Performance analysis of a novel artificial intelligence-based computational method on published ex vivo drug sensitivity data to support targeted treatment decisions in precision oncology

e13618

Background: Comprehensive molecular profiling is readily available for clinical practice. An extensive amount of published evidence provides information about the potential functional relevance of many driver genes and genetic alterations. But, due to the large number of driver genes and alterations, and the long-tail frequency distribution of these alterations and their possible combinations, very few single driver alterations have reached high-enough-level evidence alone to support clinical decisions. We hypothesized that by aggregating evidence following logical principles of reasoning by a computational system, this vast molecular information can be used to improve personalized treatment decisions. In the present study, we used previously published ex vivo drug sensitivity data to analyze the performance of an artificial intelligence-based computational method, the digital drug assignment (DDA), which has shown utility by improving treatment decisions in case of complex molecular profiles acquired in the SHIVA01 trial. Methods: We selected 111 cases with whole-genome sequencing (WGS) and ex vivo drug sensitivity data of a previously published acute myeloid leukemia study (Tyner et al, 2018). WGS variants were filtered for a preselected hematology-related panel of 446 genes and uploaded to a DDA-based software system to calculate the aggregated evidence level (AEL) values of associated molecularly targeted agents. DDA-predicted sensitivity (or resistance) was defined as AEL > 0 (or AEL < 0) in the presence of at least one actionable driver. Area under the curve (AUC) values were used for determining the ex vivo sensitivity or resistance of leukemia cells to 40 approved drugs and 53 developmental compounds. Results: The AUC values were significantly different in the drug-sensitive and -resistant groups forecasted by the DDA (167.1 and 205.5, respectively, p < 0.0001) and differed significantly from the average AUC value (194.6). Overall, sensitivity was correctly predicted in 66% of compound-sample pairs (n = 671). 88 approved drugs had AEL value over 1000; of these 73% were effective according to the ex vivo results. While forecasted resistance was confirmed in 63% of the cases. With only the most sensitive/resistant 20% of cases considered from the ex vivo data, sensitivity was accurately predicted in 75% of approved compound-sample pairs (n = 173). 37 approved drugs had AEL value over 1000; of these 81% were confirmed as sensitive. Conclusions: The DDA-based computational reasoning has a promising performance in forecasting sensitivity and resistance to a broad spectrum of targeted agents based on molecular information. Therefore, it has the potential to automate, standardize and improve complex molecular profile-based targeted treatment decisions.

A computational method for prioritizing targeted therapies in precision oncology: performance analysis in the SHIVA01 trial

Precision oncology is currently based on pairing molecularly targeted agents (MTA) to predefined single driver genes or biomarkers. Each tumor harbors a combination of a large number of potential genetic alterations of multiple driver genes in a complex system that limits the potential of this approach. We have developed an artificial intelligence (AI)-assisted computational method, the digital drug-assignment (DDA) system, to prioritize potential MTAs for each cancer patient based on the complex individual molecular profile of their tumor. We analyzed the clinical benefit of the DDA system on the molecular and clinical outcome data of patients treated in the SHIVA01 precision oncology clinical trial with MTAs matched to individual genetic alterations or biomarkers of their tumor. We found that the DDA score assigned to MTAs was significantly higher in patients experiencing disease control than in patients with progressive disease (1523 versus 580, P = 0.037). The median PFS was also significantly longer in patients receiving MTAs with high (1000+ <) than with low (<0) DDA scores (3.95 versus 1.95 months, P = 0.044). Our results indicate that AI-based systems, like DDA, are promising new tools for oncologists to improve the clinical benefit of precision oncology.

