Image-based ex-vivo drug screening for patients with aggressive haematological malignancies: interim results from a single-arm, open-label, pilot study

Summary Background Patients with refractory or relapsed haematological malignancies have few treatment options and short survival times. Identification of effective therapies with genomic-based precision medicine is hampered by intratumour heterogeneity and incomplete understanding of the contribution of various mutations within specific cancer phenotypes. Ex-vivo drug-response profiling in patient biopsies might aid effective treatment identification; however, proof of its clinical utility is limited. Methods We investigated the feasibility and clinical impact of multiparametric, single-cell, drug-response profiling in patient biopsies by immunofluorescence, automated microscopy, and image analysis, an approach we call pharmacoscopy. First, the ability of pharmacoscopy to separate responders from non-responders was evaluated retrospectively for a cohort of 20 newly diagnosed and previously untreated patients with acute myeloid leukaemia. Next, 48 patients with aggressive haematological malignancies were prospectively evaluated for pharmacoscopy-guided treatment, of whom 17 could receive the treatment. The primary endpoint was progression-free survival in pharmacoscopy-treated patients, as compared with their own progression-free survival for the most recent regimen on which they had progressive disease. This trial is ongoing and registered with ClinicalTrials.gov, number NCT03096821. Findings Pharmacoscopy retrospectively predicted the clinical response of 20 acute myeloid leukaemia patients to initial therapy with 88·1% accuracy. In this interim analysis, 15 (88%) of 17 patients receiving pharmacoscopy-guided treatment had an overall response compared with four (24%) of 17 patients with their most recent regimen (odds ratio 24·38 [95% CI 3·99–125·4], p=0·0013). 12 (71%) of 17 patients had a progression-free survival ratio of 1·3 or higher, and median progression-free survival increased by four times, from 5·7 (95% CI 4·1–12·1) weeks to 22·6 (7·4–34·0) weeks (hazard ratio 3·14 [95% CI 1·37–7·22], p=0·0075). Interpretation Routine clinical integration of pharmacoscopy for treatment selection is technically feasible, and led to improved treatment of patients with aggressive refractory haematological malignancies in an initial patient cohort, warranting further investigation. Funding Austrian Academy of Sciences; European Research Council; Austrian Science Fund; Austrian Federal Ministry of Science, Research and Economy; National Foundation for Research, Technology and Development; Anniversary Fund of the Austrian National Bank; MPN Research Foundation; European Molecular Biology Organization; and Swiss National Science Foundation.


Introduction
Genetic studies have identified several genomic alterations associated with the development of haema tological malignancies. However, barriers remain in fully translating this genomic information into direct clinical benefit for patients. Current efforts to introduce per sonalised medicine in patients with cancer, which focus on genetic and molecular patient stratification, have produced varying results. [1][2][3][4][5] In a pioneering study, 4 which used each patient as their own control, 27% of patients with recurrent metastatic cancer of any kind had a 30% longer progressionfree survival with treatment selected on the basis of genetic profiling than they did with their previous treatment. However, the SHIVA study, 5 one of the first randomised trials of genomic based precision medicine, did not show a benefit in progressionfree survival for patients assigned to genome based targeted treatment compared with treatment according to physician's choice in heavily pretreated patients with cancer. Genomebased therapy decisions are limited by our incomplete understanding of the relationship between cancer phenotype and genotype, and the complex genetics underlying cancer are the result of dynamic microevolutionary processes. 6 For instance, whereas several studies have linked cytogenetic and molecular abnormalities with distinct clinical outcomes in acute myeloid leukaemia, 7 accurate prediction of treatment response of individual patients with acute myeloid leukaemia to induction therapy remains challenging. [8][9][10][11][12] Furthermore, patients with aggressive haematological malignancies, who have failed at least two lines of therapy, are often without further standard treatment options and have a poor prognosis. 13 These patients will usually receive either best available therapy, supportive care, or will be enrolled in clinical trials. Therefore, dynamic approaches that measure drug responses in cancer cells derived from patient biopsies might complement such static genetic measurements. For example, exvivo chemosensitivity tests have been done in samples from patients with chronic or acute leukaemia and multiple myeloma, [14][15][16][17][18][19][20][21] in breast cancer derived stable cell lines, 22 in patientderived xenografts in mice, 23,24 and in gut stemcellderived organoids. 25,26 These pioneering functional assays have provided proof of concept by showing that exvivo responses might match clinical response; however, these studies have not been integrated into clinical routine because of practical limitations and scarce proof of clinical benefit. [27][28][29] Here, we investigate the clinical impact of a newly developed technology platform that combines multi parametric immunofluorescence with highthroughput automated microscopy and singlecell image analysis, called pharmacoscopy. 30 Pharmacoscopy enables tumour cell specific quantification of biological parameters of millions of adherent and nonadherent individual cells with high sample efficiency, minimal sample mani pulation, extensive automation, and fast turnaround times. We thus aimed to evaluate the feasibility of integrating pharmacoscopy into the clinic, and to assess clinical response in patients who received a treatment according to pharmacoscopy results as an individual healing attempt.

