Optimization-Based Peptide Mass Fingerprinting for Protein Mixture Identification

Transcription

1 Optimization-Based Peptide Mass Fingerprinting for Protein Mixture Identification Weichuan Yu, Ph.D. Department of Electronic and Computer Engineering The Hong Kong University of Science and Technology The First China Workshop on Computational Proteomics Beijing, Joint work with Zengyou He, Can Yang, Chao Yang, Robert Qi and Jason Tam 1 / 3

2 Outline 1 Introduction 2 Method 3 Experiments Simulation Study Real Data 4 Conclusion 2 / 3

3 Introduction Background of Mass Spectrometry Data Analysis We like to know: what proteins/peptides are in a sample? What are their expression levels? Issues in the analysis: Noise: Many sources: chemical, electrical, instrumental Physics not completely understood Low abundance signal coexists Measurement range: Is it enough to record all the information? Dynamic properties of samples: Do we obtain the data at the right time? 3 / 3

4 Introduction Motivation of Protein Identification MS data describes information directly at the peptide level. We need to identify the corresponding proteins to better understand the cellular functions. Information at the protein level is probably more robust. 4 / 3

5 Introduction Common Methods for Protein Identification 1. Peptide Sequencing Method (Using MS/MS data). Pros: More accurate Cons: Lower coverage (especially peptides with low intensity values) 2. Peptide Mass Fingerprinting (Using single-stage MS Data) Pros: Higher coverage Cons: Less accurate and cannot handle protein mixtures Can we remove the limitations of PMF? Our approach: We formulate the identification of protein mixtures as an optimization problem. 5 / 3

6 Introduction Common Methods for Protein Identification 1. Peptide Sequencing Method (Using MS/MS data). Pros: More accurate Cons: Lower coverage (especially peptides with low intensity values) 2. Peptide Mass Fingerprinting (Using single-stage MS Data) Pros: Higher coverage Cons: Less accurate and cannot handle protein mixtures Can we remove the limitations of PMF? Our approach: We formulate the identification of protein mixtures as an optimization problem. 5 / 3

7 Method PMF for Single Protein Identification PMF method for single protein identification consists of the following steps: (1) Protein purification: The 2D gel-based separation produces purified protein samples. (2) Protein digestion: e.g. trypsin digestion. (3) MS data acquisition and peak detection: record the masses of resulting peptides. (4) PMF scoring: match the MS spectrum with respect to the protein database and report the best ones. 6 / 3

8 Method PMF for Single Protein Identification Find a single protein that maximizes the scoring function: X = arg max X i D S(L) (Z, X i ), (1) Z = (z 1, z 2,...z l ): Experimental peaks D = (X 1, X 2,...X g ): Protein database S (L) (Z, X i, σ): Scoring function, σ: mass tolerance threshold. 7 / 3

9 Method PMF for Protein Mixture Identification Replace single protein X i with a set of proteins Y: Ŷ = arg max Y D S(M) (Z, Y), (2) We have the same input Z Our objective is to find a set of proteins Ŷ that best explains Z 8 / 3

10 Method Existing Approaches S (L) (Z, X uj ) Directly apply the single protein identification method to protein mixtures. The subtraction strategy: (1) First identify a protein with the highest score; (2) Remove the peaks associated with this protein from the input data (3) Go back to step (1) until the score is lower than a predefined threshold. Suppose the peak subset Z is empty. At each step j(1 j k), calculate the score as: j 1 S (L) (Z Z t, X uj ). (3) t= 9 / 3

11 Method Choice of Scoring Function (1) Virtual single protein approach. S (M) (Z, Y) = S (L) (Z, Ṽ). Ṽ: Use a set of proteins to represent a virtual single protein. (2) Peak partition approach. S (M) (Z, Y) = k j=1 S(L) (Z j, X uj ). Partitioning Z into Z j is tricky. The subtraction strategy: greedy partition Random matching is the major concern We choose the virtual single protein approach here. 1 / 3

