Patient-Customized Drug Combination Prediction and Testing for T-cell Prolymphocytic Leukemia Patients

Combination therapies are essential to address the genetic and epigenetic complexity and plasticity of cancers, due to their potential to overcome the resistance mechanisms that often hinder the long-term efficacy of conventional cytotoxic chemotherapies or targeted agents inhibiting single pathways (1, 2). However, it is increasingly appreciated that both the resistance mechanisms and combination effects are highly heterogeneous and context-specific (3–6). Such heterogeneity and redundancy limit the clinical applicability of those precision medicine approaches that are based solely on genomics-based matching of the right drugs to the right patients. This limitation can be partly addressed by functional testing of patient-derived cells exposed to large number of both targeted and conventional therapies using drug testing assays (7–12). Therefore, it seems evident that both next-generation sequencing and functional testing is required to precisely match combination drug therapies to individual cancer patients.

Systematic testing of targeted agents individually, in combination with cytotoxic treatments, and in combination with other targeted agents poses significant experimental challenges (13). Even if we consider only the approved drugs and exclude agents with redundant mechanisms, the number of two-drug combination remains in tens of thousands (1, 2, 14). Therefore, large-scale patient testing is often being done using panels of limited numbers of combinations (15). Furthermore, despite a detailed characterization of approved and investigational combination therapies (16, 17), the mechanisms behind clinically successful combinations are still poorly understood in many cancers. Since these mechanisms may also vary between individual patients, a mechanism-agonist, data-driven strategy could provide a practical option for customized combination testing in molecularly stratified subsets of patients, or even in individual cancer patients (18).

Computational data mining and modeling offer a systematic means for the characterization and prediction of combinatorial effects and for prioritizing the most potential combinations for preclinical testing (1, 13). A number of computational approaches, such as those based on network analysis, informatics measures, dynamic modeling, and machine learning, have been developed for guiding combinatorial drug designs (19, 20). Most of these in silico strategies have been developed and tested so far in established cell line models only (21, 22), or they require molecular response profiling carried out after treatments (3, 23). To the best of our knowledge, there are no computational approaches available for systematic tailoring of personalized drug synergy predictions for each patient individually based on functional compound testing in patient-derived cells combined with molecular and genomic profiling of the patients at baseline (before treatment).

We developed a drug combination prediction and testing (DCPT) platform that integrates baseline exome-seq and RNA-seq profiles with patient cell-based testing of single-drug responses ex vivo to predict the patient-specific, drug combination effects (synergy vs. nonsynergy) using a network pharmacology-based machine learning model. Subsequently, the most promising in silico synergies are tested in dose–response matrices in terms of their synergistic effects ex vivo, that is, whether they provide greater benefit in combination than expected by their individual effects in the patient sample. Importantly, DCPT makes use of healthy control cells to prioritize cancer-selective synergies that target mainly cancer cells and avoid combinations with overlapping toxicities. To facilitate mechanistic interpretation and clinical translation, the DCPT platform also implements patient-customized network visualizations to suggest potential biomarkers for the observed combination synergies among on- and off-targets of the compounds.

As the first application case, we applied DCPT platform to predict synergistic combinations for patients with T-cell prolymphocytic leukemia (T-PLL), a rare and aggressive blood cancer of mature T cells with median overall survival of less than 1 year (24). T-PLL is a highly clonal disease with a complex genetic background and most patients develop resistance to standard chemotherapeutic drugs, hence calling for novel targeted combination therapies. Because large-scale clinical trials are difficult to execute in such rare cancers, we utilized our novel computational-experimental approach to simulate an “n-of-one umbrella study” using ex vivo drug screening platform (11), combined with genomic and molecular profiling of the patient samples (12). Here, we show how the DCPT platform enabled us to make experimentally-confirmed synergy predictions for three T-PLL patient cases, each with differing resistance patterns and synergy mechanisms. The source-code and web-applications implementing the DCPT platform are freely available.

