QSARS AND VFARS TO SUPPORT RISK ESTIMATION

On June 20-21, a workshop on QSARS AND VFARS TO SUPPORT RISK ESTIMATION was held at Cincinnati, sponsored by the U.S. Environmental Protection Agency’s National Homeland Security Research Center (NHSRC) and National Risk Management Research Laboratory (NRMRL). The two-day conference featured invited speakers each morning, followed by detailed discusssions on Dr. Gerald Stelma from EPA gave an Introduction to the VFAR (Virulent Factor Activity Relationships) Concept. The VFAR premise is that architechtural & biochemical components of pathogens are closely related, based on the concept of descriptors tied to specific genes. VFAR Descriptors include:

• Genetic Elements
  
• Surface Proteins
  
• Toxins
  
• Attachment Factors
  
• Metabolic Pathway
  
• Invasion factors

In a comparison of QSAR vs.VFAR, chemicals can be considered static (although some participants objected to this characterization), while microbes are dynamic, esp. structural genes.
Among the unknowns are:

• Too many unknown virulence genes
  
• Individual virulence genes necessary but not sufficient for virulence
  
• Host susceptibility factors, dosages
  
• DNA variability among structural genes
  
• Effects of unexpressed virulence genes
  
• Are VFARs valid for viruses and protozoa?
  
• Effects of DNA from dead cells

Next Joan Rose from Michigan State University spoke on Using VFAR in the Risk Assessment Framework. Among her main points:

• Hazard identification — source identification, virulence, potential for severe outcome
  
• Dose response — potency
  
• Exposure — source persistence
  
• Characterization — sensitive populations, evolution of pathogens
  
• Health risk, remediation
  
• Host specific markers — don’t need biology, but need to know prevalence of gene in specific population
  
• Exposure Assessment — risk, duration, concentration
  
• What genes, structures are important to survival in different media?
  
• How much uncertainity in framework to be built?
  
• Water-borne outbreaks are opportunities to isolate known pathogens & identify gene function

This was followed by Syed Hashsham, also from Michigan State University, speaking on VFAR: Factors Related to Genomic Variabilities and describing the use of the Virulence and Marker Gene database. The last speaker of the day was Paul Schaudies from SAIC describing A Bioinformatic Approach to VFAR Analysis and Characterization, using Molecular Radar and FIGUR software. The afternoon featured an in-depth panel discussion of the following VFAR Charge Questions:

1. Identify selection criteria for VFs that should be considered in the VFAR approach. Should certain classes of VFs be excluded?
   
2. Compare & contract VFAR & QSAR approaches. Considering the similarities to QSAR, should the VFAR approach work with biotoxins? Viruses? Spores? Cysts? What are the strengths of the VFAR concept?
   
3. Discuss how VFARs can be used in the detection of recognized biothreat agents, newly emerging & bioengineered pathogens.
   
4. Describe technology available for examining VFs. How can we determine the presence of such VFs in water or air?
   
5. Discuss positive & negative applications of VFARs in bioengineering and the construction of highly potent pathogens inserting single genes or combinations of virulence genes into commensal organisms. Do certain classes of virulence genes lend themselves to genetic engineering?
   
6. How can VFARs be used to determine the human toxicity potential of the virulent genes? Is it possible to obtain a quantitative estimate of the virulence along with a qualitative estimate?
   
7. Can a virulence gene be altered so that it is still active but no longer detectable by the gene probes that are typically used?

The second day was devoted to QSAR, beginning with a talk by Mark Cronin from Liverpool John Moores University, on Integrating Alternative Techniques to Predict Toxicity. In dealing with prediction models for Reactive Toxicity, the question is: Can we go in chemico to in silico? Among the other points raised:

• Specific reactivity is poorly characterized in toxicology, but underpins many end points
  
• Measuring reactivity in chemico has been shown to assist in predicting toxicity
  
• There is a need for more tools including computational & in silico

The next speaker was Kannan Krishnan from Université de Montreal, discussing Integrated QSAR-PBPK Modeling for Risk Assessment Applications. Among his points:

• QSAR models are based on response-specific dose level for each species
  
• No effects on relationship between structure & internal dose
  
• Can we develop QSAR for pharmacokinetic profiles? Changing as function of route?
  
