Recursive Partitioning Method Research Articles

A Partitioning Deletion/Substitution/Addition Algorithm for Creating Survival Risk Groups

Accurately assessing a patient's risk of a given event is essential in making informed treatment decisions. One approach is to stratify patients into two or more distinct risk groups with respect to a specific outcome using both clinical and demographic variables. Outcomes may be categorical or continuous in nature; important examples in cancer studies might include level of toxicity or time to recurrence. Recursive partitioning methods are ideal for building such risk groups. Two such methods are Classification and Regression Trees (CART) and a more recent competitor known as the partitioning Deletion/Substitution/Addition (partDSA) algorithm, both of which also utilize loss functions (e.g., squared error for a continuous outcome) as the basis for building, selecting, and assessing predictors but differ in the manner by which regression trees are constructed. Recently, we have shown that partDSA often outperforms CART in so-called "full data" settings (e.g., uncensored outcomes). However, when confronted with censored outcome data, the loss functions used by both procedures must be modified. There have been several attempts to adapt CART for right-censored data. This article describes two such extensions for partDSA that make use of observed data loss functions constructed using inverse probability of censoring weights. Such loss functions are consistent estimates of their uncensored counterparts provided that the corresponding censoring model is correctly specified. The relative performance of these new methods is evaluated via simulation studies and illustrated through an analysis of clinical trial data on brain cancer patients. The implementation of partDSA for uncensored and right-censored outcomes is publicly available in the R package, partDSA.

Prediction Models of P-Glycoprotein Substrates Using Simple 2D and 3D Descriptors by a Recursive Partitioning Approach

E-mail: kyn@youai.co.krReceived October 31, 2011, Accepted December 21, 2011P-gp (P-glycoprotein) is a member of the ATP binding cassette (ABC) family of transporters. It transportsmany kinds of anticancer drugs out of the cell. It plays a major role as a cause of multidrug resistance (MDR).MDR function may be a cause of the failure of chemotherapy in cancer and influence pharmacokineticproperties of many drugs. Hence classification of candidate drugs as substrates or nonsubstrate of the P-gp isimportant in drug development. Therefore to identify whether a compound is a P-gp substrate or not, in silicomethod is promising. Recursive Partitioning (RP) method was explored for prediction of P-gp substrate. A setof 261 compounds, including 146 substrates and 115 nonsubstrates of P-gp, was used to training and validation.Using molecular descriptors that we can interpret their own meaning, we have established two models forprediction of P-gp substrates. In the first model, we chose only 6 descriptors which have simple physicalmeaning. In the training set, the overall predictability of our model is 78.95%. In case of test set, overallpredictability is 69.23%. Second model with 2D and 3D descriptors shows a little better predictability (overallpredictability of training set is 79.29%, test set is 79.37%), the second model with 2D and 3D descriptors showsbetter discriminating power than first model with only 2D descriptors. This approach will be used to reduce thenumber of compounds required to be run in the P-gp efflux assay.Key Words : ADME prediction, P-Glycoprotein, Recursive partitioning, 2D Descriptors, 3D DescriptorsIntroductionAbsorption, distribution, metabolism, excretion, and toxi-city (ADMET) properties are very important in the drugdiscovery.

Structure-Activity Relationship Modeling for Predicting Interactions with Pregnane X Receptor by Recursive Partitioning

Pregnane X receptor (PXR) is a ligand-activated nuclear factor that upregulates the expression of proteins involved in the detoxification and clearance of xenobiotics, primarily cytochrome P450 3A4 (CYP3A4). Structure-activity relationship (SAR) analysis of PXR agonists is useful for avoiding unwanted pharmacokinetics due to drug-drug interactions. To perform large-scale ligand-based SAR modeling, we systematically collected information on chemical-PXR interactions from the PubMed database by using the text mining system we developed, and merged it with screening data registered in the PubChem BioAssay database and other published data. Curation of the data resulted in 270 human PXR agonists and 248 non-agonists. After the entire data set was divided into training and testing data sets, the training data set comprising 415 data entries (217 positive and 198 negative instances) was analyzed by a recursive partitioning method. The classification tree optimized by a cross-validation pruning algorithm gave an accuracy of 79.0%, and, for the external testing data set, could correctly classify PXR agonists and non-agonists at an accuracy of 70.9%. Descriptors chosen as splitting rules in the classification tree were generally associated with electronic properties of molecules, suggesting they had an important role in the modes of interaction.

