Generative AI - Assisted Adaptive Cancer Therapy

Abstract

Adaptive combination therapy is deemed the most intuitive strategy to thwart therapeutic resistance through dynamic treatment tuning that accounts for cancer evolutionary dynamics. However, higher accuracy and reliability of treatment response predictions would be needed, in addition to the need for clinically feasible models of adaptive combination therapy that consider newly approved therapeutics and the growing multimodal data being available about cancer. Grounded in nonlinear system control theory, this review offers a perspective on exploiting GenAI learning and inferencing capabilities to predict treatment response and recommend treatments in the context of adaptive cancer therapy. Results from nonlinear system identification, control theory and deep learning are integrated within an adaptive cancer control framework to leverage the continuously expanding data about cancer and its treatment towards GenAI-enhanced adaptive therapy. The resulting models and their analysis contribute to a much-needed conceptual clarity about the research and translational pathways that would be needed to realize GenAI-assisted cancer treatments. In particular, they underscore that access to clinical data, deep learning opacity, and clinical validation present critical challenges that require adequate attention to pave the way towards acceptance and integration of GenAI in real-world oncology workflows.

Plain Language Summary

Despite the increasing number of treatment options being available to cancer patients, most of these treatments stop working after some time due to the cancer becoming resistant. One of the strategies to thwart cancer resistance is to design treatments that are one-step-ahead of cancer by using drug combinations and adapting therapies based on treatment response monitoring. This treatment approach is called adaptive therapy. This study explores the use of generative artificial intelligence (GenAI) models to enhance adaptive therapy by predicting changes of the disease and accordingly adapting treatments based on learning acquired by these models through training on large amount of data about cancer and patient treatments. The study highlights a pathway for the clinical evaluation of this GenAI-assisted approach to cancer therapy and discusses the challenges associated with its implementation and deployment in clinics.

Introduction

Cancer is complex adaptive system^1-4 with nonlinear time-varying dynamics shaped by genetic^5-7 immunological^8-12 and eco-evolutionary determinants.^13-16 These causal dimensions of cancer complexity are bound to genetic diversity and phenotypic plasticity which support a near-boundless capacity of cancer to adapt to therapeutic interventions and ultimately lead to resistance as the most pressing challenge in the treatment of cancer. Combination therapy, which involves combining different drugs and therapies, is a promising strategy for preventing the onset of therapeutic resistance^17-21 and is the object of thousands of ongoing or planned clinical trials. Recent notable additions to cancer drugs used in combination therapy include osimertnib, a third generation EGFR inhibitor, which yields a significant increase in disease free survival (DFS) when used in the adjuvant setting for patients with EGFR-mutated, early stage NSCLC (Non-Small Cell Lung Cancer). ²² Another significant advance in the treatment of cancer is the immunotherapy combination of ipilimumab and nivolumab used for metastatic NSCLC, leading to a significant improvement in overall survival (OS) compared to chemotherapy, irrespective of the expression status of PD-L1. ²³ The advent of KRAS inhibitors for the treatment of KRAS-mutated solid tumors^24-26 and the FDA (Food and Drug Administration) approval of sotorasib and adagrasib for patients with NSCLC harboring G12 C KRAS mutation offers other possible avenues for combination therapies involving immunotherapy and chemotherapy. ²⁷ Equally noteworthy is the recent FDA approval of trastuzumab deruxtecan (T-Dxd) for any solid cancer that is HER-2 positive ²⁸ and alectinib for ALK-positive NSCLC ²⁹ which have been explored in ongoing combination therapy studies.^30,31

Inter-tumor and intra-tumor heterogeneity, the multiplicity of cancer cells’ resistance pathways, and potential multi-drug synergy represent a strong rationale for the use of combination therapy^20,32 However, the large number of possible drug combinations exceeds by far the number of patients that could feasibly be recruited for their clinical studies and validations. Clearly, a brute force approach of considering all possible drug combinations is neither financially nor clinically feasible, making it imperative to pursue a more rational design of combination therapy. ²⁰ For instance, targeting synthetic lethality and focusing on multiple segments of signaling pathways are deemed rational based on their mechanisms of actions. The effectiveness of combination therapies will ultimately depend on their ability to account for the inevitable adaptation of cancer cells and their underlying eco-evolutionary dynamics. Through an explicit consideration of cancer evolutionary dynamics, adaptive therapy has shown promising clinical results and is increasingly viewed as the most apt treatment strategy to thwart therapeutic resistance.^33-39 Furthermore, the emerging view of cancer as a complex adaptive system^1-4,40 is a compelling rationale for framing cancer treatment as an adaptive control problem where therapeutic actions should be patient-tailored and adaptively tuned based on repeated monitoring of treatment response.^40-42