12 Method Scoring Function for Single Protein Identification The probability that a protein X i has r i randomly matched peaks in Z (assuming binomial distribution): Pr( M Z (X i ) = r i ) = C r i l pr i i (1 p i) l r i (4) Score: S (L) (Z, X i ) = ln C r i l r i ln p i (l r i ) ln(1 p i ) (5) M Z (X i ): subset of Z whose peaks match protein X i l: the number of observed peaks in Z r i : number of peptides in protein X i. p i : the probability for at least one match : mass range; σ: mass tolerance. p i = 1 (1 2σ/ ) n i, (6) 11 / 3

13 Method Scoring Function for Protein Mixture Identification S (M) (Z, Y) = ln C r Y l r Y ln p Y (l r Y ) ln(1 p Y ) (7) Y consists of k proteins X u1, X u2,..., X uk r Y = k j=1 M Z (X uj ) and p Y = 1 (1 2σ/ ) k j=1 n u j. 12 / 3

14 Method Maximization of Scoring Function Two cases: The number of ground-truth proteins k is known. Losak Algorithm The number of ground-truth proteins k is unknown. Losau Algorithm 13 / 3

15 Method Losak Algorithm A LOcal Search Algorithm with Known k. (1) Randomly select k proteins into Y as target proteins (2.1) In the iteration process, swap each non-target protein with the k target proteins and re-evaluate the scoring function. (2.2) keep those target proteins that achieve the best scoring values and proceed to the next protein. 14 / 3

16 Method Losak Algorithm-Detail 15 / 3

17 Method Losak Algorithm-Detail 15 / 3

18 Method Losau Algorithm A LOcal Search Algorithm with Unknown k. (1) Initialize with 2 proteins. (2) Iterate with swap, insert and delete operation. (Occam s razor principle,penalizing insert operation using ω) (3) Prune the protein list 16 / 3

19 Method 17 / 3

20 Method Filtering Procedure in the Losau Algorithm Idea: if X uj is the ground-truth protein, then the chance that other proteins in the database has a better score than X uj on M Z (X uj ) is very low. We use the number of winning proteins to measure the rank uncertainty and θ as the threshold to remove false positives. 18 / 3

21 Experiments Evaluation Criteria and PMF Algorithms We use standard performance metrics in information retrieval, including precision, recall, and F1 measure. n TP : the number of true positives. n FP : the number of false positives. n P : the number of all ground-truth proteins. precision = n TP /(n TP + n FP ) recall = n TP /n P F1 measure = 2 precision recall precision+recall 19 / 3

22 Experiments Algorthims In performance comparison, we use the following algorithms: SPA: single protein identification algorithm. Subtraction algorithm Losak algorithm Losau algorithm 2 / 3

23 Experiments Simulation Study Outline 1 Introduction 2 Method 3 Experiments Simulation Study Real Data 4 Conclusion 21 / 3

24 Experiments Simulation Study Performance Comparison 22 / 3

25 Experiments Simulation Study Performance Comparison Average Precision/Recall/F1 Measure (%) (a) Mass Error Ranges from.1 to.4 Da Mass Error (Da) Average Precision/Recall/F1 Measure (%) (b) Sequence Coverage Ranges from.1 to Sequence Coverage Average Precision/Recall/F1 Measure (%) (c) Noise Level Ranges from 1% to 7% Noise Level (%) Average Precision/Recall/F1 Measure (%) (d) Protein Number Ranges from 1 to Number of Proteins SPA Subtraction Losak Losau (Precision) Losau (Recall) Losau (F1 Measure) 22 / 3