Trash to treasure: leaf-cutting ants repair nest-mound damage by recycling refuse dump materials

Animals adjust their behaviors in response to changing environmental conditions because the costs and benefits of such behaviors change as conditions change. The reuse of materials from waste (i.e., recycling) rarely occurs in social insects because it may imply significant health risks and behavioral difficulties. However, the benefit of reusing may exceed its costs under certain circumstances. For the first time, we document that ants refuse materials to repair nest-mound damage. We conducted a series of field measurements and experiments to test the hypotheses that fluctuations in this behavior in the leaf-cutting ant Acromyrmex lobicornis depend on 1) seasonal changes in the tendency to reject refuse (a proxy of changes in their pathogen levels), and/or 2) seasonal foraging restrictions. We 1) measured the rejection of foraging ants toward experimental refuse piles among seasons and 2) analyzed how mound condition, temperatures of fungus chamber and soil surface, and foraging activity explained this behavior using a classification tree, a powerful recursive partitioning method. Foraging ants showed similar rejection levels toward refuse piles in different seasons. Colonies repaired mound damage with refuse materials only during the hottest season and when they had low foraging rates, suggesting that ants recycle their refuse by a hierarchical set of decisions dependent on the risk of fungal damage and foraging restrictions. Repairing the mounds is essential during summer, when temperatures inside damaged mounds are lethal to their fungus cultures. However, these high temperatures also restrict the foraging activity, reducing the collection of building materials. Thus, colonies with lower foraging rates apparently use their refuse to repair mounds because this substrate requires less searching and carrying time. The use of refuse did not affect the colony growth rate. This illustrates how ants integrate information about food, hygienic and nest conditions through a novel and plastic behavior: recycling of their discarded materials.

Subgroup identification based on differential effect search—A recursive partitioning method for establishing response to treatment in patient subpopulations

We propose a novel recursive partitioning method for identifying subgroups of subjects with enhanced treatment effects based on a differential effect search algorithm. The idea is to build a collection of subgroups by recursively partitioning a database into two subgroups at each parent group, such that the treatment effect within one of the two subgroups is maximized compared with the other subgroup. The process of data splitting continues until a predefined stopping condition has been satisfied. The method is similar to 'interaction tree' approaches that allow incorporation of a treatment-by-split interaction in the splitting criterion. However, unlike other tree-based methods, this method searches only within specific regions of the covariate space and generates multiple subgroups of potential interest. We develop this method and provide guidance on key topics of interest that include generating multiple promising subgroups using different splitting criteria, choosing optimal values of complexity parameters via cross-validation, and addressing Type I error rate inflation inherent in data mining applications using a resampling-based method. We evaluate the operating characteristics of the procedure using a simulation study and illustrate the method with a clinical trial example.

Recursive Partitioning Approaches to Modeling Genetics as a Source of Treatment Effect Heterogeneity

Pharmacogenomics has held the promise of delivering, “personalized medicine,” by identifying genetic determinants of variations in treatment response. While pharmacogenomics has successfully identified a handful of functional polymorphisms (relatively common variations in a single gene that affect the activity of a drug), the promise of personalized medicine has gone largely unfulfilled. Genetics researchers have come to realize that more complex interactions of genes, or of genes and environment, are more likely to underlie treatment effect heterogeneity than individual genes alone. The classical biostatistics approach to clinical trial data analysis requires that subgroups be defined ex ante. But because of the sheer number of genetic and environmental predictors and the potential for multi-level interactions, a curse of dimensionality becomes readily apparent in experimental design. Even post hoc subgroup analysis will suffer from this curse. Recursive partitioning methods (including classification and regression trees and random forests) repeatedly split observations into groups with maximally dissimilar outcomes according to the value of predictor variables. Their recursive nature allows for the exploration of multi-way interactions. However, existing, “greedy,” tree-building algorithms such as CART may fail to find highly important interactions when the direct effects of individual genes or environmental variables are small. While ensemble-based approaches such as bagging or random forests may more often correctly classify individuals, the required aggregation or averaging of classifiers makes interpretation of the models difficult. In this paper, an alternative approach is described that constructs and evaluates sets of random trees, scoring repeatedly-encountered interactions in proportion to their predictive ability and to the performance of the trees in which the interactions appear. Monte Carlo experiments are used to estimate the value of information generated by including subgroup effects identified by the algorithm in a simulated cost-effectiveness analysis of clinical trial data.