Mathematical modelling has been used to recast knowledge, insights and assumptions about cancer dynamics into mathematical models that provide a basis for the synthesis and optimization of dynamic treatment schedules.^41,43-47 The framing of cancer treatment as an adaptive control problem ⁴¹ involves three fundamental components,^48-50 namely: (1) Treatment response monitoring and disease state estimation/prediction to serve as feedback for adaptive control, (2) Planning the desired disease state trajectory, ie, the setpoints for adaptive control, and (3) Synthesis and application of controls that reflect the adaptive treatment scheme. The realization of these components and their integration into clinically feasible adaptive therapy will require advances in treatment response monitoring technologies such as liquid biopsy, disease state estimation algorithms, new therapeutic agents that can be combined with low toxicity, and innovations in clinical workflow engineering to support dynamic treatment schedules. Beyond these baseline needs, adaptive therapy depends on mathematical and computational models of cancer dynamics that are sufficiently accurate and reliable in capturing the salient elements of cancer dynamics but not too complex so as to render their analysis beyond reach. These models need not only reflect the state-of-the-art knowledge about cancer biology and related insights and perspectives but must also continuously learn from the accumulated radiomics, genomic, metabolomic, and clinical data about cancer and its treatments. Furthermore, real-world challenges such as dose-limiting toxicities and the management of immune-related adverse events (irAEs) for immune checkpoint inhibitors impose intrinsic constraints on the set of feasible adaptive control schemes that can be safely used for dosing and therapy combinations. For instance, osimertnib has been an integral part of SOC for NSCLC. However, resistance is inevitable and may be caused by different mechanisms some of which are known, such as MET amplification, histologic transformation, gene fusions (e.g., RET, BRAF) and tertiary EGFR mutation.^51,52 This would generally necessitate combination therapy and modulation thereof to prevent or delay on-target and off-target resistance, guided by pertinent factors such as tolerable toxicity, as well as clinical and molecular profiles. For example, the combination of osimertnib and gefitnib is studied for newly diagnosed EGFR-mutated NSCLC patients (NCT03122717 ) to evaluate both concurrent and 4-week alternating administration of these two agents to prevent the selection of C797X and T790 M EGFR clones. ⁵³ While the rationale behind this scheme may be tied to insight into observed patterns of emerging resistance to such combination therapy, it is not clear how this strategy would blunt the emergence of unforeseen on-target and off-target resistance to these tyrosine kinase inhibitors (TKIs) and combinations thereof. Herein lies the potential for data-driven generative artificial intelligence (GenAI) to support the paradigm of adaptive therapy in the aspects of treatment response predictions and treatment planning and adaptations. In particular, GenAI repeated predictions of clonal frequencies would provide the required feedback for schemes of therapy adaptation and dosing to delay or prevent the emergence of resistant clones. Furthermore, trained with the increasingly expanding big data on cancer and its treatment, GenAI could also provide patient-specific toxicity predictions to inform the consideration of toxicity constraints in the conception of therapy adaptation algorithms. For example, PD-L1 blockade followed by osimertnib is known to cause irAEs, while osimertnib followed by PD-L1 blockade or PD-L1 blockade followed by erlotinib does not cause severe irAEs in the treatment of EGFR-mutated NSCLC. ⁵⁴ Furthermore, the time interval separating PD-L1 blockade and EGFR-TKIs initiations and the combination scheme affects the frequency of irAEs. ⁵⁴ Informed by GenAI toxicity and clonal frequencies’ predictions, adaptive therapy would be fully equipped to modulate, in a patient-specific way, the timing of therapy initiation, the combination scheme and the dosing to thwart therapeutic resistance and prevent irAEs. However, to realize this potential, there is a myriad of real-world clinical setting factors related to the patient and the care delivery environment that would need to be adequately addressed as will be discussed in later sections.

The article shares a GenAI-driven perspective on cancer treatment that is anchored in adaptive control and nonlinear system theories. As such, it is intended to contribute to the maturity of mathematically-driven dynamic treatment approaches, which are increasingly viewed as essential in dealing with the evolving nature of cancer to improve patient outcomes.

Disease State Estimation

Cancer is a complex adaptive system whose effective control requires treatment adaptation based on repeated monitoring of treatment response. The monitoring enables an estimation of cancer evolutionary state dynamics based on appropriately chosen biomarkers of disease progression and treatment response. In order to reflect the reality of cancer evolutionary dynamics^13-16 and its tumoral and clonal heterogeneity^55,56 as determinants of therapeutic response and resistance, tumor clonal frequencies at timestep $i$ are used to defined the patient’s disease state $s_i$ . Treatment response is monitored through repeated sampling of biomarkers validated for the cancer type in question. Considering prostate cancer as an example, disease state may be estimated based on RECIST (Response Evaluation Criteria in Solid Tumors) assessments, Androgen Receptor (AR) amplification, the abundance of ctDNA (circulating tumor DNA), the alteration status of RB1, P53, and PTEN, and the state of DDR ( DNA damage repair) genes such as ATM, CDK12 and BRCA1/2 , which may be monitored using liquid biopsy (LB).^57-61 Disease state estimation may also use immunological biomarkers including the levels of cytokines such as IL-8, known to be implicated in the infiltration of MDSCs (Myeloid Derived Suppressor Cells). ⁶² The repeated monitoring of treatment response provides an objective basis for the prediction of cancer’s nonlinear dynamics. Such predictions have been extensively studied for biological, physical, economic, and social systems and are intimately linked to the broader topics of nonlinear system analysis, identification and reconstruction.^63-76 Although predictions of nonlinear dynamics are limited by their sensitive dependence on initial conditions and the real-world impossibility of measuring initial conditions with an arbitrary precision,^77-80 such dynamics may be reconstructed from time series observations using a delay-coordinate embedding as asserted by Takens’s embedding Theorem ⁸¹ and subsequent studies.^72,82-90 Grounded in these mathematical foundations, multiple methods have been developed for data driven identification of nonlinear systems. ⁷⁰ Machine and deep learning approaches have also been used to reconstruct nonlinear dynamics from data,^67-69,91-96 including the use of deep neural networks (DNN) to recast nonlinear systems within a linear framework ⁹¹ and the use of autoencoder networks ⁹⁷ to reconstruct strange attractors. ⁹⁵ These works and the mathematical foundations of reconstructing nonlinear dynamics from data are aligned with the property of neural networks as universal approximators^98-101 which has been argued to support the proposition that GPTs (Generative Pre-trained Transformers) can serve as predictors of nonlinear treatment response. ¹⁰² Many deep-learning models such as LSTM (Long Short-Term Memory) networks,^103-105 autoencoders ⁹⁷ and transformers^106-108 may be used to predict disease states. However, due to the success of transformers in a wide range of applications ¹⁰⁹ the current study builds on a recently proposed encoder-decoder transformer for oncology, called OncoGPT (see Appendix B).^102,110 Unlike the problem formulation of these works, where monitored biomarkers of treatment response are assumed to be direct measurements of disease state, monitored treatment responses are used in this study to estimate tumor clonal frequencies which represent disease state. Using OncoGPT, the prediction ${\widehat{s}}_i$ at discrete instant $i$ of disease state $s_i$ is provided by the decoder’s output, where the sequences of time-delayed monitored treatment responses $r_{i-n+1},r_{i-n+2},\dots ,{r_{i-1},r}_i$ and estimated disease states ${\widehat{s}}_{i-n},{\widehat{s}}_{i-n+1},\dots ,{{\widehat{s}}_{i-2},\widehat{s}}_{i-1}$ , are fed as inputs to the encoder and decoder, respectively. Here, $n$ is the length of historical context (see Figure 1).