26 Experiments Simulation Study Running Time Comparison 23 / 3

27 Experiments Simulation Study Running Time Comparison Average Running Time(s) (a) Mass Error Ranges from.1 to.4 Da SPA Subtraction Losak Losau Average Running Time (s) (b) Sequence Coverage Ranges from.1 to.4 2 SPA Subtraction Losak 15 Losau Mass Error (Da) Sequence Coverage Average Running Time (s) (c) Noise Level Ranges from 1% to 7% SPA Subtraction Losak Losau Average Running Time (s) (d) Protein Number Ranges from 1 to 7 SPA Subtraction Losak Losau Noise Level (%) Number of Proteins 23 / 3

28 Experiments Simulation Study Number of Iterations of Our Algorithms 24 / 3

29 Experiments Simulation Study Number of Iterations of Our Algorithms Average Number of Iterations (a) Mass Error Ranges from.1 to.4 Da Losak Losau Average Number of Iterations (b) Sequence Coverage Ranges from.1 to.4 1 Losak Losau Mass Error (Da) Sequence Coverage (c) Noise Level Ranges from 1% to 7% (d) Protein Number Ranges from 1 to 7 Average Number of Iterations Losak Losau Average Number of Iterations Losak Losau Noise Level (%) Number of Proteins 24 / 3

30 Experiments Simulation Study Sensitivity to Initialization 25 / 3

31 Experiments Simulation Study Sensitivity to Initialization Average Precision/Recall/F1 Measure (%) (a) Mass Error Ranges from.1 to.4 Da Mass Error (Da) Average Precision/Recall/F1 Measure (%) (b) Sequence Coverage Ranges from.1 to Sequence Coverage Average Precision/Recall/F1 Measure (%) (c) Noise Level Ranges from 1% to 7% Noise Level (%) Average Precision/Recall/F1 Measure (%) (d) Protein Number Ranges from 1 to Number of Proteins 25 / 3

32 Experiments Simulation Study The Effect of Filtering in Losau 26 / 3

33 Experiments Simulation Study The Effect of Filtering in Losau Average Precision (%) Average Precision (%) Average Precision (%) Average Precision (%) (a1) Average Precision Before Filtering After Filtering Mass Error (Da) (b1) Average Precision Before Filtering After Filtering Sequence Coverage (c1) Average Precision Before Filtering After Filtering Noise Level (%) (d1) Average Precision Before Filtering After Filtering Number of Proteins Average Recall (%) Average Recall (%) Average Recall (%) Average Recall (%) (a2) Average Recall Before Filtering After Filtering Mass Error (Da) (b2) Average Recall Before Filtering After Filtering Sequence Coverage (c2) Average Recall Before Filtering After Filtering Noise Level (%) (d2) Average Recall Before Filtering After Filtering Number of Proteins Average F1 Measure (%) Average F1 Measure (%) Average F1 Measure (%) Average F1 Measure (%) (a3) Average F1 Measure Before Filtering After Filtering Mass Error (Da) (b3) Average F1 Measure Before Filtering After Filtering Sequence Coverage (c3) Average F1 Measure Before Filtering After Filtering Noise Level (%) (d3) Average F1 Measure Before Filtering After Filtering Number of Proteins 26 / 3

34 Experiments Real Data Outline 1 Introduction 2 Method 3 Experiments Simulation Study Real Data 4 Conclusion 27 / 3

35 Experiments Real Data Results on Real Data Here we use a mixture of 49 standard human proteins in the ABRF sprg26 study. Table: Identification performance and running time of different algorithms on the real MS data. Here the number of reported proteins for SPA, Subtraction, and Losak is 49, i.e. the number of ground-truth proteins. Algorithms Precision Recall F1-Measure Running Time(s) SPA 24% 24% 24% 7.9 Subtraction 43% 43% 43% 24. Losak 67% 67% 67% 21.2 Losau 61% 71% 66% / 3

36 Conclusion Conclusion and Future Work Optimization-based PMF methods have great potential for protein mixture identification, especially in the analysis of low-abundance proteins, whose peptide digestion results are less likely to be covered by the peptide sequencing method. We like to combine MS and MS/MS data to further improve protein identification accuracy and robustness. 29 / 3

37 Conclusion Thank you! 3 / 3