The DCPT platform (Supplementary Fig. S1) was implemented in this study on three T-PLL index patients (1508, 1263, and 1409) and three healthy donors (616, 1188, and 1189). The study was approved by the Helsinki University Hospital (Finland) ethics committee and it followed the Helsinki Declaration principles (12). The written informed consents were obtained from all the patients and healthy donors. Biomaterial samples from Aachen were used in accordance with the regulations of the biomaterial bank and the approval of the ethics committee of the RWTH Aachen Medical Faculty. Patients had to fulfil old and current WHO diagnostic criteria of T-PLL (25). Clinical characteristics and cytogenetic abnormalities of the patients are shown in Supplementary Table S1. The size and phenotypes of proliferating populations of T cells in the blood (CD4⁺, CD8⁺, or double positive CD4⁺CD8⁺ clonal phenotypes) were determined using the IO Test Beta Mark TCR Vβ Repertoire Kit (Beckman-Coulter Immunotech; Supplementary Fig. S2).

We used the FIMM oncology compound library, which is a comprehensive and evolving compound collection of both approved drugs and investigational compounds (Supplementary Table S2 details the complete list of the compounds and their primary on- and secondary off-targets used in the polypharmacologic modeling). In this work, the library contained 218 targeted compounds, out of which approximately half are kinase inhibitors and the rest are from other compound classes, such as differentiating/epigenetic or metabolic modifiers (Supplementary Fig. S3).

Single compound testing was performed on fresh peripheral blood mononuclear cell (MNC) samples (>80% leukemic cells) from T-PLL patients and unsorted MNC cells from healthy donors as described previously (11, 12). Briefly, 20 μL of single-cell suspension were plated into each well in a 384-well plate format where the compounds were dispensed in 10-fold dilution series of five concentrations. The concentration ranges were selected for each compound separately to investigate their full dynamic range of dose–response relationships. After 72 hours incubation at 37°C and 5% CO₂, cell viability of each well was measured using the CellTiter-Glo luminescent assay (Promega) and a Pherastar FS (BMG Labtech) plate reader. The percentage inhibition was calculated by normalizing the cell viability to negative control wells containing only 0.1% DMSO and positive control wells containing 100 μmol/L benzethonium chloride (BzCl; Supplementary Table S3).

Drug sensitivity score (DSS; ref. 26) was calculated to quantify the compound responses in the samples ex vivo (Supplementary Table S3). DSS is based on integration of the area under the fitted dose–response curve in a closed-form, which effectively transforms the multiparametric dose–response relationships into a robust single response metric (27). Based on the curve fitting with the four-parameter log-logistic function, DSS quantifies the compound efficacy by summarizing the four estimated model parameters: half-maximal inhibitory concentration (IC₅₀), slope, minimal and maximal asymptotes of the dose–response curve with the DSS R-package (https://bitbucket.org/BhagwanYadav/drug-sensitivity-score-dss-calculation). The scale of DSS score is 0 to 50.

The drug combination testing was done on a subset of those combinations predicted to be synergistic in any of the patient samples. All the model-predicted compound combinations were tested in an 8 × 8 dose–response matrix, with the same positive controls (BzCl) and negative controls (DMSO) as in the single-compound testing. The percentage inhibition was calculated by normalizing the cell viability to both negative control and positive control wells (Supplementary Table S4). In the dose–response matrix, the first row and column correspond to cells treated with single agent at seven increasing concentrations, whereas the remaining cells in the matrix layout measure the combined effects at various concentration points. The combination plate design and the testing procedure are detailed elsewhere (28).

DNA was isolated from sorted CD8⁺ and CD4⁺ T-cell tumor fractions from patients' peripheral blood MNC samples with the Nucleospin Tissue Kit (Macherey-Nagel). Exome sequencing and somatic mutation calling were performed as previously described (11, 29). The mutation profiles of each patient based on the somatic mutation calling are available in Supplementary Table S5. The exome-sequencing raw data are available from the corresponding author upon a reasonable request.