Recursive partitioning of resistant mutations for longitudinal markers based on a U-type score

Development of human immunodeficiency virus resistance mutations is a major cause of failure of antiretroviral treatment. We develop a recursive partitioning method to correlate high-dimensional viral sequences with repeatedly measured outcomes. The splitting criterion of this procedure is based on a class of U-type score statistics. The proposed method is flexible enough to apply to a broad range of problems involving longitudinal outcomes. Simulation studies are performed to explore the finite-sample properties of the proposed method, which is also illustrated through analysis of data collected in 3 phase II clinical trials testing the antiretroviral drug efavirenz.

Identifying the 'inorganic gene' for high-temperature piezoelectric perovskites through statistical learning.

This paper develops a statistical learning approach to identify potentially new high-temperature ferroelectric piezoelectric perovskite compounds. Unlike most computational studies on crystal chemistry, where the starting point is some form of electronic structure calculation, we use a data-driven approach to initiate our search. This is accomplished by identifying patterns of behaviour between discrete scalar descriptors associated with crystal and electronic structure and the reported Curie temperature (TC) of known compounds; extracting design rules that govern critical structure–property relationships; and discovering in a quantitative fashion the exact role of these materials descriptors. Our approach applies linear manifold methods for data dimensionality reduction to discover the dominant descriptors governing structure–property correlations (the ‘genes’) and Shannon entropy metrics coupled to recursive partitioning methods to quantitatively assess the specific combination of descriptors that govern the link between crystal chemistry and TC (their ‘sequencing’). We use this information to develop predictive models that can suggest new structure/chemistries and/or properties. In this manner, BiTmO3–PbTiO3 and BiLuO3–PbTiO3 are predicted to have a TC of 730°C and 705°C, respectively. A quantitative structure–property relationship model similar to those used in biology and drug discovery not only predicts our new chemistries but also validates published reports.

Using dwell time as an implicit measure of usefulness in different task types

In this study, we attempted to use dwell time on content pages as a predictor of document usefulness, and evaluated the prediction performance in different types of tasks. A user study was conducted to address this research problem. A total of 32 participants conducted searches associated with 4 different tasks, varying along several task type dimensions. In the study, participants were asked to save documents that were useful for the tasks during the searches. Participants' information-seeking activities were recorded. Using the logged dataset, we used the recursive partitioning method to identify the best threshold (cutoff point) of dwell time to predict which pages would be useful, i.e., be saved by the users. We examined the differences in the threshold and compared the prediction performance for each type of task. Our results indicate that the threshold should be developed according to the type of task users are working on. While the prediction performance using dwell time as the only indicator of document usefulness is acceptable for Factual tasks, it is not good for Intellectual tasks. Our results have implications on IR system design.

Recursively Partitioned Static IP Router Tables

We propose a method-recursive partitioning-to partition a static IP router table so that when each partition is represented using a base structure, such as a multibit trie or a hybrid shape shifting trie, there is a reduction in both the total memory required for the router table as well as in the total number of memory accesses needed to search the table. The efficacy of recursive partitioning is compared to that of the popular front-end table method to partition IP router tables. Our proposed recursive partitioning method outperformed the front-end method of all our test sets.

Analysis of Dose-Volume Metrics Associated with Radiation Pneumonitis after Stereotactic Body Radiotherapy for Lung Cancer

