Assessing clinical response in early oncology development with a predictive biomarker.

In early oncology clinical trials there is often limited data for biomarkers and their association with response to treatment. Thus, it is challenging to decide whether a biomarker should be used for patient selection and enrollment. Most evidence about any potential predictive biomarker comes from preclinical research and, sometimes, clinical observations. How to translate the preclinical predictive biomarker data to clinical study remains an active field of research. Here, we propose a method to incorporate existing knowledge about a predictive biomarker - its prevalence, association with response and the performance of the assay used to measure the biomarker - to estimate the response rate in a clinical study designed with or without using the predictive biomarker. Importantly, we quantify the uncertainty associated with the biomarker and its predictability in a probabilistic model. This model estimates the distribution of the clinical response when a predictive biomarker is used to select patients and compares it to unselected cohort. We applied this method to two real world cases of approved biomarker-guided therapies to demonstrate its utility and potential value. This approach helps to make a data-driven decision whether to select patients with a predictive biomarker in early oncology clinical development.

Longitudinal multi-omics study of palbociclib resistance in HR-positive/HER2-negative metastatic breast cancer.

BACKGROUND: Cyclin-dependent kinase 4/6 inhibitor (CDK4/6) therapy plus endocrine therapy (ET) is an effective treatment for patients with hormone receptor-positive/human epidermal receptor 2-negative metastatic breast cancer (HR+/HER2- MBC); however, resistance is common and poorly understood. A comprehensive genomic and transcriptomic analysis of pretreatment and post-treatment tumors from patients receiving palbociclib plus ET was performed to delineate molecular mechanisms of drug resistance. METHODS: Tissue was collected from 89 patients with HR+/HER2- MBC, including those with recurrent and/or metastatic disease, receiving palbociclib plus an aromatase inhibitor or fulvestrant at Samsung Medical Center and Seoul National University Hospital from 2017 to 2020. Tumor biopsy and blood samples obtained at pretreatment, on-treatment (6 weeks and/or 12 weeks), and post-progression underwent RNA sequencing and whole-exome sequencing. Cox regression analysis was performed to identify the clinical and genomic variables associated with progression-free survival. RESULTS: Novel markers associated with poor prognosis, including genomic scar features caused by homologous repair deficiency (HRD), estrogen response signatures, and four prognostic clusters with distinct molecular features were identified. Tumors with TP53 mutations co-occurring with a unique HRD-high cluster responded poorly to palbociclib plus ET. Comparisons of paired pre- and post-treatment samples revealed that tumors became enriched in APOBEC mutation signatures, and many switched to aggressive molecular subtypes with estrogen-independent characteristics. We identified frequent genomic alterations upon disease progression in RB1, ESR1, PTEN, and KMT2C. CONCLUSIONS: We identified novel molecular features associated with poor prognosis and molecular mechanisms that could be targeted to overcome resistance to CKD4/6 plus ET. TRIAL REGISTRATION: ClinicalTrials.gov, NCT03401359. The trial was posted on 18 January 2018 and registered prospectively.

Multimodal perturbation analyses of cyclin-dependent kinases reveal a network of synthetic lethalities associated with cell-cycle regulation and transcriptional regulation.

Cell-cycle control is accomplished by cyclin-dependent kinases (CDKs), motivating extensive research into CDK targeting small-molecule drugs as cancer therapeutics. Here we use combinatorial CRISPR/Cas9 perturbations to uncover an extensive network of functional interdependencies among CDKs and related factors, identifying 43 synthetic-lethal and 12 synergistic interactions. We dissect CDK perturbations using single-cell RNAseq, for which we develop a novel computational framework to precisely quantify cell-cycle effects and diverse cell states orchestrated by specific CDKs. While pairwise disruption of CDK4/6 is synthetic-lethal, only CDK6 is required for normal cell-cycle progression and transcriptional activation. Multiple CDKs (CDK1/7/9/12) are synthetic-lethal in combination with PRMT5, independent of cell-cycle control. In-depth analysis of mRNA expression and splicing patterns provides multiple lines of evidence that the CDK-PRMT5 dependency is due to aberrant transcriptional regulation resulting in premature termination. These inter-dependencies translate to drug-drug synergies, with therapeutic implications in cancer and other diseases.

Expanding control of the tumor cell cycle with a CDK2/4/6 inhibitor.

