Informatics and Drug Discovery

1
Informatics and Drug Discovery

• Peter Goodfellow

2
20th Century Health Achievements

• Vaccination
• Control of infectious diseases
• Decline in deaths from coronary heart disease and
  stroke
• Family planning
• Healthier mothers and babies
• Fluoridation of drinking water
• Safer and healthier foods
• Recognition of tobacco use as a health hazard
• Motor vehicle safety
• Safer workplaces
• Source CDC MMWR April 02, 1999 /
  48(12)241-243 http//www.cdc.gov/mmwr/preview/mm
  wrhtml/00056796.htm

3
AIDS Mortality and Protease Inhibitor Use
Deaths
Deaths per 100 person-years
Therapy with a PI ( of patient-days)
Use of protease inhibitors
1994
1995
1996
1997
1998
Year
Palella et al. N Engl J Med 1998
4
Drug Discovery
Output of New Molecular Entities
120
100
Index ( of 1994 output)
80
60
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
Year
Source CMR International
5
The Drug Discovery Process
The aim is to translate new information into new
therapies
6
Complexity of Drug DiscoveryFinding a Molecule
that Satisfies Multiple Criteria
1 Drug Molecule
manufacturable
patentable
non-mutagenic
non-teratogenic
durable
reversible
non-inducing
metabolically stable
permeable
soluble
physically stable
potent
10,000 Drug Candidates
selective
Valid Biomedical Hypothesis?
targeted
7
Complexity of Drug DiscoveryFinding a Molecule
that Satisfies Multiple Criteria
1 Drug Launch
Regulatory filing
Competitive profile
Cost-effective manufacturing
Carcinogenicity studies
Long-term safety
Efficacy
Side effect profile
Dosing ranges
Patient recruitment
Trial sites and investigators
Stability
Formulation
10 Drug Molecules
Safe and active in lab and animal models
All discovery criteria met
8
Predictive Models

• A predictive model quantitatively relates a
  number of descriptors (variable factors that are
  likely to influence future behaviour or results)
  to an outcome.
• In marketing, for example, a customer's gender,
  age, and purchase history (descriptors) might
  predict the likelihood of a future sale
  (outcome).
• In drug discovery, descriptors tend to be derived
  from chemical structure, and outcomes are in
  vitro or in vivo phenomena
• the goal is to predict behaviour before synthesis
• models can be built from experimental data too
• e.g. prediction of F from solubility,
  permeability and clearance data

9
(No Transcript)
10
Statistics

• Various statistical methods are applied to find
  the mathematical relationship between the
  descriptors and the outcomes
• multiple linear regression, logistic regression
  K-nearest neighbours, PLS, linear discriminant
  analysis, decision trees, neural networks,
  Support Vector machines and many more
• Choice depends on
• data type/volume
• the objectives for the model (see later)
• personal preference

11
Modelling Decisions

• Model in vivo or an in vitro surrogate?
• in vivo ideal, but often limited data set
• in vitro is itself a model for in vivo
• but data generation is easier
• E.g. Absorption
• Caco-2 cell in vitro data vs in vivo perfusion
  data
• Use of data
• Is the data good enough to be left as real
  numbers e.g. pIC50?
• Or should it be used as a category e.g. high,
  medium, low?
• Do you want to filter bad, prioritise good or
  both?
• Do you need to avoid false positives or
  negatives?
• One is usually more important than the other

12
Uses of Predictive Models in Discovery

• Lead generation
• Filtering of structures to remove poor start
  points from screening collection
• Lipinskis rules, sub-structure filters, hard
  to remove or critical properties like poor
  solubility, permeability and hERG interaction
• Even 70 predictive models are useful, as they
  can enrich the proportion of good compounds
  coming in
• Hit to candidate
• Used to guide medicinal optimisation
• Predictive power and interpretability are key
• Interpretability can often compensate for poor
  predictive power, as gives insights to the
  chemists as to what might solve the problem
• Candidate attrition
• Predictive ADMET used as another component of
  risk assessment for taking a candidate forward,
  to aid formulation studies, or to help interpret
  the result of an experiment