Figure 1.

Prediction of Disease State Based on Repeated Monitoring of Treatment Response Biomarkers. The Histories of Treatment Response and State Estimation are fed as input to Recover the cis and Trans Long-Range Dependencies Imbedded in These Two Sequences.

Once trained as a cancer state predictor, OncoGPT is integrated in an adaptive cancer treatment scheme to provide one-step-ahead predictions of disease state. Assurances about the accuracy and reliability of treatment predictions will naturally be needed before these are used in the synthesis of the next therapeutic action. It may be argued that an adaptive therapy scheme that is constrained to deliver slowly changing therapeutic actions (e.g. drug dosing) would induce small perturbations of tumor growth dynamics that preserve local structural stability of the system, ¹¹¹ maintaining as a result the reliability of OncoGPT predictions of disease state. In all cases however, clinical evaluations of the prediction model will be the cornerstone of future deployments in real-world clinical settings.

Adaptive Treatment Planning

Cancer treatment entails steering the disease from its initial state, asserted at diagnosis time, to a desirable “cured” or “stable” state, using clinically feasible controls, i.e. clinically safe and effective therapies. The design of feasible treatments that cure or manage cancer would ultimately depend on the controllability of the disease as a nonlinear time-varying dynamical system, i.e. the possibility to steer the disease from an initial state to a desired end state. The controllability of nonlinear systems has been studied extensively with the objective to find necessary and sufficient conditions under which a given nonlinear system is controllable.^112-118 Unlike linear systems where the rank condition for controllability was worked out by Kalman^119,120 there are no general characterizations of controllability for nonlinear systems. Sufficient conditions have been established for local controllability ¹¹⁷ defined as the possibility to steer the system in an arbitrarily small time, using valid controls, from an initial state to any other point of a neighborhood that contains the initial state. This is usually labelled as small time local controllability (STLC). ¹¹⁷ In the context of cancer treatment, the problem at hand is to determine the sequence of clinically admissible controls that would steer the system’s state, assumed to be identifiable based on monitored treatment response, to a desired state. The set of desired states must therefore be accessible or reachable from the initial diagnosis state, which is a weaker condition than controllability and easier to decide. ¹¹⁶ Although algorithms have been proposed to check STLC and accessibility of biosystems such as signaling pathways, ¹²¹ it may not always be possible to develop accurate and clinically validated mathematical models of complex systems, such as cancer, for which such check would yield effective insight about its control. Nevertheless, while it may be reasonable to assume that the system is accessible, starting from the initial diagnosis state, it may not be possible to guarantee that cancer can be steered, through treatment, to all possible neighboring disease states. For instance, let us assume the desired treatment outcome is to reach a desired disease end state (cure or manageable state) associated with undetectable ctDNA and a healthy IL-8 level, with all the other biomarkers of interest having some specific values. Let us also assume that steering cancer towards the desired end state from an initial diagnosis state is sought by applying a sequence of controls (instances of drug/therapy administration) as part of a treatment cycle of $n$ timesteps. Even with the accessibility assumption, it may not be possible to guarantee that there will always exist a set of clinically admissible controls that could drive cancer along any arbitrarily selected trajectory from an initial disease state to a desired end state (see Appendix A). Hence, the desired trajectory of disease progression has to be planned taking in consideration the existence of clinically feasible controls that could drive cancer along such trajectory since not all trajectories are clinically feasible. In other words, the saying “all roads lead to Rome” does not apply to time-varying nonlinear dynamical systems such as cancer. In this respect, learning from past patient treatments could be leveraged to improve the selection of setpoints that are most likely to be reachable using clinically feasible treatments. In particular, an instance of OncoGPT trained to predict disease states from treatment response, as presented in the previous section , may be used as a time-series transformer ¹⁰⁷ to predict the next clinically plausible setpoint, given the past trajectory of setpoints and the context represented by the sequence of past treatment responses (i.e. past response trajectory) of the patient (see Figure 2). An interpolation scheme could for example be used to choose the actual next setpoint such that it is a point between the estimate of the current disease state and the disease desired end state that is the closest to the GenAI suggested setpoint in the Euclidean sense (see Appendix A).

