Introduction

Lung cancer is one of the most common cancer types and the leading cause of cancer-related deaths worldwide, with an estimated two million new cases and 1.76 million deaths per year1. Nevertheless, a trend of decreasing population-level mortality from non-small-cell lung cancer (NSCLC) was observed, probably due to treatment advances brought about by approvals and the use of targeted therapies2.

The well-established role of vascular endothelial growth factor (VEGF) in tumor angiogenesis led to the development of agents selectively targeting this pathway. Currently, three anti-angiogenic agents are FDA/EMA approved for use in advanced-stage NSCLC3. A greater understanding of disease biology and the development of targeted drugs have completely transformed the therapeutic landscape in NSCLC, spearheaded by EGFR and ALK oral tyrosine kinase inhibitors (TKIs). Molecularly targeted agents (MTAs) have been approved for genetic alterations in the oncogenes, EGFR, ALK, ROS1, BRAF, RET, NTRK1, MET, and most recently, G12C-mutant KRAS4. Besides targeted kinase inhibitors, different immune checkpoint inhibitors (ICIs) demonstrate durable responses and long-term effects on overall survival; the FDA has approved six ICIs in NSCLC by 2023 (ipilimumab, nivolumab, pembrolizumab, durvalumab, atezolizumab, and cemiplimab)5.

Precision oncology entails assigning molecularly targeted treatment for cancer, based on the individual genetic alterations of the tumor. This concept offers causal therapy by directly interfering with the biological mechanisms underlying tumorigenesis. The current paradigm mainly focuses on finding ‘actionable’ mutations or biomarkers associated with specific therapies, regardless of the full complexity of the molecular profile, resulting in underutilization of the complex genomic information provided by readily available multigene panels for next-generation sequencing in the molecular diagnostics of cancer. A recent comprehensive analysis revealed that cancer genomes typically contain 4-5 driver mutations6. Accordingly, it is increasingly evident that the simple assignment of a targeted agent to an actionable mutation results in limited benefit because of the complexity of the tumor genome (e.g.7,8). Furthermore, the interpretation of complex molecular profiles can be challenging and subjective. Thus, treatment recommendations by Molecular Tumor Boards (MTBs) are highly discordant9,10, and outcomes of MTB-recommended therapies are highly variable11.

In principle, the diversity and complexity of molecular profiles and the increasing number of targeted therapies present a data processing problem, which can be potentially addressed by computerized solutions. Indeed, their emergence in precision oncology has been anticipated12. Previously, we introduced a computational method, the digital drug assignment (DDA) system, a knowledge-based computational method to prioritize potential MTAs based on weighting and aggregating scientific evidence for the complex tumor molecular profile, rather than matching one drug to one biomarker13. As a computational reasoning model, DDA is not based on machine learning technologies, and as such, it is an ‘open-box’ explainable system. We also demonstrated that the system is predictive of clinical benefit based on the analysis of data from the SHIVA01 trial13. Importantly, patients receiving MTAs with high (1000+) DDA scores had higher benefit than those receiving MTAs with lower scores. Here, we present a study on the benefit predictivity of the DDA system by analyzing systemic lines of treatments (LOTs) received by a real-world lung cancer cohort of patients.

Results

Characteristics of the study population

To evaluate the impact of molecular diagnostics-based decision support, we followed treatments and outcomes in the everyday clinical practice of Mátraháza University and Teaching Hospital, a hospital specialized in pulmonary diseases. A study flowchart is presented in Fig. 1. Lung cancer patients enrolled in the precision oncology decision support program were included in this study without any preselection. Available samples were used for molecular diagnostic tests, tumor genomic profiles were processed by the DDA system. DDA scores were considered by the MTB that provided treatment recommendations based on the molecular findings. Treatment decisions were made by treating physicians, who considered all conditions and circumstances. Treatment histories were retrospectively collected from hospital records and analyzed.