The CDK4/6 inhibitor, palbociclib (PAL), significantly improves progression-free survival in HR+/HER2- breast cancer when combined with anti-hormonals. We sought to discover PAL resistance mechanisms in preclinical models and through analysis of clinical transcriptome specimens, which coalesced on induction of MYC oncogene and Cyclin E/CDK2 activity. We propose that targeting the G1 kinases CDK2, CDK4, and CDK6 with a small-molecule overcomes resistance to CDK4/6 inhibition. We describe the pharmacodynamics and efficacy of PF-06873600 (PF3600), a pyridopyrimidine with potent inhibition of CDK2/4/6 activity and efficacy in multiple in vivo tumor models. Together with the clinical analysis, MYC activity predicts (PF3600) efficacy across multiple cell lineages. Finally, we find that CDK2/4/6 inhibition does not compromise tumor-specific immune checkpoint blockade responses in syngeneic models. We anticipate that (PF3600), currently in phase 1 clinical trials, offers a therapeutic option to cancer patients in whom CDK4/6 inhibition is insufficient to alter disease progression.

Chemotherapy induces dynamic immune responses in breast cancers that impact treatment outcome.

To elucidate the effects of neoadjuvant chemotherapy (NAC), we conduct whole transcriptome profiling coupled with histopathology analyses of a longitudinal breast cancer cohort of 146 patients including 110 pairs of serial tumor biopsies collected before treatment, after the first cycle of treatment and at the time of surgery. Here, we show that cytotoxic chemotherapies induce dynamic changes in the tumor immune microenvironment that vary by subtype and pathologic response. Just one cycle of treatment induces an immune stimulatory microenvironment harboring more tumor infiltrating lymphocytes (TILs) and up-regulation of inflammatory signatures predictive of response to anti-PD1 therapies while residual tumors are immune suppressed at end-of-treatment compared to the baseline. Increases in TILs and CD8+ T cell proportions in response to NAC are independently associated with pathologic complete response. Further, on-treatment immune response is more predictive of treatment outcome than immune features in paired baseline samples although these are strongly correlated.

GeneTEFlow: A Nextflow-based pipeline for analysing gene and transposable elements expression from RNA-Seq data.

Transposable elements (TEs) are mobile genetic elements in eukaryotic genomes. Recent research highlights the important role of TEs in the embryogenesis, neurodevelopment, and immune functions. However, there is a lack of a one-stop and easy to use computational pipeline for expression analysis of both genes and locus-specific TEs from RNA-Seq data. Here, we present GeneTEFlow, a fully automated, reproducible and platform-independent workflow, for the comprehensive analysis of gene and locus-specific TEs expression from RNA-Seq data employing Nextflow and Docker technologies. This application will help researchers more easily perform integrated analysis of both gene and TEs expression, leading to a better understanding of roles of gene and TEs regulation in human diseases. GeneTEFlow is freely available at https://github.com/zhongw2/GeneTEFlow.

PINCER: improved CRISPR/Cas9 screening by efficient cleavage at conserved residues.

CRISPR/Cas9 functional genomic screens have emerged as essential tools in drug target discovery. However, the sensitivity of available genome-wide CRISPR libraries is impaired by guides which inefficiently abrogate gene function. While Cas9 cleavage efficiency optimization and essential domain targeting have been developed as independent guide design rationales, no library has yet combined these into a single cohesive strategy to knock out gene function. Here, in a massive reanalysis of CRISPR tiling data using the most comprehensive feature database assembled, we determine which features of guides and their targets best predict activity and how to best combine them into a single guide design algorithm. We present the ProteIN ConsERvation (PINCER) genome-wide CRISPR library, which for the first time combines enzymatic efficiency optimization with conserved length protein region targeting, and also incorporates domains, coding sequence position, U6 termination (TTT), restriction sites, polymorphisms and specificity. Finally, we demonstrate superior performance of the PINCER library compared to alternative genome-wide CRISPR libraries in head-to-head validation. PINCER is available for individual gene knockout and genome-wide screening for both the human and mouse genomes.

Comparison of the molecular and cellular phenotypes of common mouse syngeneic models with human tumors.