Figure 2.

Planning a Desired Disease State Trajectory. The Next Disease Setpoint is Chosen as the State Between the Estimated State and the Desired End State that is closest, in the Euclidean Sense, to the OncoGPT Recommended Setpoint. The Selection of the Setpoint is aided with a Basic Interpolation Scheme, an Example of Which is Provided in Appendix A (supplemental).

The rationale for the use of an instance of OncoGPT, trained as a disease state estimator, to generate control setpoints is motivated by the fact that it embodies a mapping that represents the closest to a ground truth about the actual relationship between treatment response and disease state. It may hence be argued that the setpoint generated in the context of past setpoints using this learned map are clinically feasible because they belong to the set of disease states inferred from treatment responses.

GenAI Adaptive Cancer Therapy

The framing of cancer treatment as an adaptive control problem opens avenues for the use of multiple possible adaptive strategies. ⁴¹ There are multiple potential approaches to the integration of data-driven GenAI learning and predictive capabilities to support cancer treatment decision-making. The diversity of integration approaches is shaped by the extent to which GenAI is used to assist in the realization of one or more of the three functional components of adaptive therapy, i.e. planning disease state trajectory, disease state estimation/prediction and therapeutic cancer control. In particular, there are three potential GenAI integration approaches (see Figures 3–5).

Figure 3.

Adaptive State Feedback Therapy. The Synthesis of Treatment Recommendations is Aided by OncoGPT, Which Provides One-Step-Ahead Disease State Predictions. In This Case, the Oncologist Decides About the Desired Disease State Trajectory (i.e. Desired Trajectory of Disease Progression) as Well as the Heuristic or Treatment Adaptation Scheme Underlying the Adaptive Controller and its Tuning.

Figure 4.

GenAI-Assisted Adaptive Therapy. At Each Instance of Drug or Therapy Administration, the Adaptive Treatment Planner Generates the Next Desired Disease State Based on the Treatment Objective, Past Treatment Responses and Recently Planned Disease States. Disease State is Predicted Based on Past Treatment Responses and Disease State Predictions, Respectively.

Figure 5.

GenAI-Driven Adaptive Therapy. OncoGPT-Driven Treatment Planning, Disease Estimation and Treatment Adaptation Integrated in an automatic feedback control system that carries out Treatment Objectives of the Oncologist Specified in the Form of a Desired end State of the Disease. The Treatment Planner Generates a Trajectory of Desired Disease States Based on the Target End State and The Monitored Treatment Response. At Each Instant of Drug/Therapy Administration, The Error Between The Desired Disease State and The Estimated State is Used to Drive Treatment Adaptation.

The first possible approach of leveraging GenAI’s predictive capability can be realized through the use of OncoGPT as a disease state predictor, detailed in an earlier section, to provide one-step-ahead disease state predictions as feedback for adaptive therapy (see Figure 3). In this case, planning the desired disease state trajectory is undertaken by the oncologist or a clinically validated system. The adaptive controller may embody any one of the many possible adaptive control schemes including ones that are based on PID ( proportional-integral-derivative) controllers. ⁴¹ In fact, the adaptive scheme could be a heuristic similar to the one used in the adaptive androgen deprivation clinical trial (NCT03511196) for metastatic castration sensitive prostate cancer (mCSPC), where adaptation consists in stopping treatment when the PSA (prostate specific antigen) falls below 50% and resuming it on imaging progression and PSA recovery. ³⁷ In addition to PSA and imaging, testosterone level was also used to select one among three possible therapies: Luteinizing hormone releasing hormone (LHRH) analog, new hormonal agents (NHAs), such as abiraterone, or a combination thereof. At 12 months, there was no radiological progression for any of the16 patients enrolled in this mCSPC trial and the secondary endpoints concerning median times to PSA and radiographic progressions were not met at the follow-up time of 26 months. ³⁷ These results assert the clinical feasibility of using PSA and testosterone levels as feedback for adaptive androgen depravation therapy (ADT). Furthermore, a prior clinical trial for metastatic castration resistant prostate cancer (NCT02415621) using a similar heuristic of therapy adaptation but based solely on PSA^35,45 yielded a median overall survival (OS) of 58.5 months and a median radiographic TTP (time to progression) of 33.5 months for patients in the adaptive therapy group vs medians OS and TTP of 31.3 months and 14.3 months for the standard of care (SOC) group, respectively. The success of this evolutionary-based adaptive therapy rests on maintaining a sufficiently large proportion of therapy sensitive cells to compete with and impede the proliferation of resistant cells and ultimately thwart therapeutic resistance. In other words, accurate estimates of the proportions of sensitive and resistant cancer cells is a critically needed feedback to guide therapy adaptation. This is precisely where the use of OncoGPT to predict disease states, as articulated in an earlier section, would support adaptive therapy including adaptive ADT, by providing repeated estimates of clonal frequencies, based on repeated sampling of biomarkers validated for the cancer type at hand. This approach to GenAI-assisted adaptive therapy is likely to be prioritized for clinical validation and adoption in real-world clinical settings, given the minimal extent to which control over treatment planning is delegated to GenAI. Furthermore, the treatment adaptation scheme would be transparent and explainable. Nevertheless, disease state predictions provided by GenAI still need to be clinically validated in terms of their accuracy, precision, sensitivity and reliability.

