Statistical methods for the objective design of screening procedures - - PowerPoint PPT Presentation

Statistical methods for the objective design of screening procedures for design of screening procedures for macromolecular crystallization

Daniel Hennessy, Bruce Buchanan, Devika Subramanian, Patricia A. Wilkosz and John M. Rosenberg

(Intelligent Systems Laboratory, Pittsburgh, PA 15260, USA; Computer Science Dept, Rice University, Houston TX USA and Depts of Biological Sciences and Crystallography University of Pittsburgh Pittsburgh Houston, TX, USA, and Depts of Biological Sciences and Crystallography, University of Pittsburgh, Pittsburgh, PA 15260, USA)

Anita Anburajan Anita Anburajan BBSI Class of 2008

C t ll h B k d Crystallography Background

“Crystallization is a little like hunting, requiring knowledge of your prey and a certain low cunning..” Perutz

Goal:

One converts a given macromolecule into a crystalline state

i h hi h d f i l d b i l i

y p y g

with a high degree of internal order by manipulating environmental parameters to slowly reduce solubility.

Problem Characteristics:

• Difficulty in applying thermodynamic theory ‘at the bench.’
• Many variables with many possible values
• Givens with limited predictive value
• Strong interdependence amongst variables

g p g

• Limited time, material and human resources

I t d ti Introduction

Past Methods:

Past Methods:

Carter & Carter (1979)

Jancarik & Kim (1991)

Jancarik & Kim (1991) Samudzi et al. (1992)

Use of BMCD (Biological Macromolecular

Use of BMCD (Biological Macromolecular

Crystallization Database) Di d M th d

Discussed Method:

Using a Bayesian approach, we look for ‘patterns

• f crystallization ’
• f crystallization.

M th d Methods

A Macromolecular A Macromolecular

Hierarchy

• Based on size and complexity

Restructuring of the

BMCD data BMCD data

• Data re-representation
• Attribute Abstraction
• Data Labeling
• Data Subsetting

M th d Methods

Statistical Analysis of BMCD

Via Student’s t-test – compares the means of the numeric attributes

(pH, temperature and macromolecular concentrations) between different macromolecular classes. Robust against non-normal populations. Probabilistic Screen Design – rationale

Applying BMCD data as directly and objectively as

ibl possible.

An extension and combination of partial factorial and

sparse matrix approaches. p pp

Based on previous experience and anecdotal evidence.

P D i ti Program Description

• Procedure:
• Input parameters for

crystallization into User interface for Probabilistic Screen Design program

• Given the entered

parameters, program selects set of trials best suited.

• The user specifies the number
• f desired trials, generated by

Carter & Carter algorithm.

• For each trial, necessary

frequencies are extracted from the database and b bilit f i probability of success is computed.

C t lli ti N t b k Crystallization Notebook

Functions: Functions:

• Recording and archiving
• Tools for performing chemical

and related calculations and related calculations

Features

Di l i th f t f ti

• Display in the format of entire

tray for easy readability

• Including: Phase diagram

display of results; flexible entry display of results; flexible entry

• f recipe/concentration data;

mass editing; ‘default; tray setups; Grid Screen wizard; i t d t t printed output

Probabilistic Screen Design - b bilit t ti probability computation

The probability that the diffraction limit is d th th h ld i ifi d under the threshold given specified values for the crystallization variables : Can we assume independence amongst any of the variables?

Crystallization parameter dependency graph.

any of the variables?

With each parameter computed as such : First operand in numerator Second operand in denominator is analogous.

R lt d Di i Results and Discussion

• The Student’s t-test supports

pp there are several highly significant differences in the distributions of temperature and/or pH in different classes. p

• Results support ‘patterns of

crystallization’, however results provide little guidance ‘at the provide little guidance at the bench.’

• Goal of software is to identify

y probable crystallization conditions for a given macromolecular class and to guide experiments toward g p those conditions.

R lt d Di i Results and Discussion

Striving for objectivity in future versions of the

Striving for objectivity in future versions of the

program

Selection bias / Human behavior Selection bias / Human behavior

(ie. Success at 294 K vs. 277 K)

Consideration of isoelectronic point Consideration of isoelectronic point Failure data

How effective is Probabilistic Screen Design?

Evaluation of Probabilistic S D i Screen Design

10 Test Proteins 10 Test Proteins Evaluation Criteria

Compare effectiveness of Probability Screen Design with: Compare effectiveness of Probability Screen Design with:

Commercial Screen: Hampton Screen I & II Routine Protocol: Published Grid Screen

Scoring: 1 to 5

Quality Class Probabilistic Screen Design Crystal Screen I & II Grid Screens

1 = Clear (Very Bad) 2 = Precipitates (Bad) 3 = Microcrystal / Shower (Good) 4 = Small / Poor Crystal (Very Good)

Screen Design (BCR) I & II (Sparse Matrix) 5 3 6 4 5 + 4 10 7 5 5 + 4 + 3 10 8 6

Good) 5 = High Quality Crystal (Hooray!)

5 + 4 + 3 10 8 6

Q ti ? Questions?

A k l d t Acknowledgements

John Rosenberg John Rosenberg Dan Hennessy Judy Weiber