Researchers' Corner on ATOM's Science
|
|
Integrated Design Pilot Project with Neurocrine
|
|
|
In the fall of 2020, ATOM launched a histamine receptor (H1) antagonist project with Neurocrine Biosciences. H1-antihistamines are a H1-antihistamines are a class of drugs used to relieve allergic symptoms, but first- and second-generation H1-antihistamines have undesirable side effects, mainly due to off-target activities against muscarinic receptors M1-M5 and the hERG ion channel. ATOM has executed the first demonstration of integrated active learning to identify new selective H1 antagonists through application of the generative molecular design (GMD) process followed by synthesis and experimental validation of proposed compounds.
|
|
|
In this case study, ATOM applied the GMD active learning framework to generate and evaluate selective H1 antagonists, alternating computational design with experimental testing and model retraining.
ATOM trained machine learning models for H1, M2 and hERG binding using a combination of single-concentration data provided by Neurocrine for 1,900 compounds and multi-concentration pKi data from ChEMBL and GoStar. ATOM developed a unique hybrid model architecture to enable use of noisy single-point data to improve model accuracy in specialized regions of chemical space not covered by existing datasets, without requiring the expense of performing full-curve pKi measurements.
Because the initial measurements were done at a fairly high concentration, additional data were needed to distinguish the properties of compounds that are potent at lower therapeutically relevant doses. The ATOM team ran simulations with these models to identify a subset of 160 compounds for which those measurements would be most informative, and provided the list to Neurocrine for testing at a lower concentration. The resulting data were used to retrain the models, which were used to evaluate compounds in subsequent GMD runs.
Initial GMD runs revealed another challenge for the generative design process. Optimizing compounds for predicted H1 selectivity produced chemical structures that were far beyond the applicability domain of the models. The team addressed the issue by adding applicability domain constraints to the GMD optimization criteria, to keep the generated compounds within regions of chemical space where the models are reliable.
The GMD loop was run for 250 generations on an initial 25,000-compound set, producing ~1.75 million compounds over 60 hours of elapsed time. Over 200 compounds met all quality panel design criteria (efficacy, safety, and developability). The GMD loop continued to explore new parts of chemical space even after many generations.
The team selected the top scoring compounds, clustered them by structure, and sent the best cluster representatives to Neurocrine for evaluation. Neurocrine chose 150 compounds for synthesis or procurement and tested them. The initial results show that several of the proposed compounds bind strongly and selectively to the H1 receptor. After publishing their findings, the team plans to use the experimental results to retune models and complete at least one more GMD round.
|
|
|
|
Figure 2: Predicted pKi's for binding H1 vs M2 receptors for top scoring compounds proposed by GMD run.
|
|
|
|
Figure 3: Projection of chemical space covered by initial compounds (dark blue) and molecule populations in generations 100 (red) and 200 (green). The light blue points represent a reference set of randomly selected ChEMBL compounds.
|
|
|
Prepared by LLNL under Contract DE-AC52-07NA27344. LLNL-MI-827551
|
|
Academic Outreach Programs
|
|
|
LLNL and Purdue partner on drug discovery using ATOM-developed tools
Over the Fall 2020 and Spring 2021 semesters, Lawrence Livermore National Laboratory bioinformatics scientist and ATOM researcher Jonathan Allen, Ph.D., mentored a cohort of twenty students and two teaching assistants at Purdue University, introducing them to using machine learning and other computational tools for therapeutic drug design. The engagement was conducted through Purdue’s The Data Mine learning community, led by program director Mark Ward. The students evaluated virtual compounds for drug-like potential using ATOM-developed open source software and presented their technical work at The Data Mine Corporate Partners E-Symposium 2021. Their efforts also contributed to a new set of open source tutorials on data driven drug modeling. LLNL and Purdue are looking to extend the collaboration for another year, potentially adding more ATOM researchers.
|
|
|
|
Second annual ATOM summer training program empowers and equips students
|
|
|
Trainees, mentors, and supporting scientists and staff gathered virtually for the final presentations.
|
|
|
Six graduate students from Butler University and Nova Southeastern University just completed a virtual 10-week summer training program with the Accelerating Opportunities for Therapeutics in Medicine (ATOM) Consortium. Their efforts are the latest part of ATOM’s mission to put accelerated, computer-based pharmaceutical discovery in the hands of the scientific community. During the fast-paced program, the students learned to code and to create computer-based models to predict whether certain chemical compounds would lead to cell death or pass into the brain to possibly attack brain cancer cells. The method embraced both machine learning models and models based on fundamental mechanisms of biology. The students worked on their projects under the lead mentorship of Amanda Paulson, Ph.D., data science ATOM fellow at the Frederick National Laboratory, along with Susan Mertins, Ph.D., visiting scientist at Frederick National Laboratory and assistant professor of science at Mount St. Mary’s University.