For the second approach, GenAI is used to plan the desired disease state trajectory and predict disease states (see Figure 4). The errors between desired and predicted disease states are used to drive an adaptive controller of choice. ⁴¹ The treatment planner and the state predictor are built around OncoGPT as explained in earlier sections. This approach, like the first one, has the advantage of leveraging the learning and predictive capabilities of GenAI while maintaining the transparency of the treatment decision-making process and the interpretability of the resulting treatment recommendations. Indeed, the adaptive controller may be based on a PID controller or a practically inspired heuristic whose functionality and output are fully transparent and interpretable respectively, hence ensuring that treatment recommendations will be explainable. However, unlike the first approach, OncoGPT is used to adaptively plan the desired trajectory of disease state that satisfies the treatment objective set by the oncologist. This is expected to improve treatment outcomes because OncoGPT planning relies on its learned model of disease dynamics and the continuously monitored patterns of treatment response to chart a desired trajectory of disease state that is tailored to the specific patient. On the other hand, the reliance on GenAI for both treatment planning and disease state predictions will necessitate a more comprehensive clinical validation of these components, in addition to that of the integrated adaptive therapy system.

For the third potential approach to the use of GenAI in adaptive therapy, all three components of adaptive control, i.e. planning, prediction and control are driven by different instances of OncoGPT (see Figure 5). In this case, the treatment decision-making process is fully driven by OncoGPT, including treatment adaptation whose opacity will require additional attention in the clinical validation process. However, it is important to highlight that treatment adaptation driven by GenAI models that are learned from treatment data of a large number of patients embody an extensive biological and clinical knowledge that is beyond the reach of human cognitive capabilities to translate into heuristics or for mathematically derived adaptive control laws to easily account for.

The three approaches to GenAI-assisted adaptive therapy define a profile for a systematic framework that could be utilized to leverage the learning and predictive capabilities of GenAI models and their underlying large language models (LLMs). The continuously expanding big data and knowledge about cancer and its treatment are rendered into GenAI learned models to address the challenges of adaptive therapy and help improve treatment planning, treatment response prediction and therapy adaptation towards clinical feasibility and adoption of the adaptive therapy paradigm as summarized in Table 1.

Table 1.

Comparison Between Traditional Therapy, adaptive Therapy and GenAI-assisted adaptive Therapy

	Traditional Standard of Care (SOC) Therapy	Adaptive Therapy	GenAI-Assisted Adaptive Therapy
Patient Stratification	Biomarker-based patient stratification at the time of diagnosis.122,123	Biomarker-based patient stratification at the time of diagnosis. 46	Biomarker-based patient stratification at the time of diagnosis.46,102
Dosing	Fixed treatment schedule with pre-defined dosing.124,125	Dynamic treatment adapted based on repeated monitoring of treatment response.35,37,126	Dynamic treatment adapted based on repeated monitoring of treatment response.46,102
MTD Paradigm	Treatment using maximum tolerated dose (MTD).125,127,128	• Dosage is optimized in accordance with the adaptation scheme.35,37,126 The cumulative dose is expected to be lower when compared to standard of care37,46,126• Dosage optimization is based on insight from explicit mathematical models of the disease and driven by repeated (real time) monitoring of treatment response.43,45,46	• Dosage is optimized in accordance with the adaptation scheme where the cumulative dose is expected to be lower when compared to standard of care.46,102,110• Dosage optimization is based on GenAI models, trained on big data about cancer and its treatments, and driven by repeated (real-time) monitoring of treatment response.102,110• The use of data-driven GenAI models allow an implicit consideration of disease dynamics which may not be easily identifiable through explicit data analysis (and inclusion in mathematical models), leading, potentially, to more effective adaptive therapy.102,129-131
Treatment Personalization	Limited capacity for personalized treatments due to SOC adherence to broad guidelines that abstract away cancer diversity and individual patient characteristics. 132	• High capacity for treatment personalization whereby clinically validated mathematical models may be tailored to specific patients to guide dosage optimization strategies.133,134	• High capacity for treatment personalization whereby GenAI models may be tailored to specific patients through fine-tuning based on continuously accumulated monitoring data of patients’ treatments. 102 • The data-driven personalization may lead to more effective personalization of adaptive therapy since it is tuned using the patient’s disease state dynamics learned by the GenAI models. 102
Toxicity	MTD-driven SOC prioritize the eradication of the tumor while measures are undertaken to mitigate side-effects.124,128,135,136	• Does not adhere to the necessity of an MTD treatment approach to cure or manage cancer. Given the expected lower cumulative dose compared to standard of care37,46,126 there is an expectation of lower treatment toxicity.• Mathematical models used to inform dosage optimization strategies can include toxicity as an optimization constraint to minimize side-effects. 137	• Does not adhere to the necessity of an MTD treatment approach to cure or manage cancer. Given the expected lower cumulative dose compared to standard of care37,46,126 there is an expectation of lower treatment toxicity.• Data-driven GenAI models used for therapy adaptation can consider the time-dependent cumulative dose as a proxy variable for patient toxicity to be included as a disease state component subject to control towards a desired, tolerable pattern of toxicity.• The inclusion of continuously monitored patient toxicity as feedback to the GenAI models enables an implicit accounting of toxicity patterns that are specific to the patient under treatment. This could potentially enable a more effective minimization of toxicity.
Resistance Management	SOC therapy is usually unsuccessful in achieving a long-term management of therapeutic resistance given its reactionary, catch-up strategy to cancer’s evolutionary dynamics.16,39,138	Emphasis on preventing/delaying the onset of therapeutic resistance with the goal of long-term management of the disease by modulating therapy based on monitored treatment response taking in consideration cancer evolutionary dynamics.33,35,39	• Emphasis on preventing/delaying the onset of therapeutic resistance with the goal of long-term management of the disease by adapting therapy based on monitored treatment response taking in consideration GenAI learned models of cancer evolutionary dynamics.102,110• Leverages GenAI data-driven learning capabilities to enable the continuous integration of evolving data/knowledge about cancer and the expanding space of therapeutics towards continuous improvement of disease management.