13
Modelling Retention Times on HPLC
Q. Given about half a million good quality
retention times and chemical structures, can we
build a model of retention time that would be of
use?
Pred. RT
Mean Absolute error 0.23 mini.e. 14 seconds
Abs. Error
Exptl. RT
Chris Luscombe CIX
14
Initial Filter from a Developability
AssayInterpretable rule, filters bad
compounds, with low false positive rate
143/160 compounds in the box are active
15
DEEP Detecting Adverse Events
Systems for Signal Detection DEEP Partnership
with Lincoln Technologies This system has now
been deployed at FDA, CDC,large Pharma (Pfizer,
Lilly, Bayer, BMS, JJ, Roche, AZ)
DEEP DEEP (Data Explorationand Evaluation in
Pharmacovigilance)
Scientific PublicationsNew strategies to
evaluate poly-therapy, drug interactions and
demographic risk factors for AEs
16
Safety Data Mining Enables Rapid and Systematic
Identification of Safety Signals

• With post-marketing data, it is difficult to
  distinguish signals from noise.
• Safety Data Mining (SDM)/disproportionality
  methods identify AEs that are reported with gt
  expected frequency (statistical independence)
• Frequency is assessed against the background of
  all other drugs and events. Results are used
  for hypothesis generation.
• Bayesian methodology to estimate relative
  reporting rates (risks) of AEs

Enhanced effectiveness of post-marketing
pharmacovigilance through rapid, systematic
screening of AE databases Enhanced benefit-risk
management
17
Bayesian Methods to Assess the Frequency of
Specific Drug-Adverse Event Combinations
Drug X
All other Drugs
Event of interest
C
A
All other Events
D
B

• An empirical Bayesian methodology estimates
  relative reporting rates

Is A gt C ?? AB CD
18
Interpretation

• Wonderex - Rash (16 reports in the database)
• EBGM 3.0 EB05 1.8
  EB95 4.3
• Wonderex-rash combination is reported at 3-fold
  greater frequency than if there were no
  association between Wonderex and rash
• 95 confidence that the true relative reporting
  rate is at least 1.8
• 95 confidence that the true relative reporting
  rate does not exceed 4.3

19
Enhanced Pharmacovigilance Had these tools
previously been available, critical signalsmight
have been identified years before they were
recognized with traditional pharmacovigilance.
They are now used routinely .
20
DEEP Provides Information to Reconise Product
Performance and Benefit-Risk Ratio

• Understanding the effects of litigation/publicity
  on safety signals
• Evaluating indication-specific safety profiles in
  products with multiple indications
• Evaluating rare serious events in special
  populations (i.e., children)
• Signal assessment for our co-licensed products
• Advisory committee preparation

• Benefit-risk management-Pharmacovigilance
  planning
• Competitive intelligence
• Regulatory agency queries
• Regulatory submissions for PLEs
• Characterizing factors associated with rare
  serious AEs
• In-licensing due-diligence
• Exploring drug interactions and polytherapy in
  real world use

21
NSAIDS COX-2 Inhibitors
AERS to 3Q03 (Suspect drugs)
22
AERS to 3Q03 (Suspect drugs)
23
Cardiovascular and Stroke-Related-AEs Subset
Analysis Age lt 50 yr
AERS through 3Q 2003
24
Chemical Safety Using human safety data to
determine which structural features of drugs
contribute to their toxicities
Identify associations between fragments and
signals,by calculating diagnostictest
statistics. A positive signal (EB05 ? 5 ) is
used as the gold standard. The presenceof a
fragment in drug represents a positive test.
Identify drug-event pairs with EB05 ?
5(designate as "signals").
Run datamining algorithm (MGPS).
Create a chemical fragment library for all drug
structures in AERS using MoSS to create
fragments ranging in size from 4-10 atoms.
Diagnostic test statistics For a given
fragment-event pair Odds ratio of 20 means that
the odds of having a specific "signal" are 20
times greater if the fragment is present (in the
molecule) than if it is not Positive predictive
value of 0.4 means that 40 of drugs containing
the fragment will have a signal for that
adverse event
25
(No Transcript)
26
Thanks to

• Darren Green, John Leonard, June Almenoff and
  Trevor Gibbs for sharing slides
• Colleagues who taught me about drug discovery
• SB and GSK for letting me play with a very big
  chemistry set