The objective was to find dose-volume factors associated with radiation pneumonitis (RP) after stereotactic body radiotherapy (SBRT) for lung cancer. Seventy-four patients who underwent SBRT for primary lung cancer were analyzed in the present study. Fifty-five were male, and nineteen were female. A median of age was 77 years old (range, 63-88). Chemotherapy was not administered unless disease progression was confirmed. Prescription dose for SBRT was uniformly 48 Gy in 4 fractions at the isocenter. Irradiation was performed with 6-MV X-ray beams, which were shaped into the planning target volume (PTV) plus 5-mm margins, from a linear accelerator (Clinac 2300 C/D; Varian). Dose distribution was re-calculated with the analytical anisotropic algorithm in Eclipse (Varian). RP was graded according to the CTCAE version 3, and high-grade RP was herein defined as Grade 2 or worse. Optimal cut points dividing the patient population into two subgroups based on incidence of high-grade RP were searched using the recursive partitioning method in the following dose-volume metrics: PTV volume (cc), mean lung dose (Gy), V5, V10, V15, V20, V25, V30, V35, and V40, where Vd was defined as a relative volume (%) of the lung other than PTV which received more than a threshold dose d Gy. RP Grade 2 or worse was observed in 15 patients (20.2%), including 1 patient of Grade 3, with a median follow-up duration of 31 months. Optimal cut points for pulmonary doses were found in V25 and V20. These factors were highly correlated with each other (R = 0.99), and V25 was more significant than V20. High-grade RP rate was 14.8% in the patients with V25 less than 4.2%, while the rate was 46.2% in the remainder (p = 0.02). PTV volume was found as the other significant factor. High-grade RP rate was significantly lower in the group with PTV volume less than 37.7 cc compared with the larger PTV group (11.1% vs. 34.5%, p = 0.02). Correlation between lung V25 and PTV volume was not strong (R = 0.58). When dividing the patient population into three subgroups: (1) PTV < 37.7 cc; (2) PTV ≥ 37.7 cc and V25 < 4.2%; and (3) PTV ≥ 37.7 cc and V25 ≥ 4.2%, incidence rates of high-grade RP were 11.1%, 23.5%, and 50.0%, respectively (p = 0.01). Lung V25 and PTV volume were suggested as significant factors associated with RP after SBRT.

The mean does not mean as much anymore: finding sub-groups for tailored therapeutics

The genomics revolution is still in its infancy, and there is much to learn about how to transform biological knowledge into useful medicines to further public health. At the bedside, we are asking how and why individual patients respond to different drug treatments in different ways. In addition to genetic mechanisms, there are many clinical markers (e.g. medical history, disease severity) as well as social/environmental factors (e.g. smoking habits) that can be used to identify who may or may not respond to treatment. This issue has some considerable statistical complexity, and different approaches to the analysis of clinical trials may yield more interesting insights into the problem. Novel applications of statistical methods will be discussed, and examples will be used to demonstrate sub-group identification. In order to evaluate many potential predictors of response, we use recursive partitioning methods to identify predictor variables and their cut-off values to define sub-groups of patients with differential treatment response. Validation of this variable/model selection approach was done using independent data from other clinical trials. In one example, a classification tree was developed using baseline measures to define important sub-groups of patients that responded much better than the overall mean response in the study. In a second example, a classification tree was built based on measures of response early in treatment to predict longer-term responders and nonresponders. Limitation Classification algorithms can be prone to over-fitting, and validation of results is an important consideration. Obviously, analyses are limited by the available predictor variables. Using classification trees proved to be very useful in evaluating large numbers of potential predictors to find sub-groups of patients with exceptional response. The method is easy to use, and clinicians can easily interpret and implement results. This approach can be helpful in tailoring treatments to individual patients.

Adaptive partitioning schemes for bipartite ranking

Recursive partitioning methods are among the most popular techniques in machine learning. The purpose of this paper is to investigate how to adapt this methodology to the bipartite ranking problem. Following in the footsteps of the TreeRank approach developed in Clémençon and Vayatis (Proceedings of the 2008 Conference on Algorithmic Learning Theory, 2008 and IEEE Trans. Inf. Theory 55(9):4316–4336, 2009), we present tree-structured algorithms designed for learning to rank instances based on classification data. The main contributions of the present work are the following: the practical implementation of the TreeRank algorithm, well-founded solutions to the crucial issues related to the splitting rule and the choice of the “right” size for the ranking tree. From the angle embraced in this paper, splitting is viewed as a cost-sensitive classification task with data-dependent cost. Hence, up to straightforward modifications, any classification algorithm may serve as a splitting rule. Also, we propose to implement a cost-complexity pruning method after the growing stage in order to produce a “right-sized” ranking sub-tree with large AUC. In particular, performance bounds are established for pruning schemes inspired by recent work on nonparametric model selection. Eventually, we propose indicators for variable importance and variable dependence, plus various simulation studies illustrating the potential of our method.