Study design and participants
For this singlearm, openlabel, pilot study, we collected samples and clinical data from patients with latestage haematological malignancies. Patients were eligible for inclusion if no further standard treatments or clinical trials were available for the patient; the patient had undergone at least two lines of previous therapy; the patient gave written informed consent; cancer cellcontaining samples could be biopsied after written informed consent; the clinical decision was made by a board consisting of haematologists, pathologists, pharmacists, and molecular biologists; and candidate treatments identified by pharmacoscopy were clinically available and considered safe given the patient's health condition. Pharmacoscopyguided therapy was provided to individual latestage patients as an individual healing attempt, in accordance with European Union and Austrian namedpatient use legislation. Ethical approval was granted by the Ethics Commission of the Medical

Research in context
Evidence before this study We did a systematic search of PubMed using the search terms ("functional screening" [Title/Abstract] OR "chemosensitivity test" [Title/Abstract] OR "chemoresistance test" [Title/Abstract] OR "drug profiling" [Title/Abstract] OR "drug response" [Title/Abstract]) AND ("leukemia" [Title/Abstract] OR "leukaemia" [Title/Abstract] OR "lymphoma" [Title/Abstract] OR "myeloma" [Title/Abstract] OR "hematologic" [Title/Abstract]). We did not restrict the search by date, language, or article type. We did one search before initiating this study on Sept 1, 2015, and we repeated this search on Oct 13, 2017. Several studies have shown the potential for retrospective patient stratification based on a variety of ex-vivo drug-response profiling techniques; however, no reports were found of studies in which patient treatment for haematological malignancies were adapted to ex-vivo drug-response profiling across large panels of drugs. We also searched ClinicalTrials.gov for published clinical trials using the search terms described above. This search retrieved only one other clinical trial (currently recruiting patients and with feasibility as endpoint), in which patient treatment for haematological malignancies is being adapted to the drug-response profiles of primary biopsies across at least 100 different drugs tested.

Added value of this study
To our knowledge, our study is the first prospective study showing feasibility and efficacy of ex-vivo drug-response profiling to guide personalised treatment selection across large panels of possible treatments for patients suffering aggressive haematological malignancies. We do so with a new image-based, drug-response profiling technique that we call pharmacoscopy, which uses high-throughput, automated, confocal microscopy; immunofluorescence; and single-cell image analysis.

Implications of all the available evidence
Our interim study results indicate that adapting treatment regimens of patients with aggressive haematological malignancies to pharmacoscopy is feasible, safe, and effective. More patients whose treatment protocols were selected by the haematological tumour board based on pharmacoscopy results had an overall response and had longer progression-free survival with pharmacoscopy than their previous treatment. Further studies with randomised trial designs and larger patient cohorts than our study are justified to further elucidate the clinical impact of our novel, image-based, ex-vivo drug-response profiling platform. Thus, a pharmacoscopy score of 1 represents the strongest on target exvivo response, a pharmacoscopy score of 0 indicates no exvivo effect, and negative pharmacoscopy scores indicate exvivo chemoresistance.