Real-world applications of the proposed framework will hinge on the availability of quality training data in sufficient quantity and the feasibility of clinical validation of reliability and performance of the overall GenAI-assisted cancer treatment system. Furthermore, adaptive therapy approaches that limit the scope of GenAI use to disease state prediction and treatment planning maintain the primacy of oncologists in the decision-making process, which is critical to safeguarding adherence to clinical norms and guidelines. As long as the GenAI support of the decision-making process does not replace the oncologist’s judgment, this would not significantly change the patient-clinician relationship and hence would not alter applicable ethical and legal norms, including professional accountability and liability. ¹³⁹ On the other hand, the third proposed approach may not be clinically feasible or acceptable in the short term given the implied distancing of oncologists from the treatment decision-making process. However, there are other possible configurations of GenAI-assisted adaptive therapy that may be pursued along a clinician-in-the-loop framework where the oncologist may elect to override the outputs of GenAI models based on applicable guidelines and expert recommendations, filtered through their judgments. Treatment planning and adaptation are tasks that could benefit from a hybrid approach whereby oncologists are provided with the ability to override GenAI recommendations asserted to be unsafe or inconsistent with known evidence and fallback to an acceptable alternative. For recommendations that lie outside these categories and lack support from the oncologist, a strategy would be needed to resolve disagreements. One possible approach is to rely on multiple instances of GenAI models to generate multiple recommendations, regarding the same task, as options for the oncologist to choose one that is most consistent with their experience and judgment. Augmenting LLMs’ recommendations with generated explanations would also improve transparency of the underlying knowledge model and its predictions. ¹⁴⁰ By the same token, these would provide oncologists with a tool to assess the outputs of GenAI models, potentially improving trust in their reliability.

The training of the various instances of OncoGPT would be carried out in two steps. In the first step, treatment data (drugs/therapies and doses) and treatment response data (readings of response biomarkers) from tumors of the same cancer type should be used. The second step should involve fine-tuning using data curated from phenotypically similar patients that had undergone similar types of therapies, yielding several instances of trained OncoGPT, each applicable to a specific phenotypic class of patients. Training datasets may be curated from patient-derived xenografts (PDXs) whose response to treatment are known to replicate corresponding patient treatment outcomes.^141-144 PDXs are widely used in pre-clinical trial validation of cancer drugs and represent ideal systems for studying drug combinations and resistance.^144-147 Indeed, PDXs preserve salient features of patients’ cancers with respect to the many dimensions of tumor biology and treatment, including histology, genetic alterations, gene expression, the tumor microenvironment and treatment response.^144,148,149 Furthermore, PDX systems afford an unmatched flexibility with respect to the coverage of genetic and phenotypic diversity, the controlled manipulation of drug combinations and doses, and the comprehensive monitoring of treatment response biomarkers as well as clonal composition and frequencies. The suitability of PDXs in conserving cancerous features of patient tumors and replicating patient treatment outcomes, provides a strong rationale for their use as surrogate sources of quality cancer treatment datasets for the training of GenAI models. However, data for the fine-tuning phase of training should preferably be sourced, whenever possible, from clinical trials and actual patient treatments to account for all determinants of treatment response, including the immune system and patient health state. On the other hand, training LLMs using PDXs data, prior to fine-tuning with clinical data, involves an implicit assumption of learning being transferred from the model trained on the domain of experimental PDX treatments (source domain) to the fine-tuned model ultimately used in the clinical domain of patient treatments (target domain). The correspondence between the source and target domains and associated tasks, and the risk of negative learning (i.e. unhelpful or hindering new learning) are among the issues that require attention. ¹⁵⁰ The successful use of PDXs as pre-clinical models of patients’ cancers suggests an alignment between the source and target domains in terms of similarity of the feature sets determining drug resistance and treatment response and the corresponding sets of labels, i.e., disease states. Furthermore, the similarity of tasks under consideration (e.g. disease state predictions) in both domains and the capacity of deep learning models to learn high-level features that are transferable,^151,152 diminishes the concern about the risk of negative learning. In all cases, however, there are emerging approaches that can be used for the detection and mitigation of negative transfer such as domain similarity estimation, and source data filtering.^153,154