Fig. 1: Study flowchart.
figure 1

The study involved lung cancer patients who underwent precision oncology decision support. In each case, a first Molecular Tumor Board (MTB1) reviewed the medical history of the patient, determined eligibility for molecular testing and decided the appropriate molecular tests. After completion of all tests, tumor molecular profile data were processed by the Digital Drug Assignment (DDA) system, which ranked associated molecularly targeted agents (MTAs) by DDA scores. A second Molecular Tumor Board (MTB2) reviewed the test results and the DDA scores assigned by the DDA system to MTAs by processing all alterations comprising the individual tumor molecular profile. Based on the findings, MTB2 provided a written summary including possible MTA treatment recommendations. Treating clinical oncologists then considered the MTB recommendations and all other aspects of the case to make the ultimate treatment decision. In case of relapse or treatment failure, new treatment decisions were made by clinicians (dashed arrows). Precision oncology decision support was received at different points within the journey of each patient (see also Supplementary Fig. 4). All treatment and outcome data were collected from health records by treating physicians, including treatments prior to precision oncology decision support. During data processing systemic lines of treatments (LOTs) were retained and used for analysis. LOTs were stratified according to standard chemotherapies (SC) and MTAs, which were also analyzed by score stratification (high-score (1000≦DDA score) and low-score (DDA score<1000) MTA lines). Pooled LOT data were used for response and survival analyses according to stratification.

Full size image

Patient characteristics are summarized in Table 1. Statistical details of applied molecular diagnostic test types and detected alteration categories are presented in Table 2. Molecular diagnostics findings are summarized in Supplementary Fig. 1. The most commonly mutated genes were TP53, KDR, KRAS, MUC16, MET, PIK3CA, EGFR, PIK3R1, KIT, CDKN2A. Precision oncology decision support was received prior to the first, second, and later treatment lines in 9%, 31%, and 39% of patients, respectively; 25.2% received only (neo)adjuvant or no systemic therapies. 20.7% of patients received single-gene or small-panel molecular tests (mostly KRAS, EGFR, ALK, PD-L1) before comprehensive molecular profiling and personalized DDA-based decision support.

Table 1 Patient demographics
Full size table
Table 2 Applied molecular diagnostic test types and detected alterations categories
Full size table

Clinical treatments

The administered treatment lines by compounds are available in Supplementary Table 1. Distribution of LOTs before/after precision oncology decision support is presented in Supplementary Fig. 2. The median duration of follow-up was 42.0 months (IQR: 14.0-120.0). The median OS (mOS) of the full patient population (n = 111) was 46 months (95% CI: 24.0-95.0), mOS of those who received at least one LOT (n = 83) was 32 months (95% CI: 20.0-67.0). Further OS data by subgroups is presented in Supplementary Fig. 3.

Clinical outcomes of MTA and SC treatments

The real-world nature of the study enabled comparison of MTA (including immune checkpoint inhibitors) and SC therapy outcomes. To this end, PFS data of individual treatment lines was compiled and analyzed. Distribution and duration of LOTs is presented in Supplementary Fig. 4. Median PFS (mPFS) of SC lines (n = 56) was 7 months (95% CI: 4.0-9.0) versus 11 months of MTA lines (n = 81) (95% CI: 8.0-12.0) (HR: 0.46, 95% CI: 0.31-0.67) (p < 0.001) (Fig. 2a). In lines prior to personalized decision support, mPFS of SC lines (n = 41) was 7 months (95% CI: 4.0-9.0) versus 9 months of MTA lines (n = 39) (95% CI: 7.0-11.0) (HR: 0.54, 95% CI: 0.33-0.89) (p = 0.014) (Fig. 2b). In lines following personalized decision support, mPFS of SC lines (n = 15) was 4 months (95% CI: 3.0-12.0) versus 11 months of MTA lines (n = 42) (95% CI: 9.0-17.0) (HR: 0.44, 95% CI: 0.23-0.86) (p = 0.015) (Fig. 2c). ORR and DCR data of MTA and SC lines are presented in Table 3. In terms of ORR, MTAs assigned after decision support had a clear clinical benefit over SC treatments, while a high overall DCR could be achieved with all treatment modalities used. When patient OS was analyzed by treatment types, mOS of SC treatments (patients who did not receive any line of MTA, (n = 24)) was 15 months (95% CI: 5.0-20.0) versus 67 months of MTA treatments (patients who received at least one line of MTA) (n = 59) (95% CI: 36.0-not reached) (HR: 0.17, 95% CI: 0.09-0.34) (p < 0.001) (Fig. 3a). When stratified on the basis of MTAs assigned after decision support, mOS of SC treatments (patients who did not receive MTA after decision support) (n = 39) was 16 months (95% CI: 13.0-21.0) versus 95 months (95% CI: 45.0-not reached) (HR: 0.2, 95% CI: 0.1-0.4) (p < 0.001) of MTA treatments (i.e., at least one line of MTA after decision support) (n = 44) (Fig. 3b). Taken together, these data indicate a superior clinical benefit with MTA treatments in comparison to SC by various outcome measures.

