Over 600,000 patients with solid tumors die annually in North America, including approximately 5,000 sarcoma-related deaths each year. A large segment of these patients are those who have exhausted standard clinical pathways, including patients with recurrent, refractory, or rare cancers for which no established treatment option remains. The medical community has invested heavily in genomic sequencing as a personalized medicine strategy, yet tumor DNA sequencing leaves approximately 60% of patients without an actionable result, meaning no druggable mutation is found or no approved targeted therapy exists for the identified variant.
The combination therapy problem: Even when an actionable mutation is identified, single-agent targeted therapy frequently fails to provide durable disease control. Sarcomas, which encompass more than 60 histological subtypes, are especially challenging in this regard. Their biological heterogeneity means that pathways driving cell survival in one patient's tumor may be entirely different from those in another's, and even within a single tumor, distinct spatial regions can harbor different molecular drivers. Effective multi-drug combination design demands a methodology that can account for this complexity at the individual patient level.
The PTIM framework: This 2019 paper from Oregon Health and Science University and collaborators presents proof-of-concept validation of the Probabilistic Target Inhibition Map (PTIM) framework, a computational modeling approach that integrates high-throughput single-agent drug screening data with molecular sequencing results (exome-seq, RNA-seq, phosphoproteomics, siRNA knockdown) to design personalized two-drug combinations. The pipeline was tested across three distinct sarcoma subtypes: alveolar rhabdomyosarcoma (aRMS), epithelioid sarcoma (EPS), and undifferentiated pleomorphic sarcoma (UPS), each validating a different major challenge in personalized oncology.
The paper addresses three specific unmet needs: predicting synergistic drug combinations from functional data alone, identifying consensus therapies across heterogeneous tumor regions, and preventing cancer cell resistance by targeting independent survival pathways simultaneously. The platform is designed for patients for whom DNA sequencing is uninformative and standard therapeutic options have been exhausted.
The Probabilistic Target Inhibition Map (PTIM) methodology treats cancer cell survival as a biological circuit composed of gene targets arranged in series and parallel configurations. In Boolean logic terms, targets in series require inhibition of any one member to slow tumor growth (OR logic), while targets in parallel require simultaneous inhibition of all members to achieve growth suppression (AND logic). The key insight is that cancer cells failing to respond to any single drug may nonetheless be highly sensitive to a specific two-drug combination that simultaneously blocks both parallel survival pathways.
Input data requirements: PTIM modeling requires drug screening data from at least 60 monotherapy agents, each with quantified drug-target interaction EC50 values. The promiscuity of targeted compounds is deliberately exploited: drugs that inhibit overlapping but non-identical target sets create a rich dataset from which the underlying target logic can be inferred. In this study, four compound libraries were used: the Pediatric Preclinical Testing Initiative Screen Version 2.1 (PPTI, 60 agents), the GlaxoSmithKline Open Access Orphan Kinome Library (GSK, 402 compounds with quantified drug-target EC50 profiles over 300 protein targets), the Roche Orphan Kinome Screen Library (Roche, 223 novel kinase inhibitors), and a custom investigator-selected 60-agent screen denoted Drug Screen V3. Cell viability was measured using CellTiter-Glo luminescent assay in 384-well plates at 72 hours.
Feature selection approach: The PTIM algorithm approaches sensitivity prediction as a machine learning feature selection problem where "features" are the gene targets inhibited by individual drugs. The objective function identifies target combinations (feature sets) that group sensitive and insensitive drugs into binary bins with low intra-bin variance. Each resulting block is assigned a scaled sensitivity score derived from the IC50 values of all drugs assigned to that block. Blocks are visually rendered as circuit diagrams: single targets are single points of failure, two-target combinations appear as rectangular blocks, three- and four-target combinations as circular blocks with increasing inhibitor symbols.
Molecular data integration: Secondary biological datasets refine PTIM models by modifying target inclusion probability. RNA-seq data eliminates targets with expression below 50% of matched normal tissue (removing likely false positives from drug screens). Exome-seq data increases inclusion probability for mutated or copy number-altered targets. siRNA knockdown results force inclusion of validated single-target sensitivity mechanisms. Phosphoproteomics data adds targets showing differential activation in tumor versus normal tissue. Each data type is incorporated through formal mathematical constraints on the PTIM objective function, described in full in the methods section.
