Download presentation
Presentation is loading. Please wait.
1
Determining the Number of Non- Spurious Arcs in a Learned DAG Model: Investigation of a Bayesian and a Frequentist Approach Listgarten & Heckerman
2
2 Purpose Design a vaccine for HIV By considering many patients and observing which HLA molekyles causes the T-killer cells of the imune system to react
3
3 Definitions HLA = Human leukocyte antigen Each person usally has [3;6] Epitopes = bits of protein Results of T-cell attacking HIV-peptide Peptide = “small digestible” Link between amino acids
4
4 How? Find out which HIV peptides interact with which HLA molekyles by using a graphical model.
5
5 Solution A directed acyclic graph representing HLA and peptides HLA h 1 HLA h 2 HLA h 3 HLA h 4 peptide y 1 peptide y 2 peptide y 3 HLA h N peptide y M... Model for one patient. A design of a vaccine is to identify a set of peptide-HLA-pairs, which are epitopes for a large number of the population
6
6 Properties Bi-partite model(2 levels) HLA can have zero or several outgoing archs Peptide can have zero or several ingoing archs Each patient will have [3;6] HLA nodes that are “on” Answers: which HLA molekyle(s) are(is) responsible for a given immune system reaction
7
7 Two approaches Bayesian Frequentist
8
8 Bayesian Approach cont. 1(2) true arch distribution bayesian expectation with given data D the number of archs both in G and G’ Ddata G’proposed model Gall possible graph structures
9
9 Bayesian Approach cont. 2(2) Exponentional complexity…! Can be improved by limiting |Parent set| Limit=5, gives identical results
10
10 Frequentist Approach FDR = False Discovery Rate Given a set of hypotheses Hypothesis i has a test score s: assumed to be independent in a given hypotheses
11
11 FDR cont. 1(4) Eexpected value Fnumber of false hypotheses Snumber of hypotheses with s i > t tthreshold
12
12 FDR cont. 2(4) Rewrite Where is a structure search algorithm
13
13 FDR cont. 3(4) – multiple data sets Q - – number of archs found by applying to real data, D
14
14 FDR cont. 4(4) Standard FDR: The average over multiple datasets +1 – smooths the estimate
15
15 Results PPV – positive predictive value Frequentist method: Bayesian method:
16
16 Results on non-HIV data
17
17 Results on non-HIV data
18
18 Results on synthetic HIV data
19
19 Results on real HIV data 8 results…. all matches
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.