Evaluation of a computational decision support system for molecularly targeted treatment planning by the clinical outcome data of the randomized trial SHIVA01

3642

Background: Precision oncology requires the identification of individual molecular pathomechanisms to find optimal personalized treatment strategies for every cancer patient. Incorporation of complex molecular information into routine clinical practice remains a significant challenge due to the lack of a reproducible, standardized process of clinical decision making. Methods: To provide a standardized process for molecular interpretation, we develop a precision oncology decision support system, the Realtime Oncology Molecular Treatment Calculator (MTC). MTC is a rule-based medical knowledge engine that dynamically aggregates and ranks relevant scientific and clinical evidence using currently 26,000 evidence-based associations and reproducible algorithm scoring of drivers, molecular targets to match molecular alterations to efficient therapies. To validate this novel method and system, we used data of the SHIVA01 trial of molecularly targeted therapy (Lancet Oncol 2015 16:1324-34). Molecular profiles of participants were uploaded to MTC and aggregated evidence level (AEL) values of associated targeted treatments were calculated, including those used in the SHIVA01 trial. Results: The MTC output provided a prioritized list of drugs associated with the driver alterations in the patient molecular profile, where ranking is based on AEL values. Of 113 patients who received targeted therapy with available clinical best response data, disease control was experienced in 63 cases (PR: 5, SD: 58), while disease progression occurred in 50 cases. The average AEL score for the therapies applied was significantly higher in the responsive group than in the non-responsive group (1512 and 614, respectively (p = 0.049)). In 94 cases, drugs other than those used for therapy were ranked higher by the MTC. The average AEL difference between the top-ranked and the used drugs was in an inverse correlation with clinical response, i.e. smaller differences associated with a better outcome. Conclusions: Results indicate that the aggregation of evidence-based tumor-driver-target-drug associations using standardized mathematical algorithms of this computational tool is a promising novel approach to improve clinical decisions in precision oncology. Further validation based on the results of other targeted clinical trials and real-life data using more detailed molecular profiles is warranted to explore the full clinical potential of this novel medical solution.

AI oncology algorithm and dynamic real-world learning health care system for precision oncology

35

Background: Most tumours harbor multiple driver genetic alterations and many driver alterations are linked to multiple targeted therapies with various level of evidence. In addition, a specific treatment can be linked to multiple genetic alterations in the same tumor. Several public and private databases and software solutions are available to link driver alterations to treatments options, but in clinical practice of precision oncology we need a solution to select the right treatment for our patients based on the highest level of evidence also in case of complex molecular profiles. Methods: We have developed an AI oncology algorithm and rule-engine to prioritise treatment options for every cancer patient based on the individual molecular of their tumor. This IT solution can now prioritise 1200 compounds in clinical use or clinical development based on the computing of 24,000 evidence-based associations (“rules”) between drivers, targets and compounds. The software calculates a numeric score, the “aggregated evidence level” for each driver alterations and compounds. We have linked this decision support software to a dynamic patient case management system, which records responds to therapy to create learning system to provide dynamic decision support through several lines of therapies of each patient and to use real-life evidence to further improve the algorithm. Results: Our first results indicate that system allows individualised decision of diagnostic option between single gene tests to comprehensive 600 gene NGS panels and identification of actionable alterations in 83% of cancer cases. Conclusions: This system can be a first working solution to standardise clinical decisions precision oncology, which also helps the real-life evaluation of novel multigene molecular diagnostic tests and therapies to find their best indications and accelerate their reimbursement by insurance companies and national health funds.

Growth and invasion of human melanomas in human skin grafted to immunodeficient mice

An orthotopic model of human melanoma was developed in which malignant cells were injected into human skin grafted to nude and SCID mice. Melanoma cells proliferated and invaded the human skin grafts with characteristic patterns. Three of six melanomas grew as multiple nodules and infiltered the grafts without major architectural changes in the dermis, whereas the others invaded the dermis along collagen fibers with prominent endothelial vessels. By contrast, melanoma cells inoculated into mouse skin grew as diffusely expanding nodules that did not invade the murine dermis. In human skin grafts, human melanoma cells were angiogenic for human blood vessels, and murine vessels were only found at the periphery of grafts. Tumor cells invaded the human vessels, and four out of seven cell lines metastasized to lungs, suggesting that this model is useful to determine in vivo the interactions between normal and malignant human cells.

