Transcript Use of Artificial Intelligence in the Design of Small

Artem Cherkasov, Kai Hilpert, Håvard
Jenssen, Christopher D. Fjell, Matt
Waldbrook, Sarah C. Mullaly, Rudolf
Volkmer § and Robert E.W. Hancock
ACS Chem. Biol., 2009, 4 (1), pp 65–74
MRSA
MRSA infected human tissue
Consider the following harrowing facts:




In 2002, 57.1 percent (an estimated 102,000 cases) of the staph bacteria found in U.S. hospitals
were methicillin-resistant (MRSA), according to CDC.
The total cost of antimicrobial resistance to U.S. society is nearly $5 billion annually, according
to the Institute of Medicine (IOM).
About 2 million people acquire bacterial infections in U.S. hospitals each year, and 90,000 die
as a result. About 70 percent of those infections are resistant to at least one drug, according to
the Centers for Disease Control and Prevention.
Recent CDC data show that in 2002, nearly 33 percent of tested samples from ICUs were
resistant to fluoroquinolones. P. aeruginosa causes infections of the urinary tract, lungs, and
wounds and other infections commonly found in intensive care units.
Antibiotic: a substance that kills or inhibits the growth of bacteria
(e.g. penicillin, erythromycin, anisomycin,etc)
Anisomycin
Erythromycin
Penicillin
Different types of bacteria exhibit different ways of resistance.
Some contain enzymes to change the chemical structure of the antibiotic.
Some contain enzymes capable of splitting the antibiotic molecule apart.
Some are able to “flush” the antibiotic out of the cell before it can fatally wreck the little
creature.
Each of these abilities are encoded by resistance genes often found in bacterial plasmid.



Peptide: a polymer made up of amino acid monomers (e.g. the 9-mer
KRWWKWIRW in Hancock et al)
Peptides antibiotics are simply antibiotics that are composed either partially or
wholly of amino acids.
Almost all species have evolved antimicrobial peptides capable of attacking
microbes directly, or, indirectly, by bringing about an innate or inflammatory
immune response.
Various peptide antibiotic ribbon structures
Actinomycin D

Scientific goal: Based on a antimicrobial peptide found in nature, in this
case the bovine (as in cow) neutrophil cationic peptide bactenecin
(RLCRIVVIRVCR-NH-2), that is known to serve a desirable function per
this study, perhaps we can scramble its AA sequence to determine if there
are even better antimicrobial peptides of the same length.
Cattle neutrophil bactenecin
RLCRIVVIRVCR-NH2 (from left to right)
•Each circular region
contains a synthesized
peptide.
•The tiny penciled-in
dots are the actual
specific peptides.
•Each of these can be
punched out and tested
for various biological
functions (e.g.
antimicrobial activity).




Lux assay (no graphical representation depicted in reference 20) is
accomplished by taking the peptides from the SPOT synthesis,
punching them out, transferring them to microtiter plates, and
seeing if they reduce the ability of P. aeruginosa to
bioluminescence.
An active antimicrobial peptide will destroy the P. aeruginosa and
stop its from luminescence.
An inactive antimicrobial peptide will not destroy P. aeruginosa
and thus the organism’s beautiful bioluminescent display will
persist.
Combinations of single or multiple AA substitutions led to
peptides with better antimicrobial activity than Bac2A.





Preferred AAs are found in the best antimicrobial peptides from (refs. 20, 21). They tend to be
hydrophobic and amphiphathic AAs.
Using these preferred AAs from (refs. 20, 21) the authors design sets of 943 and 500 cellulose
peptides (sets A and B respectively).
Best set A amino acid preferences were used to adjust the amino acid composition of set B.
Adjustments made to set B resulted in better antimicrobial activity than set A relative to
Bac2A.
Amino acid composition of set B thus formed to the lead amino acids to be tested in silico.
Figure 1. Occurrence of amino acids in the training and QSAR predicted data sets. The predicted activity
quartiles from the 100,000 virtual peptide library are marked as Q1−Q4.