To explore whether pharmacoscopy is predictive of clinical response, we designed a retrospective study using samples from patients with acute myeloid leukaemia collected before receiving standard firstline remission induction therapy (figure 1A). Roughly 60% of patients typically respond with complete remission to induction therapy, 31 which consists of cytarabine combined with daunorubicin and etoposide. 32,33 Each patient sample was screened through a drug combination matrix of all three firstline drugs, consisting of 125 unique drug concentration combinations in four technical repeats. To determine the exvivo druginduced cytotoxicity, we quantified the number of nonfragmented nuclei in each image after drug treatment. Druginduced cell death based on nuclear morphology was measured after overnight drug incubation, and subpopulation specificity was assessed on the cells that stained positive for CD34 or KIT (CD117). Both markers are commonly present on leu kaemic blasts in acute myeloid leukaemia. 34 Because all patient samples contained a combination of blast cells and nonmalignant cells, we calculated the RBF of drugtreated cells to compare on target druginduced cytotoxicity with that of druginduced cytotoxicity in blastmarkernegative cells ( figure 1A). The RBF is thus defined as the fraction of viable blasts surviving drug treatment relative to the average fraction of viable blasts observed in negative control samples. Drug sensitivity per patient was integrated over the drug matrix by averaging the number of RBF datapoints above (scored with +1) or below (-1) the hyperplane that best separated responders from non responders (figure 1G), weighted by the area under the receiveroperating characteristic curve (AUROC) of each cor responding concentration point in the drug matrix. We also compared clinical response with cytogenetic and molecular risk classification.
For the prospective study, eligible patients were assessed by the board (PBS, UJ, GIV, KM, CK, GH, ISK, KO, WRS) and those who met inclusion criteria were tested by pharmacoscopy as outlined above. The markers used to identify the blast populations were selected individually for each patient based on their disease indication and clinical diagnostics. For the prospective study, ontarget cytotoxicity was identified by calculating the RBF as in the acute myeloid leukaemia retrospective analysis, in which blast, in this context, now referred to any cancer cell. Thus, topscoring drugs achieved the most specific reduction of the tumourcellenriched cell fraction ex vivo, while causing minimal cytotoxicity to the markernegative healthy cells also present in the sample. The board then assessed the results, taking into account an individual patient's previous treatment outcomes to recommend the next treatment regimen. Patients assessed by the board, but who had further standard treatment options, were used as an observational cohort. Integration of data from both patient groups allowed us to test whether chemo resistance measured by pharmacoscopy (eg, exvivo survival of blast cells coinciding with death of non malignant cells) is predictive of poor clinical response. To gain an overview of the complete dataset, we first set out to cluster the drugresponse profiles. For this purpose, RBF values were normalised to pharma coscopy scores; negative values indicate drug resistance (blast survival and nonmalignantcell death), and positive values indicate on target chemosensitivity (blast death and nonmalignant cell survival; appendix p 5).
To account for the complicating fact that for most treatment regimens comprised of multiple drugs, exvivo   testing was in fact done with single drug treatments, and that multiple and varying number of blast markers were measured in different patients depending on their clinical diagnostic results, we summed the relevant pharma coscopy values over all drugs and markers per patient, resulting in an integrated pharmacoscopy (iPCY) score. We quantified overall response as 1=progressive disease, 2=stable disease, 3=partial response, 4=complete remis sion, and determined the correlation with iPCY.

Outcomes
The primary outcome measure was the proportion of patients achieving progressionfree survival, and the secondary outcome measure was the proportion of patients with an overall response (achieving either a complete remission or partial response). Progressionfree survival was calculated as the time from the first day of treatment to the date of the first reported disease progression or relapse, initiation of a new (unplanned) anticancer treat ment, or death as a result of any cause. Overall response was defined by achieving either complete remission or a partial response, defined by standard response definition guidelines. 35,36 For patients with lymphoma, responses were classified as complete remission, partial response, stable disease, or progressive disease according to the criteria proposed by the international working group on malignant lymphoma. 35 For patients with leukaemia, responses were assessed following the response criteria defined by the recommen dations of the European LeukemiaNet. 36 All patients that were included in the prospective trial had uniform followup intervals of 4 weeks.