Clinical Evaluation

Randomized clinical trials (RCTs) represent the standard and most trusted mean to verify the efficacy and safety of clinical interventions. However, clinical evaluations of AI applications in healthcare and their unique challenges may not be adequately addressed by traditional RCT designs. ¹⁵⁵ Several reporting guidelines for clinical trials of interventions involving AI have been developed, including CONSORT-AI, ¹⁵⁶ DECIDE-AI ¹⁵⁷ and SPIRIT-AI. ¹⁵⁸ Beyond their reporting objective, these guidelines contribute to delimiting the scope and aspects that require attentions in the design of RCTs for AI interventions. On the other hand, while RCTs represent the most trusted mean to demonstrate clinical impact of AI interventions ¹⁵⁹ they may not always be the most adequate approachs to generate evidence for AI use in real-world clinical settings. ¹⁶⁰ In all cases, however, a framework for the clinical evaluation of AI applications in oncology and medicine at large is needed. In this respect, the FDA guidance for the clinical evaluation of SaMDs (software as a medical devices) provide a plausible framework and a corresponding process to inform the conception of a roadmap for the clinical evaluation of GenAI-assisted adaptive therapy. ¹⁶¹ In particular, an evaluation roadmap may be conceived to mirror the three steps of the FDA supported risk-based framework, namely: (1) validation of clinical association between the outputs of the GenAI models and the relevant clinical conditions, (2) analytical validation of the GenAI inputs’ processing to be correct and of the GenAI generated outputs to be accurate, precise and reliable, and (3) clinical validation of the GenAI output to accurately, precisely and reliably achieve the desired outcomes for the target patient population in the real-world clinical care setting. ¹⁶¹ The three distinct instances of OncoGPT used in treatment planning, disease estimation and control (therapy) adaptation involve different concepts (disease state estimation/prediction, and therapeutic control), hence requiring distinct approaches and corresponding metrics to demonstrate clinical association. In particular, associations between disease states (clinical condition) and their GenAI estimates/predictions may be asserted using Spearman’s rank correlation coefficient. ¹⁶² On the other hand, since the association between disease states and therapeutic control is nonlinear, mutual information would be in this case more appropriate.^163,164 This step of the evaluation process may be undertaken using data collected from a representative group of the target patient population. For the analytical validation step, specific metrics that quantify accuracy, precision and reliability would need to be defined. Since the objective of adaptive therapy is to reduce the gap between the desired and monitored disease state (control error), precision may be defined as the average control error over a cohort of patients undergoing multiple cycles of therapy. Precision, may then be defined as the standard deviation from the average control error, while the reliability could be defined as the fraction of the control error’s variance that is due to disease state’s variance. ¹⁶⁵ Given their successful use in pre-clinical trials to validate cancer drugs, PDXs represent suitable models for use in the pre-clinical testing required to generate evidence for the second step in the evaluation process. The objective of the clinical validation step is to establish that GenAI-assisted adaptive therapy realizes the desired clinical outcomes for the target patient population. Although RCTs for AI applications are still scarce,^155,166,167 they embody the most trusted approach to validate the safety and efficacy of GenAI-assisted adaptive therapy in improving clinical outcomes such as survival and quality of life. However, RCTs must consider the unique complexity associated with the dynamic modulation of therapy, its risks to patient safety, and the “black box” nature of the LLMs underling GenAI and their vulnerabilities to hallucinations. In this respect, a series of RCTs would be needed to address the diverse aspects supporting GenAI-supported adaptive therapy. As discussed earlier, validating OncoGPT estimation of disease states to support adaptive therapy is likely to be prioritized for clinical trials where therapy adaptation is maintained under the direct control of the oncologist. Data generated from RCTs for disease state estimation would be instrumental in validating OncoGPT capabilities of planning desired disease state trajectories and recommending plausible and safe therapeutic controls. This may be undertaken using a proof-of-concept implementation to validate each one of the three aspects (planning, estimation, and control) independently, before considering the validation of the integrated system. In particular, the trajectories of disease states could be used as reference to assess the extent to which OncoGPT generates precise, accurate and reliable trajectories of desired disease states based on the oncologist specified treatment outcome (i.e. desired disease state) and an appropriately deep historical context of OncoGPT generated output. Likewise, validation of OncoGPT’s ability to generate plausible and safe therapeutic controls may use the same data generated from RCTs dedicated to GenAI disease state estimations where OncoGPT planned and estimated disease state trajectories are used as inputs to the OncoGPT-driven adaptive controller, while the oncologist’s therapeutic decisions would be used as a reference to assess the plausibility of its output. The validation steps highlighted so far will ultimately inform the design of subsequent RCTs to demonstrate the ability of GenAI-assisted adaptive therapy with respect to clinically meaningful patient outcomes such as overall survival, improvement of symptoms and quality of life.

The proposed clinical roadmap provides an overview of the evaluation activities that needs to be undertaken to generate evidence that would be instrumental for regulatory approval. The adoption of the FDA risk-based framework for the evaluation of SaMDs is a plausible starting point in paving the path for acceptance and adoption of AI-assisted adaptive therapy in the clinic. However, the many mentioned layers of complexity of such AI-assisted adaptive treatment systems, may require a pragmatic approach that prioritize achieving well proven clinical successes as a first step towards a broad and more comprehensive future adoption of AI-assisted adaptive therapy. One example of such approach is to build on the success of adaptive therapy for prostate cancer^34,37 through the conception of RCTs where GenAI predictions of tumor clonal frequencies should be used to support an adaptive therapy scheme that relies on fully explainable and interpretable heuristics of adaptation.

Discussion

Like all frameworks, ours includes some limiting assumptions. In particular, OncoGPT’s potential to support the improvement of cancer treatments is predicated on the availability of large, quality, multi-modal datasets about cancer and its treatments for GenAI models’ training. Furthermore, the framework is also predicated on the feasibility of repeated monitoring of pertinent treatment response biomarkers during treatment. Advances in treatment response monitoring using LB and the increasing number collaborative efforts to collect and curate cancer data are expected to overcome these assumptions.

