execution notebooks enable immediate experimentation without local hardware hurdles.

COLLABORATIVE RESEARCH

Interested in utilizing our models or benchmarking datasets?

Explore our full catalog of peer-reviewed publications or reach out for academic collaboration and joint research initiatives.

Publications Get in Touch