Statistical analysis
The treatment was deemed to be of clinical benefit for the individual patient who has a progressionfree survival ratio (progressionfree survival on pharmacoscopyguided therapy/progressionfree survival on prior therapy) of 1·3 or higher. In such cases, we rejected the null hypothesis, defined as 15% or fewer patients having a progressionfree survival ratio of 1·3 or higher. Thus, the individual patient was their own control. Comparisons of the overall response to previous treatment and pharmacoscopyguided treat ments were calculated using a onesided McNemar's test for paired binomial data with continuity correction. The odds ratio (OR) could not directly be calculated as one of the discordant values (those patients who did respond to the most recent treatment, but who did not respond to pharmacoscopyguided treatment) was equal to zero. We therefore calculated the overall responseassociated OR using the standard calculation for contingency tables. Significance testing for progressionfree survival dif ferences was done using the logrank (MantelCox) test. All correlations are Pearson correlation coefficients. All other p values are twotailed t tests, unless stated otherwise. Statistical analyses were done in GraphPad Prism (version 7), Matlab (versions R2015a, R2015b, R2016a, R2016b, R2017a, R2017b), and Microsoft Excel (version 2016).
The trial was registered at the ClinicalTrials.gov trial registry, number NCT03096821.

Role of the funding source
The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. PBS and UJ had access to patient annotated clinical data, BS, GIV, and GSF had access to the anonymised patient clinical data and correlated drug responses. The corresponding author had full access to all the anonymised results and final responsibility for the decision to submit for publication.

Results
The retrospective acute myeloid leukaemia study to determine whether pharmacocopy is predictive of clinical response used 20 biobanked bone marrow samples; ten samples from patients achieving stable complete remission to induction therapy, and ten from non responders to induction therapy. 32,33 The correlation coefficient (r) when examining the number of non fragmented nuclei in each image after drug treatment was 0·99 with immunofluorescence against activated caspase3 as a measure of cell death over samples from three patients, confirming the nuclear morphology readout (figure 1B). Bone marrow immunohistochemistry from clinical diagnostics confirmed the presence of CD34 and CD117 on blasts of all 20 patients. The 20 patients represented both sexes and diverse ages, had diverse genetic lesions and karyotypes, and blast fractions at time of sampling ranging from 30% to over 90% (appendix p 6). The clinical response to treatment in our retrospective cohort of patients with acute myeloid leukaemia only partially followed the cytogenetic and molecular risk classification (appendix pp 3, 6), with, for instance, all four patient who had a FLT3-ITD mutation in the nonresponders group and both inv (16) patients in the complete remission group.
The RBF was significantly different between complete remission and nonresponders groups, with significantly stronger ontarget effects observed with exvivo dauno rubicin treatment for the complete remission patient cohort (p<0·0001; figures 1C, 1D, appendix pp 3, 7). Conversely, populationaveraged cytotoxicity measure ments (total cell death) did not correctly stratify patients based on their clinical response (figures 1E, 1F), indicating the need for the relative drug sensitivity measurements. As expected, the integrated response score for drug sensitivity revealed good separation of responders and non responders ( figure 1H, appendix p 3). One particularly strong outlier was observed, complete remission in patient 10 for whom no exvivo response was measured. This discrepancy could not be attributed to differences in clinical parameters nor technical issues. Crossvalidation by leaving out and reclassifying every possible combination of two patient samples, and calculation of the ideal hyperplane based on the remaining 18 samples, revealed an average classification accuracy of 88·1% for the RBF (figure 1I), and an average AUROC of 0·97 ( figure 1J).
Consistently, we observed reduced classification power for this cohort with populationaveraged readouts: overall cell death, quantified by the total cell number, led to a    classification accuracy of 68·5% (AUROC 0·86), and cell death of markerpositive cells, quantified as the total blasts, led to a classification accuracy of 78·1% (AUROC 0·91; figures 1I, 1J).
In the prospective study of the 57 patients with aggressive haematological malignancies, nine patients were not assessed by the board for reasons given in figure 2A. Of 48 patients who were assessed by the board, 18 were not included and 13 still had further treatment options, leaving 17 patients who met the inclusion criteria to receive pharmacoscopyguided treatment (figure 2A). For these 17 patients, pharmacoscopy was always done on the same day as the biopsy procedure, and median time to report pharmacoscopy results back to clinicians was 5 days (IQR 2-8). The trial started on Sept 1, 2015, the censoring date for the interim analysis for all patients was Nov 11, 2016, and the median followup time was 7·6 months (IQR 4·5-8·7). The characteristics of the 17 patients who had pharmacoscopyguided treatment are listed in the table. The 13 patients reviewed by the board that received treatments not guided by pharmacoscopy served as an observational cohort (appendix p 8). A comparison be tween the percentage of markerpositive cells measured from the same biopsies by clinical diagnosticsbased flow cytometry, the current gold standard, and by pharma coscopy revealed strong consistency between the two methods (r=0·92, p<0·0001; figure 2B).
