LFSPROShiny: An Interactive R/Shiny App for Prediction and Visualization of Cancer Risks in Families With Deleterious Germline TP53 Mutations

Abstract

PURPOSE

LFSPRO is an R library that implements risk prediction models for Li-Fraumeni syndrome (LFS), a genetic disorder characterized by deleterious germline mutations in the TP53 gene. To facilitate the use of these models in clinics, we developed LFSPROShiny, an interactive R/Shiny interface of LFSPRO that allows genetic counselors (GCs) to perform risk predictions without any programming components and further visualize the risk profiles of their patients to aid the decision-making process.

METHODS

LFSPROShiny implements two models that have been validated on multiple LFS patient cohorts: a competing risk model that predicts cancer-specific risks for the first primary and a recurrent-event model that predicts the risk of a second primary tumor. Starting with a visualization template, we keep regular contact with GCs, who ran LFSPROShiny in their counseling sessions, to collect feedback and discuss potential improvement. On receiving the family history as input, LFSPROShiny renders the family into a pedigree and displays the risk estimates of the family members in a tabular format. The software offers interactive overlaid side-by-side bar charts for visualization of the patients' cancer risks relative to the general population.

RESULTS

We walk through a detailed example to illustrate how GCs can run LFSPROShiny in clinics from data preparation to downstream analyses and interpretation of results with an emphasis on the utilities that LFSPROShiny provides to aid decision making.

CONCLUSION

Since December 2021, we have applied LFSPROShiny to over 100 families from counseling sessions at the MD Anderson Cancer Center. Our study suggests that software tools with easy-to-use interfaces are crucial for the dissemination of risk prediction models in clinical settings, hence serving as a guideline for future development of similar models.

We developed a GUI app for genetic counselors to quantify mutation/cancer risks without programming.

INTRODUCTION

Germline mutations in cancer susceptibility genes can significantly increase the likelihood of developing certain cancer types during a person's lifetime.¹ Risk prediction models built on family cancer history and Mendelian transmission to assess such increased cancer risks have been widely developed.^2-6 A prominent example is Li-Fraumeni syndrome (LFS),⁷ a rare autosomal-dominant hereditary disorder characterized by deleterious germline mutation in the TP53 tumor suppressor gene⁷ that significantly increases the risks of a spectrum of cancer diagnoses, most notably breast cancer, soft tissue sarcoma, and osteosarcoma.^7,8 Hence, when assessing cancer risks for patients with LFS, there exist additional challenges to quantify the risk of each cancer type separately. To address this issue, Shin et al⁹ proposed a Bayesian semiparametric model that estimates cancer-specific (CS) age-at-onset penetrances for the first primary cancer. Since multiple primary cancers (MPCs) are common among patients with LFS, Shin et al¹⁰ developed a complementary model that estimates penetrances beyond the first primary cancer, but does not differentiate between cancer types. We will refer to the two models as the CS and MPC models. For risk prediction purposes, we implemented them in a publicly available R package, called LFSPRO. Given unseen patients and their family history, LFSPRO can be used to predict both the carrier status of germline TP53 mutation and the risk of developing LFS-related cancers. The cancer risk predictions from LFSPRO have been validated on independent patient cohorts,^11,12 whereas its utility to predict deleterious germline TP53 mutations has been shown to outperform the Classic¹³ and Chompret^14,15 criteria in another validation study.¹⁶

CONTEXT

• Key Objective

• To expedite the clinical utility of validated risk prediction models through the implementation of a software application with simple graphical user interface.

• Knowledge Generated

• We develop LFSPROShiny, an application that can be used by genetic counselors (GCs) to predict the probability of deleterious germline TP53 mutations and cancer risks for their counselees on the basis of the collected family history. We illustrate the easy-to-use features of LFSPROShiny and also the advantages of the underlying statistical methods over clinical criteria via a concrete example. So far, we have applied LFSPROShiny to over 100 families from counseling sessions at the MD Anderson Cancer Center.