Fig. 2: Progression-free survival of MTA and SC treatment lines.
figure 2

Treatment lines were stratified by administration of molecularly targeted agents (MTAs, blue) and standard chemotherapies (SC, red). Kaplan-Meier analysis by treatment types is shown for all lines of therapy (a), lines administered before precision oncology decision support (DS) (b), and lines administered after precision oncology decision support (c).

Full size image
Fig. 3: Overall survival by treatment types received.
figure 3

Patients were stratified by either receiving at least one line of treatment with a molecularly targeted agent (MTA, blue) or receiving exclusively standard chemotherapy treatments (SC only, red) during their treatment history. Kaplan-Meier analysis by treatment types based on all administered treatment lines is shown in (a). Patients who received treatments following precision oncology decision support (DS) were also stratified by MTA and SC treatment lines after DS (b).

Full size image
Table 3 Overall Response Rates (ORR) and Disease Control Rates (DCR) of MTA and SC treatment lines
Full size table

It is observed generally in the clinic that latter-line therapies are less effective than preceding lines. When MTA lines prior to and following decision support (mainly first versus 2+ lines) were compared, a statistically non-significant trend in PFS benefit was revealed: 9 (95% CI: 7.0-11.0) versus 11 (95% CI: 9.0-17.0) months mPFS prior (n = 39) and following (n = 42), respectively (HR: 0.77, 95% CI: 0.48-1.25) (p = 0.287) (Fig. 4). PFS ratio (PFSr) of MTA lines following decision support (PFS2) and preceding lines (PFS1) was higher than 1.3 in 43.48% of cases (n = 10/23). Modified PFSr (mPFSr) calculated according to Mock et al.14, was higher than 1.3 in 52.17% of cases (n = 12/23). Similarly, as shown in Table 3, there was no statistically significant difference between ORR and DCR of MTAs administered before and after decision support. In contrast, as expected for latter lines, both measures dropped significantly in SC lines following decision support. On par efficacy of MTA treatment lines administered after DDA-based decision support with preceding lines underscores an overall benefit of personalized approach in treatment assignment.

Fig. 4: Progression-free survival of MTA treatments administered before and after decision support.
figure 4

Treatments were stratified by administration of molecularly targeted agents (MTAs) before (red) and after (blue) precision oncology decision support (DS).

Full size image

Clinical outcome analyses by treatment DDA scores

Reanalysis of the SHIVA01 trial data previously indicated increased clinical benefit when patients were treated with drugs of higher drug scores (1000≦) assigned by DDA13. Proceeding by the same concept, to reveal increased benefit provided by high-score (1000≦DDA score) MTA lines, here, we first compared MTA lines divided by the DDA score threshold of 1000 and found a significant benefit in PFS by high-score MTAs. mPFS of high-score MTA treatments (n = 19) was 32 months (95% CI: 12.0-63.0). In contrast, mPFS of other treatments (i.e., MTA lines of DDA score<1000 and SC lines) (n = 118) was 8 months (95% CI 6.0-9.0) (HR: 0.28, 95% CI: 0.15-0.52) (p < 0.001) (Fig. 5a), and mPFS of low-score MTA lines (DDA score<1000) (n = 57) was 9 months (95% CI: 7.0-11.0) (HR: 0.36, 95% CI: 0.19-0.68, p = 0.001) (Fig. 5b). In LOTs following decision support, mPFS of high-score MTA treatments (n = 13) was 63 months (95% CI: 12.0-66.0), whereas it was 9 months both for other treatments (n = 44; 95% CI: 5.0–11.0) (HR: 0.15, 95% CI: 0.06-0.4) (p < 0.001) (Fig. 5c) and for low-score MTAs (n = 27; 95% CI: 5.0–12.0) (HR: 0.17, 95% CI: 0.06-0.48) (p < 0.001) (Fig. 5d).