Set B peptide AA preferences, representing the best amino acid
sequences (increased Ile, Arg, Val, and Trp) were used for random
computer generation of 100, 000 virtual peptides (out of an
astounding 70 billion possible 9-mer variants.
No position-specific requirements and 16 out of 20 natural AAs
used in simulation.
4 peptides were not used (3 residues were found not make for
good antimicrobials in previous libraries and cysteine results in
dimerization via disulphide formation.
QSAR solutions from sets A and B were used to help evaluate the
effectiveness of these 100, 000 virtual peptides.
QSAR?






Quantative Structure-Activity Relationships
A QSAR is a mathematical association between a biological
goings-on of a molecular system and its geometric and chemical
characteristics.
QSAR attempts to find reliable relationship between biological
activity and molecular properties, so that these “rules” can be
used to assess the activity of new compounds.
Sets A and B were used to create QSAR models relating chemical
characteristics to antimicrobial activity.
Artificial neural network was used to relate chemical descriptors
to antimicrobial activity for the 100, 000 computer generated
peptides.
100, 000 peptides were broken down into four quartiles based on
activity predicted : high, medium, low, and completely inactive.







“Chemical space” is the short qualitative answer to the following question: “How
many different types of chemical compounds are theoretically capable of existing?”
Chemical space includes: biopolymers, synthetic polymers, metallic clusters, small
carbon-based compounds, organometallic systems, etc.
Not all of chemical space may be biologically relevant. Even so, the number of
small carbon based molecules with a molecular weight of less than 500 daltons (the
molecular mass of many compounds found in living systems) is estimated to be 1 x
10^60!
The number of compounds required for synthesis in order to place 10 different
groups in 4 positions of benzene ring is 104
In silico modeling is thus necessary to search through small parts of chemical space
in a reasonable time and cost-effective manner.
A type of chemoinformatic computer modeling called QSAR is one of the methods
by which a virtual library of compounds can be generated from lead compounds
with certain desirable “drug-like” characteristics.
But first, for the purposes of this study, a lead antimicrobial peptide must be
developed and its biological activity determined (set B in Hancock et al 2009).




neural networks are attempt to make computers process information like human
neurons.
The human brain is essentially a vast array of interconnected neurons that respond
differently to different types of information
Massive interconnectivity allows for many parameters to be looked all at once as
opposed to regression analysis which typically deals with a much smaller number
of variables.
Authors refer to ANN using a “black box” metaphor—that is, they are not totally
sure how the neural network is coming up with its results. The authors leave to a
future paper an attempt to explain how the ANN is working its magic (Hancock et
al, 2009 in prep.)
Artificial Neural Network
Figure 2. Antimicrobial activity and physical parameters for antimicrobial peptides
from Training Sets A and B and peptides from the 100,000 peptide virtual library.
Tests of Candidate Peptide Antibiotic Effectiveness
Figure 3: Ability of the Peptides HHC-10 and HHC-36 to protect
mice against invasive S. Aureus infection
Starting
peptide
In silico
QSAR/ANN
analysis
Superior
training
compounds
AA Sequence
scrambled
versions
SPOT
syn/lux
assay




Not just theory: peptide antibiotic MX-226 has been shown to
significantly limit catheter colonization in phase IIIa clinical trials.
QSAR/ANN is not a one cycle process. It exhibits positive
feedback: lead compound to improved virtual compound to drug
candidate which may then be used in turn as a lead compound, ad
infinitum.
Peptide antibiotics have some negative characteristics such as
unknown toxicities, degradation by proteases (enzymes that break
down proteins), and high cost (amino acids are expensive building
blocks).
Not all of the structural characteristics of what makes a good
peptide antibiotic are known at this time.



Dr. Case
The students of Chem258
Antibiotics and bacteria