RNA was extracted from the CD4⁺ T-cell fractions (patient 1508 and 1409), CD4⁺CD8⁺ T-cell fractions (patient 1263) or CD3⁺ T-cell fractions (three healthy controls), with the miRNeasy Kit including DNAse I digestion (Qiagen). RNA sequencing and data processing was performed as previously described (29). Differential gene expression analysis was done with DeSeq2 R package (https://www.bioconductor.org/packages/release/bioc/html/DESeq2.html). The RNA-sequencing (RNA-seq) raw data are available from the corresponding author upon a reasonable request.

Since exhaustive experimentation of all possible drug and target combinations is not feasible among the 218 compounds tested, DCPT implements an efficient computational strategy for the prioritization of the most promising, selective and clinically actionable multitargeting drug synergies for each patient individually. Our target-based ensemble model is an extension of the target inhibition interaction using maximization and minimization averaging (TIMMA) model, which pinpoints the cancer-specific targets and then predicts the effects of drug combinations based on the interactions between drugs and the identified targets (30, 31). More specifically, DCPT approach combines the ex vivo sensitivity profiles of the patient's response to the 218 single compounds, with exome-seq and RNA-seq profiles of the same patient, through a systems-wide compound-target interaction network from our crowdsourcing bioactivity data platform, DrugTargetCommons (https://drugtargetcommons.fimm.fi/). For each compound-target pair, we calculated the median level of the dose–response bioactivity measurements (log-transformed K_d, K_i, or IC₅₀ endpoints), and compared the median of a particular target with the smallest bioactivity value among all the targets of the compound (32). A compound-specific cutoff [log-fold-change (FC) ≤ 2] was then applied to classify the protein targets into potent targets and nonpotent targets. Using this binary classification, we compiled a compound–target interaction matrix among 218 compounds and 366 targets (Supplementary Table S6).

For polypharmacology-based predictive modeling of combination responses, we formulated a patient/control sample-specific feature vector for each compound combination. A target feature vector of binary values defines which of the 366 targets are potent for each single compound or their combinations (Supplementary Fig. S4). For a single compound A, all of its potent targets are marked as 1, and the rest of the kinases are marked as 0. Similarly, for a compound combination A+B, all the potent targets of compound A and compound B are marked as 1 in the target feature vector. The target profiles were then combined with the genomic and molecular profiles of the same sample, extracted from the exome sequencing and patient/control gene expression data from RNA-seq (Supplementary Fig. S4). In the exome-sequencing data, genetic mutations were identified based on the somatic variant calling significance cut-off (P ≤ 0.01), in the VarScan2 algorithm (33). In total, 68 mutated genes were identified in the three patient cases (Supplementary Table S5). Thus, the patient-specific mutation feature vector is a binary vector of 68 entries in which 1 indicates a mutation in a particular patient. We further selected a set of 634 genes associated with cancer development, tumor suppressors, and drug response or resistance, and the raw counts and normalized FPKM data were utilized to construct the patient-specific gene expression vector (Supplementary Fig. S4).

We applied a supervised learning approach to predict the quantitative combination responses, using the random forest algorithm, which performs an ensemble machine learning by averaging multiple deep decision trees. Specifically, the above-defined compound, molecular and genomic vectors were used as input features to train a model that predicts single-compound DSS values in each sample separately. We performed leave-one-out cross validation (LOO-CV) to evaluate the model accuracy and to control for model overfitting (Supplementary Fig. S5, top). Using the trained, sample-specific models, we predicted the compound combination responses by assuming that a compound combination is a new “compound” with a target vector combined from those of the single compounds (Supplementary Fig. S5, bottom). For each sample, we then predicted the responses to all possible pairwise combinations among 218 compounds (23,653 combinations). To make binary classifications (synergistic vs. nonsynergistic) for each sample individually, the predicted combination responses were compared against the maximum response of the predicted single compound responses using the highest single-agent (HSA) model: if the response of a combination A+B was higher than max(A, B) response, then the combination A+B was classified as synergistic; otherwise, it was classified as nonsynergistic. Further, it was required that the combination found synergistic in the patient sample must be classified as nonsynergistic is the healthy control (Supplementary Fig. S6). To focus on nontoxic combination effects that are not dominated by any single compound, the lists of potential combinations were further narrowed based on the single-compound responses (DSS <20 both in the control and patient samples, although some exceptions were allowed where DSS <20 was achieved only when excluding the highest concentration point). The remaining combinations were then identified as patient-specific and cancer-selective combinations, which are less likely to present with severe side effects, and therefore easier to translate into clinical practice. Open-source R scripts are available for the predictive modeling (https://github.com/hly89/ComboPred).