BACKGROUND: The clinical success of immune checkpoint inhibitors demonstrates that reactivation of the human immune system delivers durable responses for some patients and represents an exciting approach for cancer treatment. An important class of preclinical in vivo models for immuno-oncology is immunocompetent mice bearing mouse syngeneic tumors. To facilitate translation of preclinical studies into human, we characterized the genomic, transcriptomic, and protein expression of a panel of ten commonly used mouse tumor cell lines grown in vitro culture as well as in vivo tumors. RESULTS: Our studies identified a number of genetic and cellular phenotypic differences that distinguish commonly used mouse syngeneic models in our study from human cancers. Only a fraction of the somatic single nucleotide variants (SNVs) in these common mouse cell lines directly match SNVs in human actionable cancer genes. Some models derived from epithelial tumors have a more mesenchymal phenotype with relatively low T-lymphocyte infiltration compared to the corresponding human cancers. CT26, a colon tumor model, had the highest immunogenicity and was the model most responsive to CTLA4 inhibitor treatment, by contrast to the relatively low immunogenicity and response rate to checkpoint inhibitor therapies in human colon cancers. CONCLUSIONS: The relative immunogenicity of these ten syngeneic tumors does not resemble typical human tumors derived from the same tissue of origin. By characterizing the mouse syngeneic models and comparing with their human tumor counterparts, this study contributes to a framework that may help investigators select the model most relevant to study a particular immune-oncology mechanism, and may rationalize some of the challenges associated with translating preclinical findings to clinical studies.

Model to improve specificity for identification of clinically-relevant expanded T cells in peripheral blood.

Current methods to quantify T-cell clonal expansion only account for variance due to random sampling from a highly diverse repertoire space. We propose a beta-binomial model to incorporate time-dependent variance into the assessment of differentially abundant T-cell clones, identified by unique T Cell Receptor (TCR) ß-chain rearrangements, and show that this model improves specificity for detecting clinically relevant clonal expansion. Using blood samples from ten healthy donors, we modeled the variance of T-cell clones within each subject over time and calibrated the dispersion parameters of the beta distribution to fit this variance. As a validation, we compared pre- versus post-treatment blood samples from urothelial cancer patients treated with atezolizumab, where clonal expansion (quantified by the earlier binomial model) was previously reported to correlate with benefit. The beta-binomial model significantly reduced the false-positive rate for detecting differentially abundant clones over time compared to the earlier binomial method. In the urothelial cancer cohort, the beta-binomial model enriched for tumor infiltrating lymphocytes among the clones detected as expanding in the peripheral blood in response to therapy compared to the binomial model and improved the overall correlation with clinical benefit. Incorporating time-dependent variance into the statistical framework for measuring differentially abundant T-cell clones improves the model's specificity for T-cells that correlate more strongly with the disease and treatment setting of-interest. Reducing background-level clonal expansion, therefore, improves the quality of clonal expansion as a biomarker for assessing the T cell immune response and correlations with clinical measures.

TNER: a novel background error suppression method for mutation detection in circulating tumor DNA.

BACKGROUND: Ultra-deep next-generation sequencing of circulating tumor DNA (ctDNA) holds great promise as a tool for the early detection of cancer and for monitoring disease progression and therapeutic responses. However, the low abundance of ctDNA in the bloodstream coupled with technical errors introduced during library construction and sequencing complicates mutation detection. RESULTS: To achieve high accuracy of variant calling via better distinguishing low-frequency ctDNA mutations from background errors, we introduce TNER (Tri-Nucleotide Error Reducer), a novel background error suppression method that provides a robust estimation of background noise to reduce sequencing errors. The results on both simulated data and real data from healthy subjects demonstrate that the proposed algorithm consistently outperforms a current, state-of-the-art, position-specific error polishing model, particularly when the sample size of healthy subjects is small. CONCLUSIONS: TNER significantly enhances the specificity of downstream ctDNA mutation detection without sacrificing sensitivity. The tool is publicly available at https://github.com/ctDNA/TNER .

BAFF (B cell activating factor) transcript level in peripheral blood of patients with SLE is associated with same-day disease activity as well as global activity over the next year.

OBJECTIVES: Quantitating gene expression is a potential method of developing biomarkers in systemic lupus erythematosus (SLE). Because of the known pathological role of B cell activating factor (BAFF) in SLE, we explored the association between BAFF gene expression and clinical activity in SLE. METHODS: A total of 275 patients with SLE completed this phase of a prospective observational study. At entry into the study, the BAFF gene expression levels were determined in peripheral blood RNA. Serum concentration of BAFF protein was also measured. We then determined clinical associations with SLE disease history, SLE activity on the same day and SLE activity over the course of the next year. RESULTS: Elevated BAFF gene expression was associated with a history of more leucopenia and serologically with more autoantibodies (anti-dsDNA, anti-Sm, anti-Ro, anti-La and anti-RNP) and low complement. Patients with higher amounts of BAFF transcript had higher measured levels of clinical disease activity. Initial high levels of BAFF gene expression also predicted increased disease activity over the course of the next year. In contrast, serum concentration of BAFF protein was not strongly associated with same-day global disease activity or with future disease activity. CONCLUSIONS: BAFF gene expression level is associated with clinical and serological SLE activity on the same day and predictive of clinical activity over the next year. BAFF gene expression is a better measure and predictor of SLE disease activity than the serum BAFF protein level.