• Relevance

• The authors present a clinical decision support tool for GCs that assists in the evaluation and counselling of patients with suspected or known Li-Fraumeni/TP53 mutations. The tool provides risk assessment for developing specific first malignancies and risk of a second malignancy occurrence.

For clinicians and genetic counselors (GCs), conversations with families regarding LFS can potentially be challenging as families with LFS present a wide spectrum of cancers. With the availability of multigene panel testing, many people receive genetic testing for TP53 regardless of meeting the testing criteria. However, GCs still need to provide their best judgment of risk of carrying a TP53 mutation and on developing cancer in the future to help aid patient's understanding and decision making. There are no computerized tools GCs can use to help quantify this risk. In addition, discussions with patients on cancer risk have been based on population-based estimates and are not tailored to their specific family history. With the high variability of LFS presentations between families, a personalized risk tool can help health care providers communicate risk with patients in a more tailored way. The good predictive performance of LFSPRO supports its potential utility to address these limitations. For this purpose, it is desirable to have an interactive software package that can be executed without programming components. In this article, we present LFSPROShiny, an R/Shiny implementation of LFSPRO that provides GCs and clinicians with an easy-to-use interface, comprehensive information on patients' risk landscapes, and effective visualizations.

METHODS

Implementation and Validation of Statistical Models

The CS model⁹ was motivated by patients with LFS being at increased risks of different cancer types.^7,8 The CS model computes the CS age-at-onset penetrance for the first primary cancer. This is defined as the probability of developing a specific cancer by a certain age given the patient's covariates (eg, sex, mutation status). We considered four outcomes: (1) breast cancer, (2) sarcoma, (3) all other cancer types combined, and (4) competing mortality. We considered four competing outcomes: (1) breast cancer, (2) sarcoma, (3) all other cancer types combined, and (4) competing mortality. The MPC model¹⁰ was motivated by the frequent cases of MPCs among patients with LFS.¹⁷ This model computes the age-at-onset penetrance for MPC. This is defined as the probability of developing the next primary cancer by a certain age given the patient's covariates and previous cancer occurrences. Because of limited data availability, we estimated penetrances for the first and second primary cancers only. The penetrance estimates from both models have been successfully validated on multiple LFS patient cohorts.^11,12

LFSPRO as a Risk Prediction Tool

LFSPRO is an R package that implements risk predictions on the basis of penetrance estimates from the two statistical models described above. The main function is lfspro(), which takes the form as lfspro(fam.data, cancer.data, counselee.id, method). Here, object fam.data contains information of family members (eg, sex, TP53 mutation status, relationship), and the cancer history data set and object cancer.data record the cancer occurrences, including the cancer types and age at diagnosis, for family members. When calling lfspro(), the user specifies the model to be used (ie, method = “CS” or method = “MPC”). We will refer to these two options as LFSPRO.cs and LFSPRO.mpc. The outputs are the predicted mutation carrier probabilities and future cancer risks for family members specified by counselee.id.

LFSPRO.cs uses the CS model-based penetrance estimates to calculate the probability of developing a specific cancer type as the first primary cancer within the next $m$ years, given that the patient has not developed cancer by age $a_p$. LFSPRO.mpc instead uses the MPC model–based penetrance estimates to calculate the probability of developing a primary cancer in the next $m$ years, given the patient’s current age $a_p$ and the age at diagnosis of the previous cancer. See the Data Supplement (Section SA) for more details. Penetrance estimation is conditioned on patient-specific covariates such as sex and genotype status, for example, mutation versus wild type. For individuals with confirmed genotypes, the probabilities described above conditioned on mutation carrier status are directly used to communicate cancer risks. Most patients, however, have not undergone genetic testing. In this case, LFSPRO.cs and LFSPRO.mpc take an intermediate step and estimate the probability that a patient carries deleterious germline TP53 mutation given family history, assuming Mendelian inheritance. In the Data Supplement (Section SB), we provide the detailed computation of this probability using a hypothetical pedigree (Data Supplement, Fig S1). The utility of LFSPRO to predict mutation carriers has been externally validated on multiple patient cohorts.¹⁶ The cancer risks for untested patients are then given by weighted sums of the corresponding age-adjusted penetrances for each genotype status, with weights being the probabilities of mutation and wild type.