In the next phase, the model-predicted synergistic combinations were experimentally-validated ex vivo using dose-response matrix testing. The degree of combination effects at various concentration ranges were quantified systematically with the computational tool SynergyFinder, using its open-source R-package (https://bioconductor.org/packages/release/bioc/html/synergyfinder.html), as well as web application (https://synergyfinder.fimm.fi/; refs. 28, 34). The synergistic or antagonistic effect of a pairwise compound combination was evaluated by comparing the observed combination response against the expected combination response, estimated with a reference model under the assumption of noninteraction between the two agents (35). In this work, the zero potency interaction (36) reference model was used to quantify the tested combinations, visualized as two-dimensional (2D) synergy/antagonism heatmaps and 3D landscapes. The overall level of synergy/antagonism was summarized over the full dose–response matrix using the delta score, which quantifies the combination effect beyond the expected interaction through potency changes of the dose-response curves between the individual drugs and their combinations (36). In the synergy testing, we focused on lower concentration ranges that are often clinically more tolerable.

In the last step of the pipeline (Supplementary Fig. S1), cancer vulnerability networks were constructed for each case separately based on the patient-specific combination predictions. The pathway cross-talk network was constructed by the shortest paths that connect the selective drug targets of the predicted combinations to the patient-specific genetic aberrations and molecular changes, including mutated or dysregulated genes, through comprehensive cancer signaling networks (37, 38). The integrated cancer signaling network includes 6,296 proteins and 79,083 interactions. The patient-customized vulnerability network enables a systematic investigation of the mechanisms of action of the drug combinations in the patient's cellular context to reveal potential biomarkers that are predictive of the synergistic responses. We have made available an easy-to-use R/Shiny web-application (https://patientnet.fimm.fi) that takes as input patient-specific exome/RNA-seq data, as well as single-compound responses and target annotations, and outputs the patient-customized network diagram visualized either in the web-browser (downloadable in an html format) or in Cytoscape (versions 2.7 and 3.4). Open-source R scripts for the network construction and visualization algorithm are also available at https://github.com/hly89/PatientNet.

Combination predictions were successfully carried out with the DCPT platform for three T-PLL patients with varying phenotypic and genetic characteristics (Fig. 1A; Supplementary Table S1), complemented with patient-customized network visualizations that connect the on- and off-targets targets of the most effective drugs with the patient's genomic and molecular alterations for finding potential mechanisms for the observed responses (Supplementary Fig. S1). The model predictions were ranked according to their predicted synergy classification (Materials and Methods), and the top predicted combinations were provided for the translational expert (S.M.). The clinically most applicable combinations were then tested on patient samples ex vivo using the dose–response matrix assay (Supplementary Table S7). The following subsections detail the successful synergistic predictions for each patient case (Fig. 1B). In total, 10 of 24 (42%) of the synergistic predictions were experimentally confirmed in the patient-derived samples (Supplementary Fig. S7).

The first patient case (1263) presented with a generally sensitive ex vivo single-compound sensitivity profile based on her treatment-naïve diagnostic sample (Supplementary Fig. S8). The DCPT platform was still able to identify selective synergies among individually potent agents (Fig. 1B). One such unexpected, beyond-additive effect included a combination between the investigational protein kinase C activator bryostatin 1 and the apoptotic modulator nutlin-3, which inhibits the interaction between MDM2 and p53 and thereby causes p53 activation at higher concentrations (Fig. 2A; Supplementary Fig. S9). To further investigate the role of MDM2-p53 axis in this patient case, we combined bryostatin 1 with three other MDM2 inhibitors (Supplementary Fig. S10). Selective synergy was confirmed already at lower concentrations of SAR405838 (MI-77301), a small-molecule agent that is currently in early-phase clinical trials for a number of malignancies. Similar to nutlin-3, SAR405838 effectively stabilizes p53 and activates the p53 pathway (39), and therefore these agents might work best on cancers that contain a wild-type p53, as in 1263 (Fig. 2B). The two other MDM2 inhibitors (idasanutlin and AMG-232) did not show synergy in this patient case.