Discovering What Dimensionality Reduction Really Tells Us About RNA-Seq Data.

Biology is being inundated by noisy, high-dimensional data to an extent never before experienced. Dimensionality reduction techniques such as principal component analysis (PCA) are common approaches for dealing with this onslaught. Though these unsupervised techniques can help uncover interesting structure in high-dimensional data they give little insight into the biological and technical considerations that might explain the uncovered structure. Here we introduce a hybrid approach--component selection using mutual information (CSUMI)--that uses a mutual information--based statistic to reinterpret the results of PCA in a biologically meaningful way. We apply CSUMI to RNA-seq data from GTEx. Our hybrid approach enables us to unveil the previously hidden relationship between principal components (PCs) and the underlying biological and technical sources of variation across samples. In particular, we look at how tissue type affects PCs beyond the first two, allowing us to devise a principled way of choosing which PCs to consider when exploring the data. We further apply our method to RNA-seq data taken from the brain and show that some of the most biologically informative PCs are higher-dimensional PCs; for instance, PC 5 can differentiate the basal ganglia from other tissues. We also use CSUMI to explore how technical artifacts affect the global structure of the data, validating previous results and demonstrating how our method can be viewed as a verification framework for detecting undiscovered biases in emerging technologies. Finally we compare CSUMI to two correlation-based approaches, showing ours outperforms both. A python implementation is available online on the CSUMI website.

Lymphotoxin-LIGHT pathway regulates the interferon signature in rheumatoid arthritis.

A subset of patients with autoimmune diseases including rheumatoid arthritis (RA) and lupus appear to be exposed continually to interferon (IFN) as evidenced by elevated expression of IFN induced genes in blood cells. In lupus, detection of endogenous chromatin complexes by the innate sensing machinery is the suspected driver for the IFN, but the actual mechanisms remain unknown in all of these diseases. We investigated in two randomized clinical trials the effects on RA patients of baminercept, a lymphotoxin-beta receptor-immunoglobulin fusion protein that blocks the lymphotoxin-αß/LIGHT axis. Administration of baminercept led to a reduced RNA IFN signature in the blood of patients with elevated baseline signatures. Both RA and SLE patients with a high IFN signature were lymphopenic and lymphocyte counts increased following baminercept treatment of RA patients. These data demonstrate a coupling between the lymphotoxin-LIGHT system and IFN production in rheumatoid arthritis. IFN induced retention of lymphocytes within lymphoid tissues is a likely component of the lymphopenia observed in many autoimmune diseases. ClinicalTrials.gov NCT00664716.

ProbeSelect: selecting differentially expressed probes in transcriptional profile data.

SUMMARY: Transcriptional profiling still remains one of the most popular techniques for identifying relevant biomarkers in patient samples. However, heterogeneity in the population leads to poor statistical evidence for selection of most relevant biomarkers to pursue. In particular, human transcriptional differences can be subtle, making it difficult to tease out real differentially expressed biomarkers from the variability inherent in the population. To address this issue, we propose a simple statistical technique that identifies differentially expressed probes in heterogeneous populations as compared with controls. AVAILABILITY AND IMPLEMENTATION: The algorithm has been implemented in Java and available at www.sourceforge.net/projects/probeselect.

Optimization of a high-throughput whole blood expression profiling methodology and its application to assess the pharmacodynamics of interferon (IFN) beta-1a or polyethylene glycol-conjugated IFN beta-1a in healthy clinical trial subjects.