Pharmacoscopyguided treatment regimens resulted in encouraging partial and complete remissions (   figure 2D). 6mercaptopurine and bortezomib were combined with antiCD20 obinutuzumab. After 28 days PETCT confirmed a partial response ( figure 2H). For patient 5, cells from an excised lymph node were tested for 139 drugs ( figure 2E). The patient achieved a complete remission ( figure 2I) to a combination of the single strongest exvivo acting drug bortezomib (RBF 0·589, p<0·0001), with cladribine ranked fifth (RBF 0·727; p=0·00029) and dexamethasone ranked 15th (RBF 0·866; p=0·050; figure 2E). And after exvivo sensitivity to cladribine was measured for patient 9 (figure 2F), complete remission was observed with cladribine treat ment in combination with CD30targeted immuno therapy brentuximab vedotin ( figure 2J). Data for patient 7, one of the two patients who did not respond to pharmacoscopy guided treatment, is shown and further discussed in the appendix (p 4). Overall response and progressionfree survival of pharmacoscopy were compared with overall response and progressionfree survival for the most recent regimen on which the patient had progressed. Four (24%) of 17 patients achieved an overall response with the most recent regimen compared with 15 (88%) of 17 patients who achieved an overall response with pharmacoscopyguided treatment (odds ratio 24·38 [95% CI 3·99-125·4], p=0·0013; figure 3A). Five (38%) of 13 patients receiving standard salvage treatment based on physician's choice achieved an overall response (appendix p 8). Notably, none of the 17 patients receiving pharmacoscopyguided treatments had progressive disease as best overall response, whereas seven patients had progressive disease in response to their most recent regimen ( figure 3A). Furthermore, pharma coscopyguided treatments also led to a significantly improved median progressionfree survival (22·6 weeks [95% CI 7·4-34·0]) compared with a median of 5·7 weeks (4·1-12·1) in the same patients with the most recent regimen (hazard ratio 3·14 [95% CI 1·37-7·22], p=0·0075; figure 3B). 12 (71%) of 17 patients had a progressionfree survival ratio of 1·3 or higher ( figure 3C). The null hypothesis was therefore rejected. Notably,   of less than -0·2, less than -0·3, or less than -0·4. Patients who had received no or only one previous treatment line showed exvivo chemoresistance (a pharmacoscopy score less than -0·1) to 9% of tested drugs, whereas patients that had received five or more previous treatment lines showed exvivo chemoresistance to 17% of tested drugs (p=0·016; figure 4D). Patient outcomes correlated positively with the integrated pharmacoscopy scores (r=0·49, p=0·0065; figure 4B). 18 (94%) of 19 responding patients (partial response and complete remission) had iPCY scores between 0 to 3, whereas four (67%) of six patients with progressive disease, all of whom did not receive pharmacoscopyguided treatments, had iPCY scores in the negative range between -0·75 and -7. Patient treatments associated with high iPCY scores combined drugs acting ontarget on all tested blast markers, or combined neutral, exvivo acting drugs with exvivo ontarget acting drugs. Conversely, patients res ponding with progressive disease as best overall response had treatments including drugs to which strong, exvivo chemoresistance was measured. One of two nonresponding (stable disease) patients receiving pharma coscopyguided treatments had an iPCY score of below -1, indicating that the pharmacoscopy test did not strongly support the final personalised treatment regimen for this nonresponding patient, due to exvivo discordance depending on the used blast markers ( figure 4B). Overall, the iPCY score separated progressive disease from patients who had achieved a partial response and complete remission with a classification accuracy of 92% and an AUC of 0·84 ( figure 4B).