“This year, we developed a set of interrelated projects. Each model is useful on its own. However, the models may be used in tandem to design new possible drugs that achieve multiple goals at once,” Paulson said. “For example, we will try to design drugs that not only pass into the brain, but also lead to cell death for cancer cells once they are in the right location.”
The students, five Pharm.D. candidates and one Ph.D. candidate, presented their final projects via web conference to scientists from Frederick National Laboratory, Lawrence Livermore National Laboratory, ATOM headquarters in San Francisco, Butler University, and Nova Southeastern University.
|
|
Meet our Technical Team Members: Belinda Akpa and Hiranmayi Ranganthan
|
|
|
Belinda Akpa from Oak Ridge National Laboratory (ORNL)
Belinda Akpa is a chemical engineer with a talent for tackling big challenges and fostering inclusivity and diversity in the next generation of scientists. She applies a problem-oriented approach that blends math, physics, chemistry, biology and computational modeling to accelerate solutions in areas including drug development and plant systems.
Her work at ORNL is part of the Accelerating Therapeutics for Opportunities in Medicine, or ATOM, consortium and focuses on modeling the complex interactions between candidate molecules and the human body to improve outcomes when potential drugs go to clinical trials. The results from Akpa’s systems models will integrate into ATOM’s larger computational framework, which aims to shorten the drug discovery timeline from five years to less than one year. A collaboration of national laboratories, academia and industry, ATOM is working to speed treatments for cancer and COVID-19.
“We’re trying to design and optimize candidate molecules without actually making the chemical compound in a lab,” Akpa said. “With computational models, we can explore whether a particular molecule will shrink a tumor or restore healthy cardiac function.”
Askey, K. (2021, April 30). Belinda Akpa: Engineering inclusive solutions. Oak Ridge National Laboratory.
|
|
|
Hiranmayi Ranganathan from Lawrence Livermore National Laboratory (LLNL)
Discovering the efficacy of new drugs has historically been a painstaking process, fraught with high failure rates. That’s why Hiranmayi is using HPC to make drug discovery faster and – ultimately – better.
A machine learning specialist at LLNL, she is part of the Accelerating Therapeutics for Opportunities in Medicine (ATOM) consortium’s data modeling team. By building deep learning models of secondary pharmacology, she hopes to give drug researchers the tools to predict adverse effects of drug candidates before they advance to animal and human trials.
“My goal is to release top performing models to the research community and make the results reproducible,” Hiranmayi explains. “This will help ATOM’s vision of transforming drug discovery from a slow, sequential, high failure process into a rapid, integrated, patient-centric model.”
For ATOM, Hiranmayi uses LLNL’s best in-class supercomputers, including its next-generation Sierra system for machine learning and algorithm development. She plans to release top performing machine learning models for 11 protein disease and related drug targets, with more expected in the future for additional protein targets as they become available.
Merritt, C. (2021, March 31). Meet Six Trailblazing Women in HPC. SC21. https://sc21.supercomputing.org/2021/03/31/meet-six-trailblazing-women-in-hpc/
|
|
Amanda Paulson, Ph.D.
- Conference: *NIH Spring Research Festival
- Date: April 28, 2021
- Title: Toward a Data-Driven DILI Prediction Platform
- Explore the Poster Here!
- *Amanda Paulson, FNLCR, won her poster session for the informatics category!
|
|
|
Jonathan Allen, Ph.D.
- Seminar: NCI DATA Science Learning Exchange
- Date: May 25, 2021
- Title: AI in Drug Development, presented by the ATOM consortium
- Explore the Presentation Here!
|
|
|
|
Sarangan "Ravi" Ravichandran, Ph.D., PMP
- Workshop: NCI Data Science Learning Exchange Hands-on Tutorial: ATOM Modeling Pipeline (AMPL) for Drug Discovery
- Date: June 8, 2021
- Title: Toward a Data-Driven DILI Prediction Platform
- Learn more about AMPL!
|
|
|
Open Position at ATOM Member Organization
|
|
ATOM Fellowship
FNLCR is accepting applications for the position of ATOM Data Science Fellow
|
|
Learn more about ATOM
Find us at an upcoming conference in 2021
|
|
|
|
|
|