Fig. 5: Clinical outcome analyses by treatment DDA scores.
figure 5

Treatments were stratified by DDA score of 1000, as high-score treatment lines (1000≦DDA score) were demonstrated to result in increased clinical benefit based on analysis of data from the SHIVA01 trial13. High-score treatments are indicated with blue, the corresponding comparator sets in red in all Kaplan-Meier plots throughout (a) to (f). Progression-free survival (PFS) probability of high-score treatments was plotted against all other treatments (SC and low-score MTAs) (a) and low-score MTAs only (b). PFS of high-score treatments versus all other treatments (c) and versus low-score MTAs (d) administered following decision support (DS). Overall survival (OS) of patients who received at least one high-score MTA line of treatment (LOT) during their treatment course versus OS of patients who did not receive any high-score MTA LOT during their treatment course (e) and OS of patients who received at least one high-score MTA LOT following DS versus OS of patients who did not receive any high-score MTA LOT following DS (f).

Full size image

The specific benefit by high-score MTAs also manifested significantly in OS. The median OS of patients who received at least one line of high-score MTA during their treatment journey (n = 15) was not reached (95% CI: 45.0-not reached), versus 24 months mOS (95% CI: 18.0-36.0) of those who only received non-high score treatments (patients only treated with low-score MTAs and/or SC, n = 68) (HR: 0.14, 95% CI: 0.03-0.58) (p = 0.002) (Fig. 5e). Median OS of patients who received at least one line of high-score MTA after decision support (n = 12) was 95 months (95% CI: 17.0-not reached) versus 32 months mOS (95% CI: 24.0-67.0) of those who only received non-high score treatments after decision support (patients only treated with low-score MTAs and/or SC) (n = 34) (HR: 0.19, 95% CI: 0.04-0.86) (p = 0.018) (Fig. 5f). The five-year overall survival rate of the 15 patients treated with at least one line of high-score MTA was 33%, and 6% for patients only receiving low-score MTA and/or SC LOTs (n = 68) (p = 0.0082).

ORR and DCR data of high-score MTA treatment lines and other treatment lines are presented in Table 4. High-score MTA treatments had a pronounced clinical benefit over other treatments in terms of ORR and while a high overall DCR was characteristic to all treatment lines, high-score MTA treatments assigned after decision support also had a benefit in DCR.

Table 4 Overall Response Rates (ORR) and Disease Control Rates (DCR) of high-score MTA treatments (1000 ≦ DDA score), all other treatment lines (MTAs with DDA score < 1000 and SC), and low-score MTA treatments (DDA score < 1000)
Full size table

Multivariate analysis was performed to assess associations with other prognostic factors (Fig. 6). Gender and age did not significatively impact outcomes. Molecular biomarkers in general did not significatively impact outcomes, but mutations that are biomarkers of FDA-approved TKIs were favorable predictors in terms of both OS and PFS. PD-L1 status did not have statistically significant predictivity. Adenocarcinoma histology was a favorable predictor, whereas small-cell carcinoma was unfavorable in terms of OS but was not significantly different in terms of PFS.

Fig. 6: Multivariate analysis forest plots.
figure 6

Multivariate analysis forest plots relative to PFS (a) and OS (b). TKI biomarker pos.: patients with driver oncogenes serving as biomarkers for FDA-approved targeted tyrosine kinase inhibitor (TKI) therapy for lung cancer; PD-L1 (overexp.): patients with molecular diagnostic results reporting positive PD-L1 expression status; any biomarker pos.: patients with any positive biomarkers for TKI or immunotherapy. SCC: squamous cell carcinoma; SCLC: small-cell lung cancer.

Full size image

All in all, 19 high-score MTA LOTs were administered to 15 patients (Supplementary Fig. 4). As high-score MTA treatments provided significant clinical benefit, to reveal the potential utility of DDA-based drug assignment, we also examined the DDA outputs for the full patient population (Supplementary Fig. 5). At least one high-score MTA approved in lung was assigned to 50% (n = 56) of patients. Within patients who received at least one line of systemic therapy (n = 83), 48% (n = 40) were assigned with lung-approved high-score MTA.