Even though the bryostatin 1 responses are interesting for biological discovery, these results may be difficult to translate into clinical treatments due to dose and exposure-dependent activity of bryostatin 1. Toward more clinically actionable combination for this patient, we further predicted and experimentally validated combination effects among approved drugs, including a selective synergy between the BCL2 inhibitor venetoclax and tacrolimus, which binds to intracellular FKBP12 (Fig. 3A; Supplementary Fig. S11). Tacrolimus is a T-cell targeted immunosuppressive drug that inhibits calcineurin, a calcium- and calmodulin-dependent serine/threonine protein phosphatase, which activates T cells of the immune system. It has been shown that calcineurin mainly activates NFAT transcription factors that are overexpressed in T-PLL tumor cells (40), including patient 1263 (NFATC4 log-FC>1.5). We further confirmed the selective synergy by combining venetoclax with another approved drug, temsirolimus, which similarly binds to FKBP12, and results in inhibited activity of mTORC1. This synergy was observed only at lower concentrations of the two drugs (Supplementary Fig. S12), partly explaining why this combination resulted in a lower overall synergy compared to the other combinations in this patient case (Supplementary Fig. S13).

As a last step of DCPT pipeline, we visualized the patient-customized pathway cross-talks by connecting the combinatorial targets to the patient-specific genetic and molecular aberrations through shortest signaling paths. The cancer vulnerability network diagrams for all the selective combinations in the patient 1263 collectively suggest that the overall sensitivity and the identified selective synergies in this case may be driven by the MAPK pathway (Fig. 2B and 3B; Supplementary Fig. S13). Even in the absence of any JAK mutations in this patient case with double-positive CD4⁺CD8⁺ clonal phenotype (Supplementary Table S1), further support for the functional importance of the JAK–MAPK pathway rewiring came from another predicted 1263-specific combination, which showed the highest selective synergy between bryostatin 1 and the investigational JAK1/2 inhibitor momelotinib (Fig. 1B; Supplementary Fig. S13). Momelotinib is being developed as a drug for myelofibrosis and it currently undergoes phase III clinical trials (ClinicalTrials.gov).

As the second application case, we demonstrate how the DCPT platform enables identification of targeted agents that enhance the efficacy of standard chemotherapy in another patient case (1409). In this case, the single-compound responses were profiled ex vivo after the patient's treatment with alemtuzumab, a monoclonal antibody that binds to CD52 on the surface of mature T lymphocytes, hence marking the CD52 proteins on the leukemic T cells to allow the immune system to pick out and kill them. We predicted and confirmed a selective synergistic combination effect in the post-treatment phase between the proteasome inhibitor carfilzomib, approved for the treatment of relapsed or refractory multiple myeloma, and STAT3 inhibitor stattic (Fig. 4A; Supplementary Fig. S14). Even though the synergistic effect was only moderate (delta score of 5.1), it was surprising given that we focused on the lower concentration end of carfilzomib (0–3 nmol/L). The target-based machine learning model was able to make the response prediction for this second-generation proteasome inhibitor by making use of its potency toward proteasome subunit beta 8 (PSMB8; see Fig. 4B).

