Deep Drug

    The Power of AI to Improve World Health

    The power of AI to improve world health

    The Work

    Deep Drug is pursuing the development of computer aided drug design software based on exhaustive graph-based search algorithms combined with machine learning-based filters to synthesize new compounds by connecting building blocks of molecules following their connectivity patterns. The software under development will automatically synthesize targeted drug molecules, filter candidates based on chemical criteria (such as being an antibiotic), toxicity, etc., involves the analysis of 3D image models  of the pathogen for possible drug repurposing, automates clinical testing for side effects,  and  predict the  candidates that are most likely to succeed. The software Deep Drug is developing is known as eSynth. The software is modular in nature being currently composed of eSynth, eMolFrag, eToxPred and several AI based filters and engines.

    eSynth creates new molecules. The building blocks from which eSynth constructs molecules are:  Rigids -  inflexible fragments often a single or fused aromatic group and Linkers - flexible fragments connecting rigid blocks. eSynth rapidly generates series of compounds with diverse chemical scaffolds complying with Lipinski’s criteria for drug-likeness (No more than 5 hydrogen bond donors, Not more than 10 hydrogen bond acceptors, A molecular mass less than 500 daltons, An octanol-water partition coefficient logP not greater than 5). Although, these molecules may have different physicochemical properties, the initial fragments are procured from biologically active and synthetically feasible compounds.

    eMolFrag decomposes organic compounds into non-redundant fragments retaining molecular connectivity patterns. Given a collection of molecules, eMolFrag generates a set of unique fragments comprising rigids with minimal torsional angles and flexible linkers. These building blocks can be subsequently used to investigate the chemical composition of small molecule libraries, to identify common fragments across sets of ligands, and to prepare virtual screening libraries for targeted drug discovery. The application of eMolFrag in recombinant drug design has been demonstrated for the adenosine receptor and coagulation factor X.

    eToxPred is a machine learning based engine that will take the molecular structure of a drug-like compound and predict a side effect score. eToxPred is integrated to the eSynth synthesizer that will eliminate any synthesized molecule with a score above a certain threshold. The FDA’s open source database is used to train the learning engine and determine the threshold.  Even after pharmaceutical companies spend years and billions of dollars in creating a new drug; it is often the case that the drug has undesirable side-effects that renders it unusable. The application of eToxPred provides an early stage analysis of success/failure for a given compound.

    Synthetic Accessibility Analysis:  Natural products are a source of ingredients for many drugs. Some of these natural products are hard to acquire. The synthesis accessibility score (SAscore) is a metric that can measure the difficulty to synthesis a drug-like molecule. A molecule’s SAscore is between 1 (easy to synthesis) to 10 (difficult to synthesis). It is calculated based on molecular complexity and fragment contributions. We have developed a deep learning algorithm that, from the molecular structure of a natural product, can predict its synthetic accessibility score.  For compounds with high scores, it is possible to synthesize them using eSynth and analyze their side-effects. We already have access to a freely available database of natural products for providing ground truths to train the machine learning engine.


    1. Misagh Naderi, Chris Alvin, Yun Ding, Supratik Mukhopodhyay, Michal Brylinski:

        A graph based approach to construct target focused libraries for virtual screening, J. Cheminformatics 8(1)

        14:16 (2016)

    2. Tairan Liu, Chris Alvin, Yun Ding, Supratik Mukhopodhyay, Michal Brylinski:

        Break Down in Order to Build Up: Decomposing Small Molecules for Fragment Based Drug Design with              eMolFrag. Journal of Chemical Information and Modeling 57(4): 627-631 (2017)

    3. Limeng Pu, Misagh Naderi,Tairan Liu,  Hsiao-Chun Wu, Supratik Mukhopodhyay, Michal Brylinski:

      eToxPred: A machine learning based approach to estimate the toxicity of drug candidates. BMC Pharmacology    vol. 20(1) 2. (2019)

    Team Leaders


    The Deep Drug team is currently competing for the IBM Watson AI XPrize and have been among the top 10 teams in the world, being nominated for milestone awards for past two years (2017 and 2018) in a row. Please see the following websites for further details: and We were also invited to present at the NeuriPS competition track.


    Heading 1