Discussion
This singlecentre study shows technical feasibility of integrating automated microscopybased, exvivo drug response profiling for patients with aggressive haemato logical malignancies into clinical practice. The testguided treatment regimens led to significantly longer progression free survival and improved overall response in patients with various haematological malignancies compared with their most recent regimens, warranting further disease specific clinical studies that include larger patient cohorts and randomised control groups. 38 Although the trial did not include a randomised control group and had a relatively small cohort size of 17 patients, our results suggest that a wide array of working chemotherapeutics and targeted inhibitors already exist, which, in principle, are capable of breaking drug resistance even in multirefractory cancers, if the right drugs are selected at the right time for each individual patient. We found that an integrative combination of chemosensitivity of the leukaemic blasts and chemoresistance of the marker negative, nonmalignant cells predicted clinical response to firstline acute myeloid leukaemia treatment with the highest accuracy. Furthermore, the same readout guided selection of treatments associated with favourable clinical responses, and predicted both good as well as poor clinical which showed that both overall response (p=0·0002) and progressionfree survival (p=0·025) remained signif icantly improved for pharma coscopyguided treatments compared with the most recent regimen. This reanalysis allowed us to exclude the possibility that the addition of antibodybased immuno therapies affected our interpre tation of the results. Taken together, pharmacoscopy guided treatment regimens de monstrated strongly im proved clinical responses and survival benefit in an initial cohort of 17 latestage patients with aggressive relapsed and refractory haematological malignancies.
To test whether chemoresistance measured by pharma coscopy is predictive of poor clinical response we did a cluster analysis including the 17 patients receiving pharma coscopyguided treatment with the 12 observation cohort patients whose subsequent treatments were also tested exvivo before treatment initiation. Hierarchical clustering of the pharmacoscopy response profiles per patient and blastmarkers as determined by clinical diagnostics, overlaid with the best overall response corresponding to drug and patient pairs, revealed extensive patienttopatient vari ability in both the number and identity of drugs to which either chemo resistance or chemosensitivity was measured (appendix p 5). Similar indications displayed remarkable heterogeneity in response profiles, indicating an absence of character istic exvivo responses for the tested indications in this partially heavily pretreated cohort. Hierarchical clustering repeatedly grouped drug classes with the same mode of action, including immuno modulatory drugs (thalidomide, lenal idomide, and pomalidomide), anthra cycline chemo therapies (daunorubicin, doxorubicin, and valru bicin), and histone deacetylase inhibitors (belinostat, pano binostat, and vorinostat). The clustering further high lighted the diversity of treatments given to the patients. 30 unique drugs, distributed across the clus tering, were tested by pharmacoscopy and subsequently administered to patients, enabling robust pantreatment statistical analysis (figure 4, appendix p 5).
Further analyses demonstrated the association between exvivo chemoresistance and poor clinical outcome. First, plotting the average pharmacoscopy scores over all markers and drugs in relation to associated overall response to those drugs showed that treatments leading to progressive disease were associated with negative pharma coscopy scores, whereas treatments leading to partial response or complete remission resulted in significantly positive pharma coscopy scores ( figure 4A). Second, the treatments to which the patient had relapsed before pharmacoscopy testing had on average negative pharmacoscopy scores (p=0·0079; figure 4C). Third, the percentage of tested drugs to which exvivo resistance was measured (at pharmacoscopy score less than -0·1) increased with the number of previous treatment rounds of each of the 29 patients (r=0·44; p=0·016; figure 4D). Similar significantly positive correlations were found when defining chemoresistance as pharmacoscopy scores responses. The positive relation observed between the number of previous treatment lines and exvivo drug resistance is intuitive, and might reflect acquired drug resistance as well as refractory disease being more resistant from disease onset.
Our investigation was designed as a prospective, non randomised study in which every patient acted as their own control. This approach allowed us to assess the overall effect across heterogeneous diseases and treatment regimens; however, the absence of random isation could have led to bias. 6,38 Future randomised trials testing pharmacoscopyguided therapies versus physician's choice are therefore warranted, and should focus on individual disease entities.