On the basis of the network diagram (Fig. 4B), we hypothesized that the synergistic effect is driven by the rewired cross-talks between PI3K/mTOR pathways in this patient case. To test this hypothesis, we carried out additional combination experiments among compounds targeting mTOR and/or PIK3CA pathways using two investigational kinase inhibitors, sapanisertib and PF-04691502, based on the DCPT predictions (Fig. 4B; Supplementary Table S7). Each of the three pairwise combinations showed selective synergy in patient 1409 cells when compared to the healthy control cells (Fig. 4C). This example case demonstrates how the DCPT platform enables making hypotheses about the selective synergies based on the patient-customized vulnerability networks, which can be then followed-up and validated (or invalidated) using additional target perturbations. Compared to the other patient cases, 1409 was not treatment-naïve (Supplementary Table S1), which may explain the somewhat lower degrees of synergistic effect in this patient case (Fig. 1B and 4C). However, it should be noted that already the added cancer cell killing effect beyond the single-agent efficacies is often considered enough for clinical therapy benefits (as scored using the HSA model).

The final index patient (1508) presented with a challenging, generally resistant ex vivo single-compound response profile based on his treatment-naïve diagnostic sample (Supplementary Fig. S8). The overall single-compound resistance phenotype may be driven by an activating mutation in β-catenin (CTNNB1, K672Q), which has been shown to induce malignant T-cell proliferation by promoting genomic instability (41). Using the DCPT platform, we predicted and confirmed a relatively strong synergistic efficacy (delta synergy score of 10.14) by combining two agents that alone had relatively low responses (Fig. 5A; Supplementary Fig. S15): the potent ERK2 (MAPK1) inhibitor VX-11E and the hormone therapy progestogen megestrol that acts also as an agonist of the glucocorticoid receptor NR3C1 (nuclear receptor subfamily 3, group C, member 1).

Based on the pathway cross-talk network constructed by connecting the four targets, MAPK1, PGR, AR, and NR3C1 to the patient-specific genetic and molecular aberrations (Fig. 5B), one can suggest that the selective synergy is linked to the IL2 and IL5 signaling pathways, which based on the exome and RNA sequencing profiles, are activated in this patient due to the gain-of-function mutations in IL2 gamma receptor (IL2RG, K315E; ref. 42) and signal transducer and activator of transcription 5B (STAT5B, P702S; Supplementary Table S1), combined with IL5 and JAK1 overexpression (Fig. 5B). The STAT5B–IL2RG mutation combination was not seen in the other patients in our T-PLL cohort (12), potentially explaining why the other patient cases showed nonsynergistic responses to this 1508-tailored combination prediction (Fig. 1B and 5A).

Using T-PLL as the first case study, we showed that the DCPT platform successfully predicted at least one synergistic combination for each of the patients, each presenting with unique resistance patterns and synergy mechanisms. In total, 10 of 24 (42%) of the model predictions were experimentally confirmed to show synergy (Supplementary Fig. S7), which is markedly higher than that expected by random sampling of drug pairs; ca. 4% to 14% compound pairs are detected as synergistic in high-throughput combination screens (22, 43–45). Importantly, the predictions from the DCPT platform are patient-specific, which poses a much more challenging prediction problem than merely identifying generally toxic combinations that can kill most cancer cells, but which may also lead to severe toxicities to normal cells. The predicted combinations from the DCPT platform were evaluated in terms of differential synergy, compared to both healthy cells and other T-PLL patient cells, which is likely to lead to clinically better tolerated combinations. Our results emphasize the importance of ex vivo validation of the predicted synergies using multiple agents that inhibit the patient-specific targets or pathways. Compared with traditional drug-centric approaches, a benefit of our target-centric approach is the possibility to follow-up the target combinations using several agents with the same primary targets, but differing mode of actions, which may lead to drastic differences in combination effects.

The selective synergies predicted among already approved drugs represent arguably the most straightforward repurposing opportunities for T-PLL. These included combinations of tacrolimus and temsirolimus with BCL-2 inhibitor venetoclax, which has been approved for relapsed/refractory chronic lymphocytic leukemia patients. Interestingly, venetoclax was recently shown to provide clinical responses already as single-agent in two late stage refractory T-PLL patients (46). Other findings that may become clinically applicable in the near future involve those selective synergies in which late-stage investigational agents sensitize approved chemotherapies, such as proteasome inhibitor carfilzomib, which is being used for the treatment of relapsed/refractory multiple myeloma. As most T-PLL patients also develop resistance to standard chemotherapeutics, and the resistance mechanisms vary greatly from patient to patient, there is a critical need for patient-customized combination therapies. T-PLL is a very rare disease, making randomized or stratified clinical trials difficult to execute in large enough patient cohorts that could show clinical efficacy of the selective combination effects. Therefore, individualized precision oncology approaches are critically needed for rare cancers such as T-PLL.