Once‐Daily Gentamicin Dosing in Children with Febrile Neutropenia Resulting from Antineoplastic Therapy

To evaluate an existing once-daily gentamicin dosing guideline in children with febrile neutropenia resulting from antineoplastic therapy and, if necessary, to develop a new simulated dosing guideline that would achieve pharmacokinetic targets more reliably after the first dose. Pharmacokinetic analysis of data from a retrospective medical record review. Hematology-oncology unit of a university-affiliated pediatric hospital in Canada. One hundred eleven patients aged 1-18 years who received once-daily gentamicin between April 2006 and January 2008 for the treatment of febrile neutropenia resulting from antineoplastic therapy, and who had plasma gentamicin concentrations determined after their first dose. Demographic data, gentamicin dosing information, blood sampling times, and plasma gentamicin concentrations were noted. Plasma gentamicin concentrations were determined at approximately 3 and 6 hours after the start of the 30-minute infusion of the first dose. Pharmacokinetic parameters were calculated according to standard first-order, one-compartment equations. The proportion of children who achieved pharmacokinetic targets after the first gentamicin dose was used as a measure of dosing guideline performance; the guideline achieved maximum concentration (C(max)) values below the target range (20-25mg/L) in 51% of patients. Ideal dosing guidelines were then developed using the mean dose required to achieved a C(max) of 23 mg/L for each patient. Univariate analysis or the Student t test was used to determine the existence of significant relationships between pharmacokinetic parameters and patient age and sex. The recursive binary partitioning method was used to determine critical values of age for dosage guideline development; analysis of variance was then used to compare the different levels obtained after use of this technique. Simulated administration of once-daily gentamicin in the following doses achieved a C(max) within or above target in 73% of patients: 1 year to < 6 years, 10.5mg/kg/dose; girls > or = 6 years, 9.5mg/kg/dose; and boys > or = 6 years, 7.5mg/kg/dose. Doses were based on actual body weight for children who weighed less than 125% of ideal body weight or based on effective body weight for children 125% or more of ideal body weight. The initial gentamicin dosing guidelines were not effective in achieving C(max). The new proposed dosing guidelines are predicted to achieve a C(max) within or above the target range in almost three quarters of patients. Subsequent dosing should be tailored according to plasma gentamicin concentrations.

Risk haplotype pattern discovery for gene mapping by recursive partitioning method based on weighted classification trees

In this paper, we present a combinatorial approach based on recursive weighted longest prefix trees (RWLPT) for mining a massive genetic marker data. Given a case and a control chromosome dataset, we develop a fast recursive permutation algorithm for building an RWLPT. This algorithm provides an optimal solution to find a privileged order of classifiers for constructing optimum haplotype clusters and biclusters. Using RWLPT model, extracting haplotype clusters and haplotype patterns will be performed in linear time on the number of markers. Employing the measure of linkage disequilibrium and the statistical test, we examine the case-control association analysis for identifying risk haplotype patterns associated with the disease permitting likely location of disease susceptibility genes. We demonstrate that the case-control association test based on haplotype patterns reduces degrees of freedom and number of tests while increasing the power of disease-haplotype association analysis. The experimental results through the real published SNP dataset show that our approach tends to be more effective and powerful than the sliding window approach, classification and regression trees and single marker analysis.

An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests.

Recursive partitioning methods have become popular and widely used tools for nonparametric regression and classification in many scientific fields. Especially random forests, which can deal with large numbers of predictor variables even in the presence of complex interactions, have been applied successfully in genetics, clinical medicine, and bioinformatics within the past few years. High-dimensional problems are common not only in genetics, but also in some areas of psychological research, where only a few subjects can be measured because of time or cost constraints, yet a large amount of data is generated for each subject. Random forests have been shown to achieve a high prediction accuracy in such applications and to provide descriptive variable importance measures reflecting the impact of each variable in both main effects and interactions. The aim of this work is to introduce the principles of the standard recursive partitioning methods as well as recent methodological improvements, to illustrate their usage for low and high-dimensional data exploration, but also to point out limitations of the methods and potential pitfalls in their practical application. Application of the methods is illustrated with freely available implementations in the R system for statistical computing.