Discussion

This retrospective study aimed to assess precision oncology decision support in lung cancer in a real-world setting, with special emphasis on the predictivity validation of the previously introduced and clinically validated Digital Drug Assignment system13. To this end, clinical data of lung cancer patients in a hospital specialized for pulmonary diseases, who received tumor molecular testing and precision oncology decision support were collected. Several targeted and immunotherapies are approved in lung cancer, thus this tumor type represents a relevant use case for digitally supported precision oncology.

Tumor genomic profiling revealed mutations in genes that are typically associated with lung cancer, e.g., TP53, KRAS, MET, EGFR (Supplementary Fig. 1). Larger tests evidently detected a larger number of alterations, most importantly they detected an average of 4.5 alterations categorized as driver by the DDA system (Table 2), in line with published average driver counts6. The actual complexity of tumor genomes thus appears sufficiently well represented by a 600-gene NGS test in the clinical context. In addition, many additional variants were detected, most of these were classified as variant of unknown significance (VUS) by the DDA system. Overall, the unique complexity of tumor genomic profiles underscores the importance of personalized approach in treatment assignment but also highlights issues with interpretation. Molecular diagnostics provides a snapshot of the tumor evolutionary process to guide therapy decisions. DDA can facilitate this process, but as tumors dynamically evolve under changing selective pressures, it is reasonable to perform further genomic testing (including liquid biopsy) to guide more accurate decisions for subsequent lines of treatments. Nevertheless, it should be noted that various resistance mechanisms remain hidden with current clinical diagnostic methods15.

MTA treatments demonstrated a general benefit over SC (Figs. 2 and 3), supporting their clinical role in lung cancer. In terms of selecting effective MTAs, the utility of precision oncology decision support is reflected by multiple clinical outcome measures. There was an increase in the use of MTAs in lines following decision support. Efficacy of these subsequent personalized treatment lines is demonstrated by the high PFSr: 43.48% higher than 1.3 of PFS2/PFS1; and 52.17% modified PFSr, calculated according to Mock et al.14. Although the total number of cases where PFSr for subsequent MTA lines could be calculated was relatively modest (n = 23), PFSr was higher than reported for various clinical trials14,16. Similarly, a numerically higher mPFS, 9 versus 11 months, was achieved by MTAs administered before and after decision support, respectively (Fig. 4). In line with these findings, while ORR of SC therapies administered after decision support was strikingly lower than of those prior to (0% versus 37%, respectively, p = 0.001; Table 3), there was even a numerical increase in ORR of MTA therapies administered after decision support versus of those before (37% versus 30%, respectively, p = 0.514). A similar trend was observed in DCR data. Taken together, having at least equivalent clinical benefits (mPFS, ORR, DCR) in later lines, where a significant decline in benefit with later SC lines is well demonstrated, strongly supports the role of multigene testing and appropriate decision support for MTA treatments.

Beyond the general trends of MTA benefits, we also aimed to analyze the specific benefit of MTAs assigned with high scores by the DDA system. DDA drug scores were considered by the MTB for treatment recommendations to treating clinicians. The final therapy decision was made by the clinicians, who, beyond molecular profiles, also had to consider other clinical and accessibility aspects (e.g., guidelines, comorbidities, drug interactions, toxicities, reimbursement, etc.). As not all patients received high-DDA-score MTAs, we could retrospectively analyze the effectiveness of high-DDA-score MTAs in comparison to other therapies (i.e. low-score MTAs and SC). Previously, we demonstrated better clinical outcomes within participants of the SHIVA01 trial, who received MTAs with DDA scores higher than 100013. In line with this, here we also found that high-score (1000+) treatment lines resulted in significantly longer PFS than other lines (Fig. 5a–d). Similarly, ORR of such lines was also significantly higher than other lines (Table 4). Moreover, mOS of patients who received at least one line of MTA with 1000≦DDA score was also significantly higher than of patients not receiving such therapies (Fig. 5e, f), along with a significantly higher five-year survival rate. These results strongly support the notion that DDA score is predictive of clinical benefit. The increased five-year survival rate with high-score therapies may indicate that computational precision oncology can overcome the issues with resistance related to the complexity of cancer genomes. Furthermore, 50% of the patient population were assigned with high-score MTAs approved in lung cancer, highlighting a meaningful potential utility of computational processing of complex tumor genomic data for efficient drug assignment.