• QSAR-PBPK models facilitate internal dose-based risk assessment
  
• Influence of exposure concentration, route can be examined
  
• Effects on specific sub-populations can be evaluated
  
• Modeling of multi-route exposures for risk assessment

Andrew Maier from TERA analyzed Integration of Mode of Action (MOA) and Weight of Evidence (WOE) Concepts in Predictive Toxicology. WOE is increasingly used in hazard screening algorithms and is characterized by “totality of evidence” in making decisions about causality. The emphasis on Totality has opened the door for predictive toxicology tools. These evolving concepts are driven by impoved biological understanding and by improved sophistication of validation tools. Tools for evaluating WOE include Expert judgement (Peer review & consultation, Expert classification techniques) and Quantitative tools (Bayesian analysis). Currently SARs & QSARs are used as independent tools. Biological understanding is necessary for interpreting results. For risk assessment, we distinguish between mechanism of toxicity (detailed understanding of cellular & sub-cellular processes) vs. mode of action (less detailed, generalized cellular response). The goal is maximizing use of biology. WOE evaluation represents a maturation in chemical risk assessment and is of critical use in resolving conflicting data. Advances in basic biology, computational chemistry & statistical/mathematical methods should be used in risk assessment.
Next William Welsh from the Robert Wood Johnson Medical School presented Activities at the New UMDNJ Computational Toxicology Center: Advanced QSAR-based Methods of Rapid Hazard Identification, Prediction and Characterization. He described the DORIAN (Dose Outocme Response) computational toxicology system, the Environmental Bioinformatics KnowledgeBase and his integrated structure- & ligand-based scrrening approaches, using:

• Decision Forest (for fast consensus modeling: they improve classification by combining individual models)
  
• Shape Signatures (enabling large-scale screening of query chemicals against databases based on similarity in shape and other bio-relevant molecular factors) 1-D and 2-D Shape Signatures have been computed for receptor binding sites of PDB-extracted ligands in bio-active conformation. These Shape Signatures are rotationally invariant and encode surface charge and polarity. These Shape Signatures employ the same ray-tracing technology we use in PEST descriptors.
  
• Polynomial Neural Networks (PNN, for optimal linear/nonlinear QSAR)
  
• VHTS (to predict ligand binding affinity and MOA)

.
The Heirarchial Screening Framework employs:

• Rejection Filters (e.g. MW, structural properties)
  
• Active/Inactive Assessment (Shape Signatures, Pharmacophores)
  
• Quantitative Predictions (based on PLS/QSAR/PNN)
  
• Knowledge-Based Approaches

More details at http://www.ebCTC.org/
The last speaker of the morning was Andrew Worth, Scientific Officer from the European Commission, discussing The Role of the European Chemicals Bureau in Promoting the Regulatory Implementation of Estimation Methods. The rest of the day, until the conclusion of the meeting, was devoted to a panel discussion of the QSAR Charge Questions:

1. In light of the complexities of developing QSAR models, what is the potential for emerging technologies (e.g. genomics, proteomics, bioinformatics) to reduce the uncertainities involved in the application of QSAR methods to EPA’s risk assessment/risk management process?
   
2. How can genomic, proteomic & bioinformatic data be used in predictive models? There are examples of commercial organizations (e.g. ICONIC, GeneLogic) developing predictive signatures for chronic hepatotoxicity based on genomic expression patterns. What lessons have been learned by these approaches? Are there examples where the “omics” technologies in combination with QSAR models have proven to be able to predict, both qualitatively and quantitatively, acute/chronic toxicity across multiple chemical classes?
   
3. Since rule-based and expert models are based on congeneric groupings of chemicals (i.e. the training set is a congeneric data set), how can statistical models, that are generally based on non-congeneric training set, be improved? Can such models incorporate MOA data if available? Can statistical models provide some insight re:MOA for a chemical query?
   
4. The toxicity of a chemical for any given health endpoint is in general due to an adverse interaction between the chemical and/or its metabolite and the tissue/organ/DNA, etc. associated with the endpoint. In developing statistically-based QSAR models for chemicals with different MOA, the descriptor pool contains descriptors that are chemical specific (i.e. they depend on the structure of the chemical alone). Are there any descriptors that can describe the tissue/organ/DNA characteristics and its interaction with a chemical and/or its metabolites?
   
5. Current methodology on the statistically-based QSAR development for toxicity prediction calls for the inclusion of as many (classes of) descriptors in the descriptor pool as possible to explain the variance in the dependent variables (some measure of toxicity). In developing these QSARs, are there any (class of) descriptors that one should definitely include in the potential descriptor pool (e.g. partition coefficients to account for transfer from blood to tissues, etc.)?
   
6. Qualitative SAR models (i.e. models yielding dichotomous or graded responses such as yes/no or low/med/high, etc.) do not provide a quantitative measure of a chemical’s toxicity while quantitative SAR models (i.e. models yielding numerical potency estimates) do not provide a qualitative measure of the activity of a chemical for any given health endpoint. How does the panel view the feasibility of applying hybrid QSAR models (i.e. capitalizing on the benefits of SAR & QSAR by minimizing the disadvantage, if any, of each approach) for toxicity prediction? If feasible, how does the panel envision EPA applying such models?
   
7. The current state of the science for the extrapolation of acute to chronic toxicity is the use of default uncertainty factors described in Agency risk assessment guidance documents. Can QSAR methods be used to reduce the uncertainty in extrapolating from acute and short-term benchmarks (such as LD50) to sub-chronic and chronic LOAEL? What are the issues that must be dealt with in order to do this?