The first proof-of-concept validation used a low-passage primary tumor cell culture (designated U23674) established from a genetically engineered mouse model (GEMM) of alveolar rhabdomyosarcoma (aRMS). This Myf6Cre/Pax3:Foxo1/p53 mouse model is a well-characterized genetic facsimile of the human disease. Three kinase inhibitor libraries were screened on U23674: the GSK library (305 compounds, of which 40 or 13% were "hits" producing at least 50% cell growth inhibition), the Roche library (223 compounds, 21 hits or 9.4%), and the PPTI screen (60 compounds, 28 hits or 46.7%). The higher hit rate on the PPTI screen reflects its targeted composition of clinically staged agents known to be broadly active against pediatric cancers.
Molecular sequencing findings: Whole exome sequencing of U23674 identified six genes with activating mutations (Fat4, Gm156, Mtmr14, Pcdhb8, Trpm7, Ttn, Zfp58) and one high-impact frameshift indel (Ppp2r5a), but none were druggable. Copy number gain was identified in four druggable targets (Gsk3a, Epha7, Psmb8, Tlk2), but RNA-seq showed neutral or reduced expression for three of these, and the fourth had no available clinical-stage inhibitor. Critically, this demonstrates the 60% problem explicitly: exome sequencing alone provided no actionable therapy for U23674.
PTIM-guided combination selection: The GSK screen, with its 24 compounds per target on average and thorough drug-target quantification, was selected as the primary dataset for PTIM modeling. Baseline, RNA-seq-informed, exome-seq-informed, siRNA-informed, and phosphoproteomics-informed PTIM models were generated. The RNA-seq-informed model highlighted the Igf1r and Pik3ca target combination as the highest-scoring two-drug block, corresponding to the clinical agents OSI-906 (an IGF1R/INSR inhibitor) and GDC-0941 (a PI3K/mTOR inhibitor). A baseline model also identified a three-target combination (Igf1r, Insr, Pka) targetable by OSI-906 plus SB-772077-B (a PKA inhibitor).
In vitro and in vivo validation: In vitro validation of OSI-906 + GDC-0941 demonstrated synergy by non-constant ratio Combination Index (CI) analysis using CompuSyn software, with CI values calculated across functionally relevant dose ranges (approximately 10 nM to 5 uM for OSI-906, 5 nM to 1 uM for GDC-0941). Low-dose experiments at 175 nM OSI-906 and 50 nM GDC-0941 confirmed the predicted mechanism of action. In vivo, a four-arm orthotopic allograft study (n=8 mice per arm) treated mice with vehicle, 50 mg/kg OSI-906, 150 mg/kg GDC-0941, or the combination. Kaplan-Meier analysis showed a significant survival benefit for the combination arm versus vehicle (p=0.005, Bonferroni-corrected) and versus OSI-906 alone (p=0.014). Neither monotherapy arm differed significantly from vehicle (p greater than 0.5).
A critical insight from the aRMS experiments was the demonstration of rapid tumor cell rewiring following drug exposure. After treating U23674 cells with low-dose monotherapy or combination therapy using the PTIM-predicted OSI-906 and GDC-0941 agents, the researchers screened the surviving cell populations again using the Roche kinome library. Within hours of drug exposure, the surviving cell populations showed evidence of altered drug sensitivity profiles, indicating activation of secondary signaling pathways as a survival mechanism.
Implications for treatment design: This rewiring phenomenon is not unique to sarcoma, but it is particularly relevant in the context of PTIM modeling. The PTIM circuit representation specifically addresses rewiring through its parallel-block architecture: when a cancer cell is exposed to a drug that blocks one survival pathway (one PTIM block), the cell can up-regulate alternative pathways (other PTIM blocks) to maintain proliferation. This observation strongly supports the rationale for simultaneous multi-pathway targeting rather than sequential monotherapy escalation.
The rewiring experiments also highlighted a limitation of synergy-focused two-drug combinations that target a single PTIM block: while achieving synergy, they may still leave parallel survival pathways unaddressed, allowing resistant clones to expand. This finding directly motivated the third proof-of-concept experiment on UPS, where the goal was to block two independent PTIM blocks simultaneously rather than target a single synergistic combination block.
The PTIM framework can theoretically extend to three- or four-drug combinations by targeting additional blocks, providing a path toward more durable disease control. However, the authors note that the practical feasibility of such regimens depends on demonstrating adequate safety and tolerability in Phase I studies, which presents a significant translational hurdle for novel drug combinations not previously evaluated together in human subjects.