It has been long understood that the resistance mechanisms are due to complex biological factors, such as pathway rewiring, redundancy and cross-talks (13). This translates into the need of polypharmacologic approaches, such as DCPT, which make the drug target combination predictions based on overlapping target profiles among hundreds of agents targeting various pathways. Single-compound functional testing is becoming increasingly popular in the next-generation precision oncology approaches (7–12), enabling systems-wide drug-target interaction network modeling for each patient sample separately. In our approach, signaling networks are used only in the very last phase, for mechanistic interpretation using patient-specific network, which is different to those approaches that consider pathway cross-talk already in the prediction phase (22, 47). Signaling pathways and metabolic networks are still incompletely mapped in most cancer-contexts, in which wild-type interactions are often rewired due to genetic and epigenetic aberrations. Therefore, we chose not to use in this pilot such incomplete information in the prediction phase, but once comprehensive and accurate enough network information is available for a given cancer type, DCPT can use it also in the prediction phase (e.g. in terms of extending from target proteins to target pathways).

Our original TIMMA prediction model has been systematically evaluated previously in cell line models using RNAi and CRISPR/Cas9 target validations (30, 31). This is the first application of the extended model to patient-derived samples. DCPT platform extended the original model by using exome/RNA-seq profiles with ex vivo testing of single-drug responses to predict patient-specific drug combination effects, and importantly, by making use of healthy control cells to prioritize cancer-selective combinations. DCPT shares similarities with the iterative experimental-computational algorithm TACS (48), in terms of using the healthy control cells to avoid toxic effects, or effects related only to the cell phenotype. The multistep TACS is limited to <50 compounds, due to being able to search for optimal combinations among an arbitrary number of dugs (not only two-drug combos), whereas DCPT can deal with larger compound panels (here, >200), because of its effective two-stage design (single-compound profiling, followed by combination predictions and experimental validation). It should be noted that although the use of healthy controls can improve the cancer-selectivity of the predictions, control samples are not always available in clinical applications, or in drug sensitivity profiling of cell line panels.

An additional case study in another leukemia type, acute myeloid leukemia (AML), demonstrates how the DCPT approach is applicable also without healthy controls for the purpose of finding common synergistic combinations across diseased samples (Supplementary Fig. S16). In this case study, DCPT predications were experimentally validated in eight AML patient samples for three given combinations (ruxolitinib and venetoclax, WEHI-539 and venetoclax, and ruxolitinib and quizartinib); the first two combinations were found synergistic and the last one nonsynergistic in our previous study (49). This case study enabled us not only to confirm a high true positive rate of the DCPT platform in another leukemia type [true positive rate (TPR) = 0.625], but also to estimate the false negative rate of the prediction algorithm [false negative rate (FNR) = 0.375]. These results indicate that the DCPT platform identifies synergistic combinations in AML cases with similar false positive rate as in the T-PLL cases [false positive rate (FPR) = 0.13 and FPR = 0.19, respectively; see Supplementary Fig. S7]. We note these results are only AML-specific, not patient-selective, as we lack exome and RNA sequencing data for the healthy AML samples. As expected, there was also a patient-to-patient variability in these established synergistic combinations (Supplementary Fig. S16), further highlighting the need of patient-customized drug combination predictions for leukemia patients.