BACKGROUND: Clinical trials offer a unique opportunity to study human disease and response to therapy in a highly controlled setting. The application of high-throughput expression profiling to peripheral blood from clinical trial subjects could facilitate the identification of transcripts that function as prognostic or diagnostic markers of disease or treatment. The paramount issue for these methods is the ability to produce robust, reproducible, and timely mRNA expression profiles from peripheral blood. Single-stranded complementary DNA (sscDNA) targets derived from whole blood exhibit improved detection of transcripts and reduced variance as compared to their complementary RNA counterparts and therefore provide a better option for interrogation of peripheral blood on oligonucleotide arrays. High-throughput microarray technologies such as the high-throughput plate array platform offer several advantages compared with slide- or cartridge-based arrays; however, manufacturer's protocols do not support the use of sscDNA targets. RESULTS: We have developed a highly reproducible, high-through put, whole blood expression profiling methodology based on sscDNA and used it to analyze human brain reference RNA and universal human reference RNA samples to identify experimental conditions that most highly correlated with a gold standard quantitative polymerase chain reaction reference dataset. We then utilized the optimized method to analyze whole blood samples from healthy clinical trial subjects treated with different versions of interferon (IFN) beta-1a. Analysis of whole blood samples before and after treatment with intramuscular [IM] IFN beta-1a or polyethylene glycol-conjugated IFN (PEG-IFN) beta-1a under optimized experimental conditions demonstrated that PEG-IFN beta-1a induced a more sustained and prolonged pharmacodynamic response than unmodified IM IFN beta-1a. These results provide validation of the utility of this new methodology and suggest the potential therapeutic benefit of a sustained pharmacodynamic response to PEG-IFN beta-1a. CONCLUSIONS: This novel microarray methodology is ideally suited for utilization in large clinical studies to identify expressed transcripts for the elucidation of disease mechanisms of action and as prognostic, diagnostic, or toxicity markers.

A computational framework for boosting confidence in high-throughput protein-protein interaction datasets.

Improving the quality and coverage of the protein interactome is of tantamount importance for biomedical research, particularly given the various sources of uncertainty in high-throughput techniques. We introduce a structure-based framework, Coev2Net, for computing a single confidence score that addresses both false-positive and false-negative rates. Coev2Net is easily applied to thousands of binary protein interactions and has superior predictive performance over existing methods. We experimentally validate selected high-confidence predictions in the human MAPK network and show that predicted interfaces are enriched for cancer -related or damaging SNPs. Coev2Net can be downloaded at http://struct2net.csail.mit.edu.

Lymphotoxin-beta receptor blockade reduces CXCL13 in lacrimal glands and improves corneal integrity in the NOD model of Sjögren's syndrome.

INTRODUCTION: In Sjögren's syndrome, keratoconjunctivitis sicca (dry eye) is associated with infiltration of lacrimal glands by leukocytes and consequent losses of tear-fluid production and the integrity of the ocular surface. We investigated the effect of blockade of the lymphotoxin-beta receptor (LTBR) pathway on lacrimal-gland pathology in the NOD mouse model of Sjögren's syndrome. METHODS: Male NOD mice were treated for up to ten weeks with an antagonist, LTBR-Ig, or control mouse antibody MOPC-21. Extra-orbital lacrimal glands were analyzed by immunohistochemistry for high endothelial venules (HEV), by Affymetrix gene-array analysis and real-time PCR for differential gene expression, and by ELISA for CXCL13 protein. Leukocytes from lacrimal glands were analyzed by flow-cytometry. Tear-fluid secretion-rates were measured and the integrity of the ocular surface was scored using slit-lamp microscopy and fluorescein isothiocyanate (FITC) staining. The chemokine CXCL13 was measured by ELISA in sera from Sjögren's syndrome patients (n = 27) and healthy controls (n = 30). Statistical analysis was by the two-tailed, unpaired T-test, or the Mann-Whitney-test for ocular integrity scores. RESULTS: LTBR blockade for eight weeks reduced B-cell accumulation (approximately 5-fold), eliminated HEV in lacrimal glands, and reduced the entry rate of lymphocytes into lacrimal glands. Affymetrix-chip analysis revealed numerous changes in mRNA expression due to LTBR blockade, including reduction of homeostatic chemokine expression. The reduction of CXCL13, CCL21, CCL19 mRNA and the HEV-associated gene GLYCAM-1 was confirmed by PCR analysis. CXCL13 protein increased with disease progression in lacrimal-gland homogenates, but after LTBR blockade for 8 weeks, CXCL13 was reduced approximately 6-fold to 8.4 pg/mg (+/- 2.7) from 51 pg/mg (+/-5.3) in lacrimal glands of 16 week old control mice. Mice given LTBR blockade exhibited an approximately two-fold greater tear-fluid secretion than control mice (P = 0.001), and had a significantly improved ocular surface integrity score (P = 0.005). The mean CXCL13 concentration in sera from Sjögren's patients (n = 27) was 170 pg/ml, compared to 92.0 pg/ml for sera from (n = 30) healthy controls (P = 0.01). CONCLUSIONS: Blockade of LTBR pathways may have therapeutic potential for treatment of Sjögren's syndrome.