Lung cancer is already one of the key areas for precision oncology, with multiple approved targeted therapies available, and the repertoire of targeted drugs keeps increasing. For example, the KRAS p.G12C inhibitors, sotorasib and adagrasib have been recently approved in lung cancer17. Due to their experimental status during this study, none of the patients received KRAS inhibitor therapies, although 21 tumors (19%) harbored KRAS p.G12C mutation, highlighting further potential for personalized therapies. MRTX1133, another specific KRAS inhibitor, effective against the p.G12D mutation, is currently in clinical trial (NCT05737706). These examples illustrate a trend of ever-increasing complexities in personalized oncology, further increasing the burden of identifying the right drug for each individual molecular profile, while interpretation of complex molecular profiles can already be challenging in clinical practice.

The need for appropriate genomic interpretation has been addressed by the establishment of MTBs, thus incorporating molecular biologists and geneticists into the clinical decision-making process11. This entails labor-intensive manual research, which heavily relies on databases, presenting heterogeneous data18. Accordingly, concordance between treatment recommendations by different MTBs on the same molecular profiles is low9,10. Digital tools are obvious solutions to data processing problems, and their emergence in personalized oncology has been already anticipated12. Advances in large language models (LLMs) have gained much attention, and a recent study examined their utility in precision oncology decision support on fictional cases of patients with advanced cancer with genetic alterations19. However, the results indicate that LLMs are not yet applicable as a routine tool in clinical decision support. Machine learning (ML) tools that use various clinical data types, which may include some kind of genetic or omics data to predict survival and/or efficacy of specific drugs (e.g. ICI or chemotherapies) in lung cancer have been presented, although none of these is designed to process tumor molecular profiles or to rank targeted therapies associated with the individual tumor genomic data20,21,22,23,24,25,26. In general, training and testing deep-learning-based artificial intelligence (AI) in the domain of precision oncology is hindered by limitations of appropriate training data, training expertise, experimentation, and simulation27. Moreover, the black-box problem of ML systems is a cause for caution, explainability in medicine has a special relevance28. DDA is an ‘open-box’, explainable knowledge-based reasoning system13, and our results demonstrate that digital solutions can be incorporated into day-to-day practice of an MTB and can potentially improve outcomes. Analysis of Medicare fee-for-service data revealed substantial underuse of molecular testing and targeted therapies in NSCLC, with variation by practice type and patient socioeconomic characteristics29. Increased digitalization of molecular data processing can contribute to standardization and improve equity in access to precision oncology.

The retrospective and real-world nature of this study poses limitations in terms of data quality derived from unstructured health records. Adherence to oral medications could not be fully followed in this setting. Accessibility (reimbursement) issues also influenced clinicians’ decisions. In case of combination therapies, the actual contribution of individual compounds and possible synergistic or additive effects could not be evaluated. Despite its limitations, our results are relevant to personalization of precision oncology, where molecular diagnostics is becoming standard, especially in lung cancer. This study provides evidence that a digital tool can aid addressing the various challenges brought about by molecular-data-based drug-assignment approach in the clinical setting. Thus, the results warrant further prospective, randomized investigation of the utility of the DDA system in the clinical settings focused on improving the benefits of precision oncology.

Methods

Study design and participants

A study flowchart is presented in Fig. 1. Between 2018 and 2022, 111 lung cancer patients from Mátraháza University and Teaching Hospital (Mátraháza, Hungary) were involved in the precision oncology program provided by Oncompass Medicine (Budapest, Hungary). An MTB consisting of medical oncologists, molecular biologists, and a genetic counselor was responsible for evaluating patients for molecular diagnostics and precision oncology decision support. Eligibility criteria included histologically proven lung cancer, available tumor specimen with proof of malignancy, as well as having satisfactory performance status according to the Eastern Cooperative Oncology Group (ECOG) criteria at molecular diagnostics.

All patients provided written informed consent to use anonymized data for research purposes. Prior to conducting the study, ethics approval was obtained from the National Institute of Pharmacy and Nutrition (approval no. OGYEI/50268/2017), in accordance with the principles of the Declaration of Helsinki.