As with most combination prediction algorithms (50–52), the current DCPT version does not predict the optimal concentrations of compounds in the combination ex vivo, rather it summarizes the synergy/antagonism over the whole dose-response matrix. Prediction of drug-dose-combinations may require the use of global optimization search algorithms (48, 53, 54), which will increase both experimental costs (as they often require iterative rounds of combination testing), as well as computational complexity (especially for larger compound sets), but might lead to clinically more tolerable low-concentration combinations when tested in parallel in healthy controls. However, it is important to note that the ex vivo predicted doses may not translate to relevant concentrations in vivo, even though our previous works on leukemia patient-derived samples have shown that, in general, ex vivo drug testing results correlate with the clinical responses (11, 55, 56). The clinical implementation of the current DCPT pipeline takes up to four weeks' time, but the process can be done in parallel for multiple patients simultaneously, and streamlined by faster exome-seq analysis, which is the current bottle-neck in the implementation. Once optimized, the same modeling principles could be used also when predicting targeted combinations inhibiting various cancer or resistance-driving cell populations, as soon as clone-level drug sensitivity profiling becomes available. Modeling of cancer heterogeneity is considered extremely important when implementing treatment predictions in clinical applications (57).

In conclusion, this work demonstrated how an integrated use of functional and genomic platforms enables patient-customized prediction and testing of selective synergies for T-PLL patients with the help of the DCPT platform. Since experimental testing of the exponentially increasing number of all drug/target combinations is infeasible in practice, and especially in clinical treatment applications, our approach guides the search of patient-specific synergies by making use of compound testing combined with exome/RNA-seq profiling of patient samples at baseline to narrow down the massive search space of all potential combinations. We stress the importance of careful ex vivo validation of the in silico predictions before clinical translation.

S. Koschmieder reports receiving a commercial research grant from Novartis, Bristol-Myers Squibb, Janssen and is a consultant/advisory board member for Novartis, Ariad/Incyte, Bristol-Myers Squibb, Janssen, AOP Pharma, CTI, Shire, Bayer. S. Mustjoki reports receiving a commercial research grant from Novartis, Pfizer, BMS, Orion, has received speakers bureau honoraria from BMS and Pfizer, and is a consultant/advisory board member for Novartis. No potential conflicts of interest were disclosed by the other authors.

Conception and design: L. He, J. Tang, E.I. Andersson, S. Mustjoki, T. Aittokallio

Development of methodology: L. He, J. Tang, T. Aittokallio

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): E.I. Andersson, S. Timonen, S. Koschmieder, S. Mustjoki

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): L. He, J. Tang, S. Koschmieder, K. Wennerberg, S. Mustjoki

Writing, review, and/or revision of the manuscript: L. He, J. Tang, E.I. Andersson, S. Koschmieder, K. Wennerberg, S. Mustjoki, T. Aittokallio

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): L. He, S. Timonen

Study supervision: J. Tang, S. Mustjoki, T. Aittokallio

Other (performing the drug combination screening): S. Timonen

The authors thank Aleksandr Ianevski (FIMM), Matti Kankainen (FIMM and HUSLAB), and Samuli Eldfors (FIMM) for technical assistance; Riikka Karjalainen (FIMM), Minxia Liu (FIMM), and Juho Miettinen (FIMM) for the AML case study data; and Carlos Cuesta-Mateos (Hospital Universitario de la Princesa, Spain) for T-PLL patient material. This work was supported by funding from the Integrated Life Sciences (ILS) doctoral program (to L. He), Biomedicum Foundation (to E.I. Andersson), Finnish Cultural Foundation (to E.I. Andersson), European Union's Horizon 2020 Research and Innovation Programme 2014-2020 under Grant Agreement No. 634143 (MedBioinformatics; to T. Aittokallio), the European Research Council (ERC) Starting Grant Agreement No. 716063 (DrugComb; to J. Tang), and ERC consolidator grant agreement No. 647355 (M-IMM; to S. Mustjoki), the Academy of Finland (grants 292611, 269862, 272437, 279163, 295504, 310507 to T. Aittokallio; 277293 to K. Wennerberg; 287224 to S. Mustjoki), Sigrid Jusélius Foundation (to K. Wennerberg, S. Mustjoki), Cancer Society of Finland (to T. Aittokallio, K. Wennerberg, S. Mustjoki), Finnish Cancer Institute (to S. Mustjoki), Signe and Ane Gyllenberg Foundation (to S. Mustjoki), and State funding for University-level Health Research in Finland (to S. Mustjoki).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.