Epithelioid sarcoma (EPS) is a rare soft tissue sarcoma of children and adults for which chemotherapy and radiation provide minimal survival benefit beyond wide surgical excision. The biological hallmark of EPS is loss of the INI1 protein (encoded by SMARCB1), which occurs in approximately 93% of cases. The researchers acquired a resected tumor from a 22-year-old female patient with a large proximal (shoulder) EPS tumor (assigned identifier PCB490). Because of the tumor's size and the biological expectation of spatial heterogeneity in solid tumors, the approximately 3 cm2 resected mass was divided into five spatially-distinct regions (PCB490-1 through PCB490-5), each of which was cultured separately to generate heterogeneous cell models.
Confirming heterogeneity: Three of five regions (1, 2, and 5) were confirmed as EPS by western blot for INI1 loss, consistent with the disease's molecular hallmark. Cultures were maintained at low passage (passage 2 or below) to minimize biological drift from the original tumor. Drug screening via Drug Screen V3 (60 agents) and the Roche screen was performed on the three successfully grown cultures (PCB490-3, PCB490-4, PCB490-5). Spearman correlation analysis revealed that correlation within PCB490 cultures (mean r=0.7671) was significantly higher than correlation between PCB490 and the second patient-derived culture PCB495 or established EPS cell lines (mean r=0.4601, p less than 0.001), validating that each patient's tumor has a unique drug sensitivity profile. Importantly, sensitivity differences existed between cultures from regions millimeters apart within the same tumor, confirming clinically relevant spatial heterogeneity.
Sequencing results and the 60% problem: Whole exome sequencing of PCB490 identified germline and tumor variants in five genes (ABL1, NOTCH1, MDM4, PAK4, MAP4K5), all of unknown clinical significance and all also present in matched normal tissue, indicating germline rather than somatic origin. Drug screening results showed no pathway-specific sensitivity that correlated with these variants. Once again, DNA sequencing alone yielded no actionable therapeutic target, reinforcing the study's central premise.
Heterogeneity-consensus drug combination: PTIM models were generated separately for PCB490-3, PCB490-4, and PCB490-5 (the last with RNA-seq integration). The models identified common high-scoring blocks across all three heterogeneous sites: epigenetic modifiers (HDAC, EHMT), PI3K/mTOR signaling, and VEGF receptor (KDR) signaling. The investigators selected BEZ235 (a PI3K/mTOR inhibitor) and sunitinib (a poly-kinase inhibitor targeting KDR and AXL among others) as the consensus two-drug combination. This selection was made solely from PTIM modeling data, without prior knowledge of sunitinib's activity in EPS.
Having identified BEZ235 + sunitinib through PTIM modeling alone, the investigators bypassed in vitro validation and proceeded directly to in vivo testing in a patient-derived xenograft (PDX) model. This decision was deliberate: by skipping cell culture-based validation, the study more closely replicated the intended clinical workflow where a patient's tumor is screened, PTIM analysis is performed, and a combination therapy is selected for direct administration, minimizing the time and tissue required.
PDX model development: The PCB490 PDX was established at The Jackson Laboratory (model J000078604) by implanting fresh surgical tumor tissue into immunodeficient NSG (NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ) mice within approximately 24 hours of surgery, without prior in vitro culturing. Engrafted tumors were validated at each passage by STR fingerprinting to ensure model provenance, by histological review by a board-certified pathologist for morphological concordance with the primary tumor, and by immunohistochemistry for INI1 loss and human markers (Ki67, vimentin). Human CD45 staining was performed to screen for lymphomagenesis, a known artifact in PDX models.
Treatment results: For the in vivo study, female athymic nude mice (Envigo Foxn1nu) were implanted with PCB490 PDX fragments and treated once tumors reached 150-250 mm3. Treatment arms were vehicle control, sunitinib alone (30.0 mg/kg PO QD for 21 days), BEZ235 alone (25.0 mg/kg PO QD for 21 days), and BEZ235 + sunitinib combination (n=3 mice per group). Tumor volume was measured twice weekly with digital calipers. At Day 19, the combination arm showed 92% slower tumor growth compared to vehicle (p=0.01). In statistical analysis restricted to treated animals, BEZ235 + sunitinib significantly outperformed both BEZ235 alone (p=0.01) and sunitinib alone (p=0.01), confirming the combination's superiority over either monotherapy.
The result is particularly notable because the combination was selected based entirely on computational modeling of drug screen data from primary tumor cell cultures, with no prior clinical or preclinical evidence specifically guiding the suitability of this pairing for EPS. It represents a direct demonstration that PTIM ensemble modeling can identify clinically relevant heterogeneity-aware combinations without relying on established disease-specific drug knowledge.