Sample preparation and molecular profiling

Samples were collected from the primary tumor or metastatic samples or liquid biopsy. The available samples were reviewed by the MTB, which ranked them for molecular testing and determined the need for additional tests. Tumor samples were analyzed by targeted next-generation sequencing (NGS) panels, immunohistochemistry (IHC), and fluorescent in situ hybridization (FISH). Results from NGS tests with 591-gene panels included copy number variations (CNVs) and tumor mutational burden (TMB). 50-gene (Cancer HotSpot Ampliseq Panel, Thermo Fisher) and 591-gene (Eurofins Genomics) panels were used for targeted sequencing; the latter panel also included 22 gene fusions and four promoter regions of key cancer-specific genes, as well as 24 miRNA genes. IHC test was applied to determine the expression of the programmed cell death 1 ligand (PD-L1) protein (22C3 pharmDx, DAKO, M365329-2). To ascertain the rearrangements of ALK and ROS1 genes, and to examine the amplification of EGFR, FGFR1, HER2, MET, and PIK3CA, FISH tests were carried out with ZytoLight Direct Label system (ZytoVision). Individual tumor molecular profiles are presented in Supplementary Data 1.

Bioinformatics

Bioinformatic filtering was conducted using the QCI (QIAGEN Clinical Insight Interpret 8.0.2021 0827) software with the following parameters: call quality of at least 30.0, and pass upstream pipeline filtering, and read depth of at least 10.0, and allele fraction of at least 1.0 with genotype quality of at least 30.0. Variants that exhibited an allele frequency of 10.0%≤ of the genomes in the 1000 genomes project, or at least 10.0% of the ExAC Frequency or had a frequency of 10.0%≤ in gnomAD were excluded. Based on ACMG guidelines classification, benign and likely benign alterations were also ruled out. Finally, only exonic and splice-site mutations were kept.

Digital drug assignment and precision oncology decision support

As previously described, the Digital Drug Assignment (DDA) system is an algorithmic computational reasoning model that ranks associated targeted therapies based on the totality of individual tumor genomic data13. Briefly, this is a rule-based knowledge engine, which is able to classify genetic alterations and prioritize target genes and agents using a mathematical algorithm that connects the unique molecular profile of the tumor to related scientific evidence from the database. The system is based on the principle of simultaneously considering pieces of evidence connecting alterations, targets, and drugs. These evidence relations can be either positive (e.g., representing drug efficacy) or negative (resistance) and of various levels of strength. The structured evidence relations comprise a computable network with unique nodes and edges for each individual molecular profile. The algorithm performs a weighted aggregation of these evidence relations, which can eventually be expressed by a mathematical score, the aggregated evidence level (AEL) score13. For clarity, ‘AEL score’ is referred to as ‘DDA score’ throughout this article. The aggregated processing of the available scientific evidence simulates expert logic and makes the method objective and reproducible.

The software versions of the DDA system used in this study were the Realtime Oncology Treatment Calculator v1.47-1.71 (Genomate Health, Cambridge, MA, USA). Individual tumor molecular profiles were processed by the DDA system and DDA scores assigned to MTAs were considered by the MTB for treatment recommendations in the precision oncology report. Individual DDA drug scores are presented in Supplementary Data 2. Treatment decisions were ultimately made by the treating physicians, outcome analyses were performed post hoc, following treatments and data collection.

Statistical analysis

Clinical data for outcome analysis were collected from health records by treating physicians in 2023. Overall survival (OS) data were collected for all 111 patients. In total, 155 systemic lines of treatments (LOT) were assigned, their distribution is presented in Supplementary Fig. 4. Each individual LOT was treated as an entity in all PFS and response analyses, as they provide direct information about drug efficacy and thus are suitable to statistically analyze the relationship between clinical benefit and DDA score tiers. Combined treatment LOTs were classified as MTA if comprising of at least one molecularly targeted agent. In case of combined MTA treatments, the higher DDA score was considered for stratification. Patient-level treatment and outcome data are presented in Supplementary Data 3. PFS and OS data were calculated by Kaplan-Meier (KM) estimation, log-rank test was employed to compare the survival endpoints. Hazard Ratio (HR) and corresponding 95% lower and upper confidence interval values were calculated by Cox proportional hazards regression model. Median follow-up time was calculated by the reverse KM estimator method. Level of significance in differences of clinical outcome data such as disease control rate (DCR) and overall (objective) response rate (ORR) was determined by using Fisher’s exact test. Statistical analyses were performed using the NumPy, SciPy, and lifelines modules of Python 3.9. KM plot figures were generated with the lifelines Python module; p values refer to log-rank test.