We will deploy a proprietary deep learning-based framework to rapidly screen multi-billion small molecule libraries. The performance of the proposed framework is tested on several curated as well as publicly-available unbiased benchmarking datasets. To demonstrate the actual application of the framework, we have screened 1.37 billion molecules to discover new inhibitors of the epigenetic protein BRD9 bromodomain. We have identified and prioritized 17 molecules for in vitro testing, four of which were active, and three of which were chemically distinct from known binders. A novel, first-in-class hit PS-902 that demonstrated an IC50 of ~1.0 µM was discovered.

The proposed framework utilizes a lightweight convolutional neural network architecture highly optimized for handling low-level molecular characteristics capable of discovering novel chemotypes while screening one billion commercial, synthesizable libraries (e.g. ENAMINE Real) per day for both specific and multi-targeted hits identification. The successful validation is already done on one target. We are currently also evaluating the performance of the framework on multiple other targets in collaboration with different partners. 

For the CACHE #3 challenge, we are preparing to screen the current (November 2022, 1st week) ENAMINE Real Database with over 5.5 billion molecules. 

Maestro (protein preparation)
Glide for docking/hits prioritization (but we'll make a decision later whether to use it or Autodock4/Autpodock Vina or SMINA)

Python, TensorFlow, Keras, RDkit, Autodock4/Autdock Vina/SMINA, Gromacs (if required), PyMOL

We have not published the work yet as the validation studies are currently undergoing. 

Merged selection will be rather done by combined use of molecular docking and MD simulations. 

The approach used for merged selection with fewer hundred molecules will not be our proprietary approach rather we will be applying the established methods to make the rank selections. We will be focussing more on docking and dynamics calculations. 

We may use Desmond from Schrödinger but at this point, we're not sure as we may also use the open-source package Gromacs. The decision will rather be based on the output of the docking step and based on how many top hits we receive, we may have to select between Desmond or Gromacs.

Autodock4, Gromacs

Noncanonical binding of Lck to CD3ε promotes TCR signaling and CAR function
New insights into the structural dynamics of the kinase JNK3
4-Acyl Pyrrole Capped HDAC Inhibitors: A New Scaffold for Hybrid Inhibitors of BET Proteins and Histone Deacetylases as Antileukemia Drug Leads
Hyaluronic acid–GPRC5C signalling promotes dormancy in haematopoietic stem cells
 

The approach used for hit-to-lead optimization will largely use molecular docking, MD simulations and free energy calculations/perturbations.

The approach used for hit-to-lead optimization will largely use molecular docking, MD simulations and free energy calculations/perturbations. It will not be a proprietary approach but rather a combination of established methods in the field of hit optimization. 

We may use Desmond from Schrödinger but at this point, we are not sure as we may also use the open-source package Gromacs. 

Autodock 4, Gromacs

4-Acyl Pyrrole Capped HDAC Inhibitors: A New Scaffold for Hybrid Inhibitors of BET Proteins and Histone Deacetylases as Antileukemia Drug Leads