1. Mince breast tissue into small (1-2mm) pieces using a sterile razor blade and scissors or mince

Transcription

1 Purification of all cell types from breast tissue 1. Mince breast tissue into small (1-mm) pieces using a sterile razor blade and scissors or mince using food processor.. Digest tissue in ~5x tissue volume of DMEM/F1 medium containing mg/ml collagenase I (Sigma C0130), mg/ml hyaluronidase (Sigma H3506) at 37 o C with constant agitation. The digestion time should be limited to 1- hours. 3. Spin down cells in 50 ml tubes and centrifuge at 3,000 rpm for 10 minutes. Then wash with 1x with HBSS (Hank s Balanced Salt Solution) and centrifuge again. For invasive tumors go to step 5, since invasive cancers do not form organoids. For normal and DCIS breast tissue for the purification of epithelial and myoepithelial cells use the organoid fraction (will remain in the filter), while in the case of invasive cancer all cell types are in one fraction. 4. For normal and DCIS breast tissue: resuspend pellet in 10-0 mls of HBSS and filter through a 500, a 50, a 100 µ, and then a 40 µ cell strainer (500, 50 and 0 µ filter meshes are from Tetko, cat # , and respectively, while the 100 µ and 40 µ cell stariners are from Fisher cat #: and , respectively). Collect organoids which remain in 50 µ and 100 µ strainer, and pour flow through into the 40 µ strainer. Combine all organoids in 40 µ strainer with organoids from 100 µ strainer and 50 µ mesh. Whatever flows through the 40 µ restrainer filter through a 0 µ filter to ensure nearly complete removal of breast ducts. Collect organoids as well as flow through (the flow through is the stroma fraction) by cf. With the stromal cell fraction go directly to step Resuspend cells (invasive cancers from step 3) or organoid (normal and DCIS breast tissue) pellet in 10x vol. trypsine/edta, incubate at 37 o C for 5 min. then dilute with ice cold 1

2 medium+10%fbs, then filter through a 100, 40, and 0 µ filter mesh, save flow through (these should to be single cells) and collect cells by cf. 3,000 rpm for 5 min. 6. Percoll gradient cf. (Percoll-Pharmacia cat# , mix: 9.5 ml percoll+1.5ml 10xPBS, 4 ml water, layer cells in 10 ml PBS underneath the Percoll, cf for 3,000 rpm 15 min.) to remove dead cell/erythrocytes, you can see a cloudy cell layer in the upper ~10 ml, gently collect this with a 10 ml pipette and wash cells in HBSS 1x to remove Percoll, collect them by cf. at 3,000 rpm for 5 min. 7. Resuspend cell pellet in µl (volume depends on the total number of cells, you don t want to have more than ~1- million cells/00 µl) PBE (PBE is PBS, 0.5 % BSA and mm EDTA). (a) In the case of the stromal cell fraction from normal and DCIS breast tissue go directly to step 8. (b) Purification of epithelial cells from the organoid (normal or DCIS tissue) or from the total (invasive cancers) cell fraction: add 50 µl (or 10-0 µl for small tumor tissue) pre-washed BerEP4 beads to the cells (Dynal cat# 161.0), incubate for 8 min. on ice with occasional flicking of the tube then dilute to 3-5 ml volume with PBE, capture bound cells on magnet for min. Captured fraction contains the luminal EPITHELIAL CELLS. Wash bound cells 3xs00 µl PBE then freeze captured cells immediately on dry ice. Take supernatant into new eppendorf tubes and repeat magnetic capturing to ensure that all bead bound cells are removed. Take supernatant and in the case of normal or DCIS organoids go directly to step step For the stromal (normal or DCIS tissue) and total (invasive tumor tissue) cell fraction collect cells by cf. and resuspend in µl PBE the same way as before and add µl of 1:1 mix of pre-washed CD45 and CD15 beads (Dynal cat# and , respectively, incubate 8 min. on ice with occasional flicking of the tube. then Dilute and capture cells as before. The captured

3 cells are the LEUKOCYTE FRACTION. Wash bound cells 3xs00 µl PBE then freeze captured cells immediately on dry ice. Go directly to step From step 7-take sup. pellet cells by cf. and resuspend in µl PBE the same way as before and add 50 µl of pre-washed/bound CD10 beads [bind 0 µl antibody (Dako cat#m077) to 50 µl beads (Dynal Pan mouse IgG cat# 110.3) in 100 µl PBE and incubate on ice for at least 1 hour] to the cells and incubate for 10 min. on ice with occasional flicking of the tube. Dilute/capture cells as before. The captured cells are the MYOEPITHELIAL CELL FRACTION. Wash bound cells 3xs00 µl PBE then freeze captured cells immediately on dry ice. 10. Take supernatant from step 8 and pellet cells by cf. and resuspend in µl PBE the same way as before and add a 10-0 µl of P1H1 antibody (Chemicon cat#mab16985) to the cells and incubate on ice for min. then add 10-0 µl pre-washed anti-mouse antibody coupled magnetic beads (Dynal cat# 110.3) and incubate for 10 min on ice with occasional flicking of the tube. Dilute/capture cells as before. This will be the ENDOTHELIAL CELL FRACTION. Wash bound cells 3xs00 µl PBE then freeze captured cells immediately on dry ice. 11. Take sup. pellet cells by cf. and save this leftover fraction. Although this is likely to contain multiple cell types since the removal of these is not 100% efficient, it is highly enriched for stromal fibroblasts. 3