Let's not forget tautomers.

A compound exhibits tautomerism if it can be represented by two structures that are related by an intramolecular movement of hydrogen from one atom to another. The different tautomers of a molecule usually have different molecular fingerprints, hydrophobicities and pKa’s as well as different 3D shape and electrostatic properties; additionally, proteins frequently preferentially bind a tautomer that is present in low abundance in water. As a result, the proper treatment of molecules that can tautomerize, ~25% of a database, is a challenge for every aspect of computer-aided molecular design. Library design that focuses on molecular similarity or diversity might inadvertently include similar molecules that happen to be encoded as different tautomers. Physical property measurements might not establish the properties of individual tautomers with the result that algorithms based on these measurements may be less accurate for molecules that can tautomerize—this problem influences the accuracy of filtering for library design and also traditional QSAR. Any 2D or 3D QSAR analysis must involve the decision of if or how to adjust the observed Ki or IC50 for the tautomerization equilibria. QSARs and recursive partitioning methods also involve the decision as to which tautomer(s) to use to calculate the molecular descriptors. Docking virtual screening must involve the decision as to which tautomers to include in the docking and how to account for tautomerization in the scoring. All of these decisions are more difficult because there is no extensive database of measured tautomeric ratios in both water and non-aqueous solvents and there is no consensus as to the best computational method to calculate tautomeric ratios in different environments.

Lifecourse Priorities Among Appalachian Emerging Adults: Revisiting Wallace's Organization of Diversity.

We examine how social demographics (gender, age, or race-ethnicity), census tract characteristics, and family environment during childhood relate to variability in the lifecourse priorities of 344 Cherokee and white youth during emerging adulthood (age 19-24). Analyses were performed using recursive partitioning and random forest methods to examine determinants of prioritizing education, family formation, economic establishment, self characteristics and close relationships, youth independence, conspicuous consumption, and community reliance. Overall, characteristics of census tracts were the most common and influential predictors of lifecourse priorities. Childhood family poverty, parental relationship problems, parental crime, and stressful life events were also important predictors. Race-ethnicity or cultural group (Cherokee vs. white), age, and gender were relatively unimportant. At this developmental stage and in this population, community characteristics and childhood family experiences may be better proxies for developmental settings (and resulting enculturated values and preferences) than social demographic variables (e.g., ethnicity or gender).

Proposed Metrics for the Determination of Rheumatoid Arthritis Outcome and Treatment Success and Failure

Patients with rheumatoid arthritis (RA) and their physicians often disagree as to the success of RA treatment or RA outcomes. However, guidelines (such as EULAR criteria for DAS scores) are heavily weighted toward joint counts and laboratory tests, and no guidelines exist for patient reported outcomes. Our aims were (1) to provide a patient-based definition of successful RA outcome or of treatment success and failure; (2) to describe the characteristics of patients meeting this definition; (3) to describe how external states such as disability and comorbidity influence definitions of health outcome; and (4) to derive surrogate-measure cutpoints for the definition. A total of 20,268 patients with RA (5132 without comorbidity) were studied by recursive partitioning and regression methods to determine best dividing points between RA treatment and outcome success and non-success using 0-10 visual analog scales (VAS) for patient global assessment, pain, fatigue, and RA activity, and a Health Assessment Questionnaire (HAQ) scale. 14.5% of all patients and 22.9% of those without comorbidity were very satisfied with their health (success). Patient global at a level<or=1.25 best separated success from failure. Mean and median scores for those who were very satisfied were HAQ (0-3 scale) 0.36, 0.12; pain (0-10) 1.1, 0.5; global (0-10) 0.9, 0.5; and fatigue (0-10) 1.5, 1.0. VAS scores increased by approximately 0.5 units for each comorbid condition. Patient global at a level<or=1.25 best separates patients who are very satisfied with their health from those not very satisfied, regardless of the presence of comorbidity. All scores increase with increasing comorbidity, which must be accounted for when assessing individual patients. Values identified here suggest patients require better outcomes than are found in patients who are in Disease Activity Score-28 remission or OMERACT low disease activity states.

