Transcript [Poster title] - Oakland University

With quantum theory, who needs drugs?
UnCoRe 2007
Alaesa Hearn1; Shabana Khan2; Mohamed A. Zohdy, PhD3,
1Texas A&M Corpus Christi, 2Cal Pomona, 3Oakland University
ABSTRACT
METHODS – PHARMACEUTICAL SIDE
Modern drug discovery process involves mapping pharmaceutical knowledge
about target proteins and ligands as well as sophisticated computer science
data mining. This process produces the so-called hit to lead which is followed by
lead optimization to take the drug to clinical trials. In this project, we have made
contributions to both the pharmaceutical and computer science domains which
includes: 1. Generalized data structure for drug effectiveness descriptors, 2.
Novel neural network evolved from supervised and unsupervised learning to
narrow down the choice of effective drugs, 3. Applied our methods preliminary
cancer-causing proteins with reasonable success.
The target proteins this project focused on are all involved in various forms of cancer.
MOTIVATION
If a new disease suddenly
emerges, the current system of
drug discovery will take years
to develop a cure. Our system
is intended to reduce the
response time dramatically.
Akt1
JAK2
METHODS – NEURAL NETWORK SIDE
RESULTS
ALIGNMENT 2 - Unsupervised Neural Network: 2D
The networks developed in this project are based on the competitive design.
This selects the node most similar to the given input, and also gives weight to a
mathematically designated neighborhood
DHFR
The protein at the right is dihydrofolate
reductase (DHFR). It is shown here with
ligands. It is crucial in cell division and
proliferation. We trained our neural
network using measurements of various
conformations of an inhibitor molecule of
this protein.
The above diagrams show a neural network with a winning node 13, and a circular
neighborhood of radius 1 (left) and a rhombus-shaped neighborhood (right).
Additional features were then added to the basic structure to adjust the efficiency
and effectiveness of the various networks.
The Drug Discovery Process:
ALIGNMENT 3 - Unsupervised Neural Network: 3D
RESULTS
The network before training. A 4-by-5 Self Organizing Feature Map
This research is based on many interacting fields:
CONCLUSIONS
RESEARCH OBJECTIVES
• Pharmacology Side
– Drug targets (macro molecules, key proteins)
– Ligands (micro molecules, drug compounds)
– Consider the docking, binging, and reacting of the targets and ligands
– Focus on proteins involved in various types of cancer
– State-of-the-art databases
• Neural Network Side
– Develop new neural network paradigms by combining known features
into new structures, i.e. Self-Organizing Vector Quantization
– Use the competitive network at the core, and add features of other
– Apply neural networks to discover effective drug molecules.
FUTURE RESEARCH OBJECTIVES
•
•
•
•
•
•
•
Drug Delivery
Drug interaction
Protein interaction
Personalized medicine
Drugs for developing countries
Orphan diseases
Drug addiction
We are drowning in information but starved for knowledge.
- John Naisbitt
As scientific research advances, the collective pool of data
grows exponentially. All of these facts must be collected
and organized in order to be useful information. However,
the sheer quantity of data presents difficulties in searching,
integrating, and applying this knowledge. Neural networks
are among the best tools available for recognizing and
analyzing relationships in vast amounts of information.
• Target Discovery
• Target Validation
• Assay Development
• Screening
• Hit to lead
ALIGNMENT 1 - Supervised Neural Network: 2D
In this time-constrained research, we have learn and applied so-called
rational drug discovery, in which computer science is essential to
complement pharmaceutical analysis to produce hits (promising molecules)
and leads (proven effective molecules). The core of this process utilizes
small molecule descriptions, which can be geometric, chemical, physical, or
a computer learning algorithm. We focuses on a class of neural networks
that are based on competitiveness and self-organizing. We then applied the
neural network appropriately to one specific protein, DHFR, and have been
able to show preliminary effectiveness of extended neural learning to binding
and reaction mechanisms of many possible inhibitor ligands.
• Lead Optimization
• Preclinical evaluation
Protein / Kinetic
Parameters (Trees)
• Non-human toxicology
REFERENCES
• Clinical trials
Surface Areas /
Radii (Huffman)
• Market
Annema, Anne-Johan. Feed-Forward Neural Networks: Vector Decomposition Analysis, Modelling,
and Analog Implementation. Norwell, MA: Kluwer Academic Publishers, 1995.
Yi, Zhang, and K.K Tan. Convergence Analysis of Recurrent Neural Networks. Norwell, MA: Kluwer
Academic Publishers, 2004.
Neelakanta, Perambur S., and Dolores F. De Groff. Neural Network Modeling. Boca Raton: CRC
Press, 1994.
Wade, L.G. Organic Chemistry. 3rd ed. Upper Saddle River, NJ: Prentice Hall, 1995.
Conformations
(Number)
Chemoinformatics
Descriptors
Warmuth, Manfred K. "Active learning with support vector machines in the drug discovery process."
Journal of chemical information and computer sciences 43.2 (2003): 667. 14 June 2007
<http://pubs3.acs.org/acs/journals/doilookup?in_doi=10.1021/ci025620t>.