Not all patients in our study had correlation between pharmacoscopy results and outcome, in particular one outlier patient (patient 10 in the retrospective acute myeloid leukaemia study). Identifying the causes for such outliers will thus require repetition with larger cohort sizes and integration with systematic molecular data. In our comparison of con ventional genetics with response in the retrospective acute myeloid leukaemia cohort, our results matched those in previous studies. 39 A benefit of pharmacoscopy resides in the analytical power derived from monitoring with computeraided precision millions of individual singlecell drug responses, which combined with the ability to discriminate cell types allows us to score specific rather than general and averaged cytotoxic effects. Pharmacoscopy will likely be instructive for the personalised identification of clinically effective therapies for other malignancies beyond those tested here. The selection of personalised therapy by pharmacoscopy benefits from the ability to measure hundreds to thousands of drug exposures using small patient samples, in which each exvivo treatment includes healthy cell controls from the same patient sample. Pharmacoscopy detects cancer cells with fluorescently labelled antibodies against clinically used diagnostic markers, which means the test synergises with, and uses similar antibodies as, clinical flow cytometrybased diagnostics. Both microscopy and flow cytometry or optometry share the limitations of detection of cancer cells by antibodybased immuno fluorescence, whereas pharmacoscopy allows for reduced sample processing and increased throughput and automation. Our results show that singlecell detection of blast markers by pharma coscopy enables a clinically useful comparison of ontarget and offtarget cytotoxicity, while the minimal exvivo culturing of cells, and compatibility with clinical diagnostic markers, ensure fast and relevant feedback. Specifically, the platform allowed us to test 768 conditions for almost all of the 17 patient samples, returning results to the clinic within 5 days of receiving a sample. A crucial tradeoff nonetheless remains between the number of different drugs, technical replicates, concentration ranges, timepoints, and drug combinations that can be tested from one biopsy. In that regard, the observation made in this study that exvivo testing of single treatments can aid selection of clinically beneficial combination treatments suggests that not every drug combination needs to be tested in combination ex vivo, thus allowing for larger drug panels to be tested; future studies are needed to further refine optimal, exvivo drugtesting regimens.
Comprehensive drug response profiles of individual people, as generated here, represent the outcome of interplay between various molecular parameters of the responding cells, including not only the genetic, proteomic, and metabolic state of the cells, but also the direct and indirect molecular interactions with other cells. 30 We therefore hypothesise that such comprehensive drug response profiles can offer novel functional insight into the underlying health status of an individual, with potentially wideranging implications in preventive and participatory medicine. Given the fast throughput of the method, both experimentally and analytically, future studies can include higher patient numbers that will be of great interest to investigate the translatability of pharma coscopy further.
Pharmacoscopy provided useful treatmentguidance in an initial latestage patient cohort, warranting further investigation in larger and indicationspecific clinical trials. It is likely that the approach will synergise well with molecular profiling techniques such as genomics and proteomics for personalised treatment identi fication. Such studies could lead to improved patient treatment and be a useful route to mechanistic elucidation of clinically relevant genotypetophenotype relationships.

Declaration of interests
BS is a shareholder of Allcyte, has a patent WO2016046346 licensed to Allcyte, and is a scientific cofounder of Allcyte. GIV is a shareholder of Allcyte, has a patent WO2016046346 licensed to Allcyte, and is a scientific cofounder of Allcyte. NK is a shareholder of Allcyte, has a patent WO2016046346 licensed to Allcyte, and is a scientific cofounder of Allcyte. KM reports personal fees from Chugai, Kyowahakko Kirin, outside the submitted work. GH reports grants and personal fees from Gilead; and personal fees from Novartis, Amgen, and Ariad, outside the submitted work. KO reports personal fees from Novartis, outside the submitted work. RK reports personal fees from and has been an advisory board member for AOP Orphan, PharmaEssentia, and Qiagen; and has received personal fees from Novartis, outside the submitted work. UJ reports personal fees from Janssen Cilag, during the conduct of the study; grants and personal fees from Roche, Celgene, Gilead, Novartis, and True North Therapeutics; and personal fees from Amgen, Takeda, AbbVie, Infinity, outside the submitted work. GSF reports grants from ERC Proof of Concept, during the conduct of the study; and is a shareholder of Allcyte, has a patent WO2016046346 licensed to Allcyte, and is a scientific cofounder of Allcyte. All other authors declare no competing interests.