Development of LFSPROShiny

Programming knowledge in the R statistical software (v4.2.2; R Core Team 2022) is required to install and run LFSPRO, thus making it inconvenient in clinical settings. Furthermore, LFSPRO gives numerical output without any visualization capability. These limitations motivated our development of LFSPROShiny, a critical component in collaboration with GCs and clinicians at the MD Anderson Cancer Center (MDACC). Following agile software development practices, we used an iterative approach to build LFSPROShiny, with the first version (started in 2019) offering a basic graphical user interface (GUI) and functionality and the next version improving on the last on the basis of response from the end users. In each cycle, the development team kept regular contact with GCs, who ran LFSPROShiny right after their counseling sessions, to gather their feedback and accordingly improve the application. Although most of the discussions took place in regular meetings, GCs also reached out immediately when they encountered errors or unexpected results when running LFSPROShiny. If the input data did not run properly, the corresponding data files were shared with the development team so that issues were identified and fixed right away. Overall, most of the updates were to add functionalities and to improve visualization. For example, overlaid bar charts were added for quick comparisons between a patient’s predicted risks and the average cancer risk in the general population. These plots were further made interactive for easy download and zoom-in views of the numerical values. LFSPROShiny has been robustly functioning in clinical settings without technical issues for more than 1 year. See the Data Supplement (Section SC) for further information on the implementation of the Shiny app.

User Interface of LFSPROShiny

Following the instructions in Figure 1A, GCs supply two input data sets that are formatted as csv files (Fig 1B). The first is a family data set that provides information such as age-at-last-contact, sex, and TP53 mutation status (if known) of the family members, each of whom is uniquely identified by an identification number (ID). For each family member, the data set contains the IDs of his or her parents, which are used by LFSPROShiny to construct a pedigree plot on the basis of the inferred familial relationship (Fig 1C). The second data set describes the cancer occurrences, including the cancer types and age at diagnosis. The cancer types must be coded according to the LFSPRO cancer spectrum. These csv files are automatically generated from the pedigree data pulled from Progeny, a family medical record system at MD Anderson, in which each pedigree is uniquely identified by the medical record number of the proband. This automated process makes it straightforward and more appealing for GCs to use LFSPROShiny. Table 1 provides descriptions of the input data sets.

FIG 1.

(A) Instructions for using LFSPROShiny. (B) Control panel for uploading the input data sets, running LFSPROShiny, and customizing the output. (C) Pedigree plot describing the relationship among family members, as constructed from the input family data set. (D) Output describing the risk profile of each family member, including the probabilities of being the TP53 mutation carrier from the CS and MPC models. (E) CS risk prediction from the CS model for a patient with no cancer history. (F) Risk of a second primary cancer for a patient with one primary cancer occurrence in the past. CS, cancer-specific; ID, identification number; LFS, Li-Fraumeni syndrome; MPC, multiple primary cancer.

TABLE 1.