Causal modeling using network ensemble simulations of genetic and gene expression data predicts genes involved in rheumatoid arthritis.

Tumor necrosis factor α (TNF-α) is a key regulator of inflammation and rheumatoid arthritis (RA). TNF-α blocker therapies can be very effective for a substantial number of patients, but fail to work in one third of patients who show no or minimal response. It is therefore necessary to discover new molecular intervention points involved in TNF-α blocker treatment of rheumatoid arthritis patients. We describe a data analysis strategy for predicting gene expression measures that are critical for rheumatoid arthritis using a combination of comprehensive genotyping, whole blood gene expression profiles and the component clinical measures of the arthritis Disease Activity Score 28 (DAS28) score. Two separate network ensembles, each comprised of 1024 networks, were built from molecular measures from subjects before and 14 weeks after treatment with TNF-α blocker. The network ensemble built from pre-treated data captures TNF-α dependent mechanistic information, while the ensemble built from data collected under TNF-α blocker treatment captures TNF-α independent mechanisms. In silico simulations of targeted, personalized perturbations of gene expression measures from both network ensembles identify transcripts in three broad categories. Firstly, 22 transcripts are identified to have new roles in modulating the DAS28 score; secondly, there are 6 transcripts that could be alternative targets to TNF-α blocker therapies, including CD86--a component of the signaling axis targeted by Abatacept (CTLA4-Ig), and finally, 59 transcripts that are predicted to modulate the count of tender or swollen joints but not sufficiently enough to have a significant impact on DAS28.

iWRAP: An interface threading approach with application to prediction of cancer-related protein-protein interactions.

Current homology modeling methods for predicting protein-protein interactions (PPIs) have difficulty in the "twilight zone" (<40%) of sequence identities. Threading methods extend coverage further into the twilight zone by aligning primary sequences for a pair of proteins to a best-fit template complex to predict an entire three-dimensional structure. We introduce a threading approach, iWRAP, which focuses only on the protein interface. Our approach combines a novel linear programming formulation for interface alignment with a boosting classifier for interaction prediction. We demonstrate its efficacy on SCOPPI, a classification of PPIs in the Protein Databank, and on the entire yeast genome. iWRAP provides significantly improved prediction of PPIs and their interfaces in stringent cross-validation on SCOPPI. Furthermore, by combining our predictions with a full-complex threader, we achieve a coverage of 13% for the yeast PPIs, which is close to a 50% increase over previous methods at a higher sensitivity. As an application, we effectively combine iWRAP with genomic data to identify novel cancer-related genes involved in chromatin remodeling, nucleosome organization, and ribonuclear complex assembly. iWRAP is available at http://iwrap.csail.mit.edu.

Convergent Random Forest predictor: methodology for predicting drug response from genome-scale data applied to anti-TNF response.

Biomarker development for prediction of patient response to therapy is one of the goals of molecular profiling of human tissues. Due to the large number of transcripts, relatively limited number of samples, and high variability of data, identification of predictive biomarkers is a challenge for data analysis. Furthermore, many genes may be responsible for drug response differences, but often only a few are sufficient for accurate prediction. Here we present an analysis approach, the Convergent Random Forest (CRF) method, for the identification of highly predictive biomarkers. The aim is to select from genome-wide expression data a small number of non-redundant biomarkers that could be developed into a simple and robust diagnostic tool. Our method combines the Random Forest classifier and gene expression clustering to rank and select a small number of predictive genes. We evaluated the CRF approach by analyzing four different data sets. The first set contains transcript profiles of whole blood from rheumatoid arthritis patients, collected before anti-TNF treatment, and their subsequent response to the therapy. In this set, CRF identified 8 transcripts predicting response to therapy with 89% accuracy. We also applied the CRF to the analysis of three previously published expression data sets. For all sets, we have compared the CRF and recursive support vector machines (RSVM) approaches to feature selection and classification. In all cases the CRF selects much smaller number of features, five to eight genes, while achieving similar or better performance on both training and independent testing sets of data. For both methods performance estimates using cross-validation is similar to performance on independent samples. The method has been implemented in R and is available from the authors upon request: Jadwiga.Bienkowska@biogenidec.com.