Undifferentiated keratinocytes control growth, morphology, and antigen expression of normal melanocytes through cell-cell contact

BACKGROUND

Melanocytes in the normal human epidermis are generally dendritic and neither proliferate nor express melanoma-associated antigens. In culture, on the other hand, melanocytes are bi- to tripolar, proliferate with 2 to 4 day doubling times, and express melanoma-associated antigens. This observation prompted us to investigate the regulatory role of keratinocytes for growth, morphology, and antigen expression of melanocytes.

EXPERIMENTAL DESIGN

Melanocytes and keratinocytes were cultured under three different co-culture conditions: (a) separated by a semiporous membrane, (b) in monolayer cultures allowing direct contact between cells, and (c) in three-dimensional epidermal reconstructs.

RESULTS

Melanocytes separated from keratinocytes by semiporous membranes remained di- and tripolar and could not proliferate in medium optimal for keratinocytes. When cell-cell contact was established between melanocytes and undifferentiated, but not differentiated, keratinocytes, melanocytes proliferated at a rate similar to keratinocytes and they developed multiple dendrites. In co-cultures allowing the multi-layered growth of keratinocytes, melanocytes were nonproliferative when juxtaposed to undifferentiated keratinocytes in the basal layer, but proliferated when surrounded by differentiated keratinocytes in the intermediate and upper layers. Expression of melanoma-associated antigens on melanocytes decreased to similar levels as in normal skin when melanocytes were in direct contact with undifferentiated, but not differentiated, keratinocytes.

CONCLUSIONS

Undifferentiated, but not differentiated, keratinocytes control growth, morphology, and antigen expression of melanocytes through direct cell-cell contact. These results suggest that the phenotypic characteristics of nevus and melanoma cells in the dermis, i.e., proliferation and expression of tumor-associated antigens, may be due to their loss of contact with undifferentiation keratinocytes.

Phenotypes and interactions of human melanocytes and keratinocytes in an epidermal reconstruction model

The morphologic and antigenic phenotype of normal human melanocytes and keratinocytes was investigated in monolayer and 3-dimensional cultures in an effort to develop an epidermal model that resembles the normal human epidermis. When cultured for several passages in optimal growth medium, pure cultures of either cell type could be established as demonstrated by light and electron microscopy and with monoclonal antibodies defining melanocyte- and keratinocyte-associated antigens. Three-dimensional growth of keratinocytes on polycarbonate filters was induced by increasing calcium concentrations in the culture medium and exposing cultures to air. After 30 to 35 days incubation, the 3-dimensional keratinocyte cultures reached a total of 12 to 25 layers and keratinocytes of various stages of differentiation formed three morphologically and antigenically different strata. The basal layer of these constructs consisted of ovoid cells with desmosomes and hemidesmosome-like structures. These cells expressed low molecular weight cytokeratins similar to basal cells in situ. The intermediate layer, representing the stratum spinosum in situ, contained flat cells with keratohyaline granules and many desmosomes. These cells expressed gp 80 kilodaltons, gp 40 to 50 kilodaltons, involucrin, and filaggrin. The upper layer, the stratum corneum equivalent, contained large, flattened cells with keratohyaline granules. The majority of these cells were anucleate. When melanocytes were cocultured with keratinocytes in monolayer or in epidermal reconstructs, they assumed a multidendritic morphology and donated pigment to surrounding keratinocytes. The majority of pigmented cells localized singly within the basal layer of the reconstructs and their dendrites were intimately associated with keratinocyte plasma membranes. Pigment donation to keratinocytes appeared to occur through the uptake of melanosome-containing dendrite fragments and phagocytosis of individual melanosomes by keratinocytes. It is hypothesized that keratinocytes produce unique microenvironmental factors that regulate the melanocytic phenotype.