Undifferentiated pleomorphic sarcoma (UPS) is a high-grade soft tissue sarcoma with limited systemic treatment options. The third PTIM validation addressed a fundamentally different clinical challenge: rather than predicting synergy within a single target block, the goal was to identify two drugs targeting independent biological pathways (separate PTIM blocks) that together prevent cancer cell resistance by closing off all major survival routes simultaneously.
Cross-species experimental design: The researchers screened two UPS models using the PPTI kinase inhibitor library: a human-derived UPS cell culture (PCB197, from a 75-year-old male) and a canine UPS cell culture (S1-12, from Oregon State University's veterinary medicine program). The rationale for including a canine model was that cross-species consensus would strengthen confidence that any identified combination reflects true biological mechanisms rather than model-specific artifacts. PTIM models of both species identified common sensitivity blocks corresponding to two independent pathways: HDAC7 inhibition and MCL1 inhibition.
Drug selection and experimental arms: Panobinostat (a pan-HDAC inhibitor) and obatoclax (an MCL1 inhibitor) were selected as the two drugs targeting the consensus cross-species PTIM blocks. To demonstrate resistance abrogation, a six-arm in vitro experiment was conducted for each species in quadruplicate (n=4 wells per arm, 10,000 cells per well on 6-well plates). The six arms were: vehicle control; obatoclax monotherapy for 6 days; panobinostat monotherapy for 6 days; obatoclax for 3 days followed by panobinostat for 3 days; panobinostat for 3 days followed by obatoclax for 3 days; and simultaneous obatoclax + panobinostat for 6 days. Drug concentrations were set at 1.5 times the EC50 of the respective PTIM target, and both concentrations had to fall below maximum clinically achievable exposures (Cmax).
Results: All five arms except simultaneous combination treatment showed cellular regrowth within 100 days, indicating the development of resistance. Both monotherapy arms and both sequential arms permitted tumor cell recovery. Crucially, in both the human PCB197 and canine S1-12 models, the simultaneous panobinostat + obatoclax combination arm showed zero cellular regrowth over the full 100-day observation period. This cross-species concordance strongly supports that the resistance-abrogating effect reflects the underlying biology of UPS rather than species-specific artifact, and that blocking two independent survival pathways simultaneously prevents the cellular rewiring that enables resistance.
Sample size and proof-of-concept scope: The authors are explicit that these three validation experiments constitute proofs-of-concept rather than clinical validation. Each in vivo experiment used only 3 to 8 mice per arm, insufficient for definitive conclusions about efficacy at scale. Each represents a single patient's biology (with the exception of the cross-species UPS study). The critical next phase requires n-of-1 prospective pilot testing in partnership with physicians and veterinarians treating individual human patients and animals with spontaneous cancers, where real-world variability and tumor evolution can be assessed.
Cost and regulatory constraints: The PTIM pipeline is designed to be cost-accessible. At the time of publication, functional drug screening in 384-well plate format could be performed for under $300 per compound library, and CLIA-certified sequencing experiments cost under $500 per analyte. The PTIM computational analysis itself can be completed in under two weeks, faster than the typical turnaround for clinical whole-genome sequencing. However, translating PTIM-predicted combinations to patients faces two substantial barriers: most PTIM-guided combinations involve FDA-approved drugs used off-label in combinations not validated in Phase I trials, raising safety concerns; and the financial costs of modern targeted therapy regimens may be prohibitive for individual patients.
Disease-specificity of screening panels: The paper notes that optimal PTIM modeling requires compound libraries designed around the known biology of specific tumor types. Kinase inhibitor libraries are highly relevant for sarcoma, where kinase signaling is central to disease biology, but different cancer types may require different compound classes to achieve adequate target coverage. Similarly, the selection and weighting of secondary molecular datasets (RNA-seq, exome-seq, proteomics) will depend on the available tissue quantity and the expected predictive utility of each data type for a given disease.
Future directions: The authors identify several critical extensions of the PTIM platform: incorporation of personalized toxicity and dosing prediction for minimizing adverse effects of novel combinations; application to direct-to-plate tumor screening to minimize dependence on cell culture establishment; integration of routine proteomics data as CLIA-certified proteomic workflows become available; and expansion of disease-consensus PTIM modeling to develop subtype-specific drug combinations through pooled analysis of multiple patient datasets. The MATLAB software package implementing the base PTIM algorithm has been published and made publicly available to support broader adoption and further development by the research community.