Column-by-Column Descriptions of the Input Family Data Set and Cancer History Data Set

Family Data Set
id	An ID that uniquely identifies a family member. The numbering should start from zero
fid	ID of the father. The value is set to NA if the father is unknown
mid	ID of the mother. The value is set to NA if the mother is unknown
test	TP53 mutation status. The value is set to 0 for noncarriers and 1 for mutation carriers. If the person has not undergone genetic testing, the column should be left empty
gender	0 for female and 1 for male
age	Age of the person at the last follow-up. The column should be left empty if the age at last contact is unknown
vital	“A” for alive and “D” for dead
Proband	“Y” if the person is the proband (ie, the person who receives genetic counseling from the GC), and “N” if not
PedigreeNotes (optional)	Handwritten notes from the GCs that may contain additional information regarding the age at last contact of the person. If provided, LFSPRO will use this column to fill in the missing age information

Cancer History Data Set
id	An ID that uniquely identifies a family member. This should be the same as id in the family data set
cancer.type	The type of cancer which must be coded according to the LFSPRO cancer spectrum: ost: osteosarcoma, sts: soft-tissue sarcoma, breast, lung, brain, leukemia, acc: adrenocortical carcinoma, and non.lfs: all other cancer types combined
diag.age	The age at cancer diagnosis

Abbreviations: GC, genetic counselor; ID, identification number; NA, not available.

The output is displayed in a tabular format (Fig 1D). The main results of LFSPROShiny are the carrier probabilities for untested patients and the cancer risk predictions (as detailed below). By default, LFSPROShiny not only incorporates the mutation status of the genetically confirmed family but also provides an option on the left panel (Fig 1B) to disable this information. The MPC model is preselected to compute the carrier probabilities, but LFSPROShiny has an option for the users to choose the CS model (Fig 1B). A patient is predicted to be a germline TP53 mutation carrier if the output probability is ≥0.2. This decision threshold achieved good trade-offs between sensitivity and specificity on multiple LFS data sets.¹⁶ If necessary, GCs can set their preferred cutoff values (Fig 1B). Prediction results from the Classic¹³ and Chompret^14,15 criteria, which are the current standards for detection of TP53 mutation carriers, are also provided. Although our multicohort validation studies indicated that the MPC model outperformed the CS model on average^16,18 (hence selected as the default option in LFSPROShiny), the CS model might be the better choice in certain cases (eg, families that are affected with many LFS-related cancers). To provide specific guidance, additional analyses are needed to accurately determine the scenarios in which one model is likely to perform better than the other.

For patients without cancer, LFSPROShiny invokes the CS model to provide separate probabilities for sarcoma, breast cancer, and all other cancer types combined, in the next 5, 10, and 15 years. For those who have had one primary cancer, LFSPROShiny invokes the MPC model to calculate the probability of a second primary. LFSPROShiny plots the patient’s predicted risks, along with a hypothetical noncarrier of the same age and sex, in an overlaid interactive bar chart (Figs 1E and 1F). Since LFS is a rare disorder,¹⁹ it is reasonable to assume that the risks of noncarriers are representative of the general population.

Deployment of LFSPROShiny

LFSPROShiny is supported across all major operating systems. The source code is available on GitHub.²⁰ We host the live app on Shinyapps.io to provide a convenient way to access LFSPROShiny.²¹ For clinical use at the MDACC, LFSPROShiny is deployed on a virtual server, which is part of a VMware cluster. The virtual server has 16 processor cores and 64 GB of memory and operates on the Red Hat Enterprise Linux 7 operating system. This internal hosting structure, as opposed to public cloud-based hosting services, ensures sufficient security controls for protected health information and allows for immediate input of patient data without access issues. Moreover, LFSPROShiny can seamlessly accept patients’ family history directly exported from the hospital’s medical record system, all within the protected MDACC intranet. Since December 2021, we have applied LFSPROShiny to over 100 families from counseling sessions at the MDACC.

RESULTS

We provide a walk-through of how GCs can run LFSPROShiny. Because of the confidentiality of patient data, we create hypothetical family and cancer history data sets. Figure 2A shows an example family data set, and Figure 2B shows the family’s cancer history. GCs may include additional columns that contain their notes regarding the patients’ age at death. These columns are based on the handwritten notes they take when additional information becomes available after the counseling sessions. LFSPROShiny automatically extracts information from the notes to complete the missing age at last contact. For example, ID 2 is estimated to have died at age 65 years.

