The Knowledge-Gradient Algorithm for Sequencing Experiments in Drug Discovery

Question

1. What are the contributions mentioned in the paper "The knowledge-gradient algorithm for sequencing experiments in drug discovery" ?

2. What is the recent ranking-and-selection algorithm?

3. What is the effect of the KGCB algorithm on the computation and memory requirements?

4. What is the common method of comparing compound activity with the results of a laboratory test?

Accepted Answer

Beginning with a base compound, the authors consider the problem of searching for a chemical derivative of the molecule that best treats a given disease.. The authors apply a recently developed algorithm, known as the knowledge-gradient algorithm, that uses correlations in their Bayesian prior distribution between the performance of different alternatives ( molecules ) to dramatically reduce the number of molecular tests required, but it has heavy computational requirements that limit the number of possible alternatives to a few thousand.. The authors develop computational improvements that allow the knowledge-gradient method to consider much larger sets of alternatives, and they demonstrate the method on a problem with 87,120 alternatives.

Accepted Answer

A recent ranking-and-selection algorithm that uses correlated Bayesian beliefs to take advantage of structure within the set of alternatives is the knowledge-gradient algorithm for correlated beliefs (KGCB) (Frazier et al. 2009).

Accepted Answer

Because the number of compounds is exponential in the number of substituents per site, maintaining a belief in this way reduces computation and memory requirements substantially and allows the KGCB algorithm to be used on problems with even hundreds of thousands of candidate compounds.

Accepted Answer

In both methods, multiple linear regression is used to correlate activity on laboratory tests with collections of compound features.

Accepted Answer

Although pure exploration is able to use the full Bayesian machinery to guide its implementation decisions and thus take advantage of the linear structure of the model, its measurement decisions are chosen purely at random.

Accepted Answer

In drug discovery, families of molecules often contain tens or hundreds of thousands of compounds, which makes this algorithm for computing KGCB computationally infeasible.

Accepted Answer

In this problem, because the prior distribution is independent and measurements of one alternative do not affect the other alternatives, KG can also be understood as a generalization of EI to the noisy case.

Accepted Answer

For a set of compounds with five sites and nine possible substituents at each site (to make 10 possible choices at each site, including the possibility of attaching no substituent), the number of possible compounds is 105 compared with only 45 total substituents.

Accepted Answer

observe that the beliefs about those bx not included in n+1 will not change as a result of measuring xn, and so ¶ 4èn1xn5, which is the standard deviation of the change in beliefs about the values of the alternatives, is not affected by these deviation terms not included in n+1.

Accepted Answer

that empirical study assumes thatmeasurements are free from noise, and the versions of these BGO policies they employ require modification for the noisy case before they can be applied to the problem the authors consider here.

Accepted Answer

Given these suppositions, their prior distribution is given by taking 00 and C 0 00 to be the value and measurement variance, respectively, of their previous measurement of the base molecule, and by taking 0i and C 0 ii for i>0 to be the estimated mean and variance, respectively, of the population of substituents.

The Knowledge-Gradient Algorithm for Sequencing Experiments in Drug Discovery

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What are the contributions mentioned in the paper "The knowledge-gradient algorithm for sequencing experiments in drug discovery" ?

2. What is the recent ranking-and-selection algorithm?

3. What is the effect of the KGCB algorithm on the computation and memory requirements?

4. What is the common method of comparing compound activity with the results of a laboratory test?

5. Why does pure exploration perform better than KGCB?

6. What is the computationally demanding step in the KGCB algorithm?

7. Why is KGCB a generalization of EI to the noisy case?

8. How many substituents are there for a set of compounds?

9. what is the standard deviation of the change in beliefs about the values of the alternatives?

10. Why does the KGCB policy require modification for the noisy case?

11. How do the authors determine the prior distribution of the base molecule?

Figures

Citations

The Design and Analysis of Experiments

Efficient Global Optimization of Expensive Black-Box Functions

A Tutorial on Bayesian Optimization

Bayesian reaction optimization as a tool for chemical synthesis.

Predictive Entropy Search for Efficient Global Optimization of Black-box Functions

References

The Design and Analysis of Experiments

Response Surface Methodology: Process and Product Optimization Using Designed Experiments

Advances in Neural Information Processing Systems 14

Efficient Global Optimization of Expensive Black-Box Functions

Efficient Global Optimization of Expensive Black-Box Functions

Related Papers (5)

Efficient Global Optimization of Expensive Black-Box Functions

The Knowledge-Gradient Policy for Correlated Normal Beliefs

A Knowledge-Gradient Policy for Sequential Information Collection

Practical Bayesian Optimization of Machine Learning Algorithms

A Tutorial on Bayesian Optimization of Expensive Cost Functions, with Application to Active User Modeling and Hierarchical Reinforcement Learning