4 Significance and Clustering Analysis of SAGE Data Poisson Assumption In a SAGE experiment, a set of transcripts from a cell or tissue is sampled for tag extraction. Considering the numerous types of transcripts present in a cell or tissue and the small probability of sampling a particular type of transcript, we assume that the number of sampled transcripts of each type is approximately Poisson distributed. Statistically, when this actual sampling process is random enough, Poisson would be the most practical and reasonable assumption compared to other probability models. This assumption leads to the following probability models used for significance analysis and clustering analysis of SAGE data. Significance Analysis of SAGE data Based on Poisson assumption, we developed a significance analysis algorithm ( SA algorithm ) to detect differentially expressed tags in SAGE data. The input to the SA algorithm is a tab-delimited file containing multiple sage libraries. The SA algorithm can simultaneously compare two or more SAGE libraries. The output of SA algorithm is a set of p values of test for the significance of the difference in gene expression. Genes with significantly small p values are identified as differentially expressed across different libraries. The p values are calculated in the following way: Letting X i be the number of copies of tag i in library, we define three sums: 1. M i. = X i ; The sum of counts of tag i over all libraries.. M. = X i ; The sum of tag counts over all tags in library. i 3. M = X i i, ; The sum of tag counts over all tags and libraries. 4

5 Under the null hypothesis that there is no expression difference across libraries, M i. M. / M copies are then expected to be observed for tag i in library. Further, considering that the tags are extracted from a random sample of transcripts in cell, it is reasonable to assume X i is Poisson distributed with mean λ i = M i. M. / M. The Chi-square statistic is used to test the deviation of observed counts from expected counts: k ( X i λ i ) TS i =, =1 λ i where k is the number of libraries compared. When k is large or λ i is not small (greater than 5), TS i is approximately Chi-square distributed with degree of freedom of k-1 ( χ k 1 ), the SA algorithm calculates the p-values using the approximated χ k 1. However, when k and λ i are small, there is a large departure of TS i from χ k 1, the SA algorithm calculates exact p-value of observedts i based on the Poisson distribution of X i. Clustering Analysis of SAGE Data Let Y i (t) be the count of tag i at condition t, and Y i (t)~poisson(_ i (t)_ i ). Here _ i (t)_ i denotes the expected count for tag i at condition t under the Poisson process. The expected count _ i (t)_ i consists of two factors: _ i and _ i (t). _ i is the expected total count of transcript i (mapped to tag i) over all conditions; _ i (t) is the contribution of transcript i at condition t to the total count (_ i ) expressed by percentage. Thus λ i (t) = 1 when T conditions are considered. T t =1 5

6 The goal is to group together the transcripts with similar expression variation over different conditions; that is to cluster tags by their _ i (t)s. We thus assume that the tags within a cluster share the same _=(_(1), _(),, _(T)), where _ represents the cluster model. Further, based on the Poisson assumption and letting Y i =(Y i (1), Y i (),, Y i (T)), we have the following oint likelihood function for tags i 1, i,, i m within a cluster: exp( λ(t)θ ik )(λ(t)θ ik ) Y i (t ) m T k f (Y i1,y i,...,y im λ,θ i1,θ i,...,θ im ) =. (1) Y ik (t)! t =1 The maximum likelihood estimators of _s and _s are: θ ik = Yik (t), () t and λ (t) = m Y ik (t) = m θ ik m m Y ik (t). (3) Y ik (t) t Thus for a set of tags assumed to be in the same cluster, we can estimate the cluster model _ and the total count _ i of each tag by (). Further, we can use Chi-square test statistics to evaluate how well the observed tag count fits the estimated cluster model, which is to calculate S= k t (Y ik (t) λ(t)θ ik ) λ(t)θ ik. The larger the value of S is, the less likely that the tags share same patterns of expression. Using Chi-square test statistics, the penalty for deviation from large expected counts is much smaller than that for small expected counts. This is consistent with the oint likelihood function f (Y i1,y i,...,y im λ, ) under expected Poisson models, since a Poisson distribution has the property of mean=variance. We choose 6

7 Chi-square test statistics as distance for measuring deviation from expected models due to two reasons: 1). The derivation of oint likelihood is very time consuming, and thus the corresponding clustering algorithm is very slow and impractical for large datasets; ) Simulation study shows that the performance of clustering algorithm using Chi-square statistics is almost as good as using oint likelihood function directly. Based on the probability model described above, and K-means clustering method, we developed the PK clustering algorithm. K-means cluster algorithm [0] generates good clusters by specifying a desired number of clusters, say, K, and then assigns each obect to one of K clusters so as to minimize a measure of dispersion within the clusters. We modified K-means clustering algorithm by using the Chi-square statistics as similarity measurements instead of using the Pearson correlation or Euclidean distance or other similarity measurements. The modified algorithm is called the PK clustering algorithm. A brief description of the PK algorithm: 1. All SAGE tags are assigned at random to K sets. Estimate the total count θ of each tag by formula (). (0). Set cluster centers λ k by formula (3). (0) λk θ then represents the expected expression pattern of the tag in cluster k. Current iteration i=0. 3. In the ith iteration, assign each tag to the cluster with the nearest expected expression pattern. We use Chi-squared statistics to measure the distance between the observed expression pattern of tag (i) (Y ) and the expected expression pattern of tag in cluster k ( λ k θ ): t (i) (Y (t) λ k (t)θ ) (i) λ k (t)θ (i+1) 4. Set new cluster centers λ k. 5. Go to step 3 until convergence. 7

8 In total, if c() denotes the cluster number that tag is assigned to, the PK-clustering algorithm is to minimize the within-cluster dispersion: (Y (t) λ c( ) (t)θ ). λ c( ) (t)θ t Implementation The PK algorithm is implemented in both C++ and Java. The routines for the EM algorithm for reassigning cluster members are from the work by Michiel de Hoon et. al. in the Human Genome Center at the University of Tokyo. The SA algorithm is implemented in C++. Both algorithms here described are available from: 8