Diverse Patterns of T Cell Reactivity Directed against Multiple Y Chromosome-Derived, HLA a*0201 Restricted, Peptides in Patients with chronic Gvhd

Minor histocompatibility antigens (mHA) are targets of graft versus host disease (GVHD) and graft versus leukemia (GVL) responses. In male patients with female donors, Y chromosome-encoded mHA are known to be highly immunogenic but few MHC class I presented epitopes have been identified and T cell responses to HY mHA have not been examined in large cohorts of patients. We recently developed a novel method for rapid identification of putative mHA based on in-silico prediction of potential MHC class I restricted peptides. Criteria for HY mHA prediction included the following: Y chromosome encoded gene sequence; high affinity binding to HLA A*0201 (≤100nM); lack of identical sequence in homologous X gene; validated protein expression not restricted to male-specific tissues; and amino acid disparities between Y and X chromosome homologues predicted to be immunogenic. In this study, 43 peptide epitopes (9–10 amino acids) representing 5 Y-encoded proteins (DBY, SMCY, UTY, PCDH11Y, USP9Y) were identified. The only previously known Y-encoded (SMCY) HLA A*0201 restricted mHA (FIDSYICQV) ranked first in the list of predicted epitopes. Two peptides failed synthesis and 41 HY peptides were tested for T cell reactivity in post-transplant samples from 21 male patients with female donors (M-F), 9 male patients with male donors (M-M) and 19 healthy donors (7 males; 12 females). All patients and donors were HLA A*0201+ and all patients had chronic GVHD. T cell reactivity was determined by ELISPOT. Peripheral blood mononuclear cells were stimulated with individual peptides for 7 days. IFN-g secreting T cells were enumerated after 18 hour restimulation with HLA A*0201 T2 cells pulsed with the same peptide. Known HLA A*0201 restricted peptides from EBV, CMV, HIV and HCV were used as positive and negative controls. T cell responses to individual peptides could be blocked by anti human HLA class I antibodies but not by anti HLA class II antibodies. Using the recursive partitioning method, a minimum of 20 spots/2×105 PBMC above background level was identified as an optimal cut-off point to classify positive responses for the majority of peptides. T cell responses to some HY peptides were also detected in M-M patients and healthy donors, but responses were most frequent and of greater magnitude in M-F patients. The median number of positive peptide responses per patient was 8 in the M-F patient group, 1 in M-M patients and 2 in healthy donors (p=0.002 for M-F vs M-M, p=0.02 for M-F vs healthy donors). All M-F patients responded to at least 1 HY peptide whereas 3/9 (33%) M-M patients and 7/19 (37%) healthy donors did not respond to any peptides. 18/21 (86%) M-F patients responded to 3 or more peptides but this level of reactivity was only detected in 2/9 (22%) M-M patients and 7/19 (37%) healthy donors. High frequency responses (>50 spots/2×105 PBMC) were detected in 19/21 (91%) M-F patients but only in 2/9 (22%) M-M patients, 1/7 (14%) healthy males and 4/12 (33%) healthy females. T cell responses were detected in at least 1 M-F patient for 36 of 41 peptides and responses were detected against all 5 Y-encoded proteins. However, a subset of 20 peptides appeared to be highly immunogenic with T cells responses detected in >25% of M-F patients. No single HY peptide elicited responses in all M-F patients including the previously known HY mHa, which was only positive in 40% of M-F patients. In fact 9 other peptides derived from DBY (1), SMCY (3), UTY (3), PCDH11Y (1) and USP9Y (1) elicited responses in 40–71% of M-F patients. Each of these 10 peptides elicited high frequency responses (>50 spots/2×105 PBMC) in at least 3 M-F patients but not in any M-M patients or healthy males. Within the M-F patients, the frequency of response was not associated with severity of cGVHD, underlying hematologic disease, age, stem cell source, transplant conditioning regimen or donor type. There was a correlation between time post transplant and the number of HY peptide responses (r=0.53, p=0.002 for all F-M and M-M patients combined). The functional application of bioinformatic models represents a new approach for identifying large numbers of novel HY peptides and assessing T cell responses after transplantation. These studies demonstrate a highly diverse T cell response despite identical mismatch and HLA type. Extending this method to other HLA alleles and to autosomal genetic disparities will improve our understanding of the role of mHA in GVL and GVHD after allogeneic transplantation.