FIG 2.

(A) Family dataset with the GCs' notes in the last three columns and (B) cancer history data set that describes the cancer occurrences among family members. GCs, genetic counselors; ID, identification number; NA, not available.

These data sets, once formatted as csv files, can be uploaded to LFSPROShiny. We provide examples to show how LFSPROShiny robustly handles incomplete data sets (Data Supplement, Figs S2-S4) and provides warnings when the family history information is deemed insufficient (Data Supplement, Figs S5 and S6). LFSPROShiny infers the relationships among the family members on the basis of their IDs and graphically displays the family as a pedigree tree (Fig 3).

FIG 3.

Pedigree tree that describes the relationship among family members. By convention, square denotes male and circle denotes female, and deceased family members are crossed with a diagonal. The proband is displayed in red. Individuals with at least one cancer occurrence (ie, affected) are color-filled.

One of the most important outputs is the probability of carrying a deleterious germline TP53 mutation for probands and their family members, especially in families with no previous genetic testing. LFSPROShiny is able to factor the proband’s personal cancer history and his or her blood relatives’ cancer history into the computation of TP53 mutation probability to determine the likelihood of having a deleterious germline TP53 mutation (as described below). In addition, in families with a TP53 mutation, although we recommend genetic testing for all close relatives, LFSPROShiny could provide additional information about their risk to test positive to complement Mendelian inheritance information. Figure 4 shows the output of the MPC model to predict TP53 mutation carriers in the example family, which has no previous testing results. The predicted mutation probability for the proband (ID 0) is close to 1.00 because of her two primary cancer diagnoses at age 15 years and age 23 years. Family members, who are closely related to the proband, are also predicted by LFSPROShiny to be mutation carriers although some of them have not developed cancer (ie, ID 11 and ID 13).

FIG 4.

TP53 mutation probabilities (ProbLFSPRO) produced by the MPC model for (A) individuals with IDs 0-11 and (B) individuals with IDs 12-21. The binary decision (LFSPRO-carrier) is determined by the default cutoff probability of 0.2. The Shiny app allows users to modify the cutoff probability. Prediction from the Chompret criteria, which is part of the current NCCN guideline, is also provided. ID, identification number; MPC, multiple primary cancer; NCCN, National Comprehensive Cancer Network.

LFSPROShiny provides interactive overlaid bar charts to visualize the cancer risks of their patients relative to the general population. For ID 9, who had osteosarcoma at age 11 years, LFSPROShiny invokes the MPC model to compute the risk of a second primary tumor (Fig 5, top left). This individual has a very high mutation probability (Fig 4), and thus, the predicted cancer risks (blue) are well above the general population (red). ID 1 has not had cancer, so LFSPROShiny invokes the CS model to compute the risks of breast cancer, sarcoma, and all other cancer types combined for the first primary cancer (Fig 5, top right). This person is predicted at near-zero probability to carry a mutation (Fig 4), and hence, the predicted cancer risks are only slightly higher than the general population. Risk prediction is not available for ID 4, who died at age 87 years (Fig 5, bottom left). Currently, LFSPROShiny does not predict CS risks beyond the first primary, and hence, risk prediction is not available for ID 0, who had osteosarcoma and breast cancers (Fig 5, bottom right).

FIG 5.

LFSPROShiny predicts the risks of a second primary tumor for individuals with one cancer (top left) and predicts cancer-specific risks (breast cancer, sarcoma, and all other cancer types combined) for those without cancer history (top right). The personalized risks are shown in blue, whereas the risks in the general population are shown in red. The predicted risks are provided for the next 5, 10, and 15 years. Risk predictions are not available for deceased individuals (bottom left) and for those who have had more than one cancer occurrence (bottom right). ID, identification number; NA, not available.

DISCUSSION

In this article, we present LFSPROShiny, a GUI tool that is built on the validated LFSPRO R package to provide a much-needed interface for medical professionals to calculate, visualize, and assess a patient's personalized mutation probability and cancer risks on the basis of their family history. In families who have not previously undergone genetic testing, LFSPROShiny can incorporate a counselee's personal cancer history and the cancer history of their relatives into the estimation of TP53 mutation probability, hence supporting an enhanced genetic counseling experience than traditional risk assessment. In families with a known mutation, LFSPROShiny can provide more personalized risk to close relatives instead of basing risk only on Mendelian inheritance. While the general guidelines are to recommend testing for all at-risk blood relatives, these complexities of LFSPROShiny allow GCs to more accurately perform risk assessment and guide patients' decision making. Accurate risk predictions help reduce the anxiety and stress for low-risk patients, while allowing high-risk patients to pursue timely genetic testing, prepare emotionally for a possible positive result, and if positive, undergo the LFS screening protocols for timely detection of cancer. Furthermore, the utility of LFSPROShiny to provide personalized cancer risks can help patients better understand their diagnosis and motivate them to follow the cancer screening regime, which remains an intensive and sometimes overwhelming process.

The current version of LFSPROShiny has several limitations. First, it does not automatically suggest a specific model, MPC or CS, to GCs for prediction of deleterious germline TP53 mutations. In the future, we wish to design a questionnaire that can lead to a suggestion for which the model may perform best for the specific patient. Although our validation study¹⁶ supports the MPC model as the default option, an in-depth evaluation of the two models in different clinical scenarios is needed to determine those in which one model is better than the other. Another limitation is that LFSPROShiny currently does not support prediction of CS risks beyond the first primary cancer. The software can be updated to address both limitations when a single model for both MPC and CS is available. Indeed, we recently developed a new model to address this challenge.²² Once this model has been sufficiently validated across multiple patient cohorts like the currently incorporated models, it will be incorporated into LFSPROShiny. LFSPROShiny currently treats all deleterious TP53 mutations identically, hence ignoring the heterogeneous effects across different variants.^23,24 TP53 mutations that confer reduced penetrance can pose a challenge since mutation carriers might not have cancer history that is highly suggestive of LFS. This is an active research topic with many recent discoveries.^25-28 When future findings on the differential outcome effect of groups of TP53 mutations are fully quantified and validated, we will incorporate them into LFSPROShiny. Another potential improvement is to extend LFSPROShiny to calculate risk in the case of adopted probands. This can be achieved by splitting the pedigree into two disjoint parts on the basis of the adopted proband and then running LFSPROShiny on the subpedigrees separately. Finally, our pipeline can be improved further to offer better user experience. Although LFSPROShiny is able to use genetic testing results for computing cancer risks, GCs currently manually enter this information into the csv files generated by Progeny. We will implement changes at the backend to automate this process. Finally, when false-positive or false-negative predictions occur, health care providers and GCs should counsel the patients on the basis of their clinical assessment of the patient's personal/family history.

AUTHOR CONTRIBUTIONS

Conception and design: Nam H. Nguyen, Elissa B. Dodd-Eaton, Gang Peng, Jessica L. Corredor, Jacynda Woodman-Ross, Wenyi Wang

Financial support: Wenyi Wang

Provision of study materials or patients: Banu K. Arun

Collection and assembly of data: Gang Peng, Banu K. Arun

Data analysis and interpretation: Nam H. Nguyen, Jessica L. Corredor, Wenwei Jiao, Banu K. Arun, Wenyi Wang

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

DATA SHARING STATEMENT

The conducted study has been approved by the institutional review board at the MD Anderson Cancer Center. The source code of LFSPROShiny and example data sets are publicly available on GitHub (https://github.com/wwylab/LFSPRO-ShinyApp). The live app is hosted on Shinyapps.io (https://namhnguyen.shinyapps.io/lfspro-shinyapp-master/). The underlying LFSPRO package can be found on GitHub (https://github.com/wwylab/LFSPRO).

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).