The quality of training data is critical for LLMs to yield sufficiently accurate predictions of treatment response to support adaptive therapy. Such data quality would have to involve an adequate sampling of the phenotypic diversity of the target patient population to achieve a uniform performance of deep-learning-assisted adaptive therapy over a wide range of patient characteristics. The need for training data to be representative of patients’ phenotypic diversity is aligned with the results of a recently proposed deep reinforcement learning (DRL) framework for personalized adaptive therapy, whereby the use of data from a cohort of characteristically diverse patients led to a DRL network that is robust to variations of tumor model-fitted parameters. ¹⁶⁸ The extent to which training data span the space of possible trajectories of tumor progression under treatment is another quality aspect that is critical for LLMs to learn faithful models of treatment response dynamics across a wide range of treatments. In fact, training data has to embody no less than the richness resulting from probing tumors in a systematic and specific way akin to nonlinear system identification methods.^70,169 In this respect, MTD (maximum tolerated dose) treatments may not entail sufficient selective probing to uncover the range of expected tumor progression dynamics under adaptive therapy. One potential remedy is to complement MTD treatment data with data from probed PDXs using nonlinear system identification approaches^70,169 to yield quality training data for GenAI models. The collection and curation of PDX and clinical data will require concerted efforts that consider the unique needs of LLM training in oncology. Existing initiatives such as the NCI (National Cancer Institute) Patient Derived Xenograft Network (NCI-PDXNET), the European Organization for Research and Treatment of Cancer (EORTC) and the Prostate Cancer Clinical Trials Consortium (PCCTC) may provide collaborative avenues for data curation and sharing. However, the potentially transformative impact of GenAI in oncology is a strong rationale for the creation of a consortium, not unlike the TCGA (The Cancer Genome Atlas) program, dedicated to the collection and curation of data for the training of GenAI models in oncology. Aside from the need for quality training data, the training of LLMs faces the challenge of data overfitting and the consequent susceptibility to memorize high frequency patterns and learn invalid correlations that are not generalizable, requiring hence mitigations that includes regularization, data augmentation and hyperparameter tuning.^170-172 Since OncoGPT’s space of training, learning and inferencing is assumed to be spanned by the dynamics of cancer as a nonlinear system, augmenting training data using synthetic data sourced from mathematical models of tumor growth dynamics under treatment may be a suitable data augmentation approach in this context.

Another aspect of concern is the inherent diversity of patient characteristics, such as age and comorbidity, which may necessitate patient-tailoring of GenAI-assisted adaptive therapy. In particular, dedicated instances of personalized OncoGPTs, obtained through fine-tuning, ¹⁰² may need to be used for distinct groups of patients classified based on biomarkers, phenotypic similarity, ¹⁷³ and patient characteristics. Beyond the personalization of GenAI models to address aspects of performance and efficacy across the target patient population, patients’ inputs and their preferences would need to be considered as core components of the clinical decision-making process. ¹⁷⁴ This may also be addressed through fine-tuning of the relevant instances of OncoGPT using patients’ values and treatment preferences. Such personalization must however be supported with transparent and patient-centric treatment approaches involving patient-doctor interactions that honor patient preferences and earn satisfaction regarding selected treatment strategies. ¹⁷⁴

Given the magnitude of cancer treatment challenges, the utility of GenAI learning and inferencing capabilities is paramount to the conception of a more effective cancer treatment paradigm. The different instances of OncoGPT are trained to learn mappings between one or more of the pertinent variables, i.e. disease state, treatment response, control and control error, where the objectives are to predict or inference the next disease state, plan the next treatment setpoint, and control the patient’s disease given the relevant input and historical context. However, the inevitable hallucination problem of LLMs underlying GenAI^175-177 represents a venerability that must be addressed. LLM hallucinations are manifested in the form of output content non-aligned with relevant reference knowledge or is outright imagined and cannot be asserted by evidence. A significant number of techniques have indeed been explored to mitigate LLM hallucinations ¹⁷⁸ and may inform the development of mitigation approaches applicable to OncoGPT and its mission-critical healthcare delivery environments. In the case of adaptive cancer therapy, the tokens being processed by OncoGPT are discrete values of one or more variables representing treatments, treatment responses and the states of cancer as a time-varying nonlinear dynamical system. One potential remedy to LLM hallucinations, in this case, is to check the clinical plausibility of OncoGPT’s outputs based on additional biomarkers or clinical variables. ¹⁰² In the events of a refuted clinical plausibility, a rectification of the output based on maximizing a pertinent reward is one possible mitigation technique. ¹⁷⁹ Perhaps the most promising remedy for LLM hallucinations in the safety-critical context of cancer treatments is to complement OncoGPT with a RAG ( retrieval-augmented generation) ¹⁷⁹ model to condition the generation of OncoGPT’s output on retrieved non-parametric factual and up to date clinical knowledge from genomic, radiomic and clinical databases such as GENIE. ¹⁸⁰ Clinical knowledge that would power RAG need also to include cancer management guidelines such as the 6^th and 7^th advanced breast cancer guidelines, ¹⁸¹ classification systems such as ESMO ( European Society for Medical Oncology) scale of clinical actionability for molecular targets ( ESCAT) ¹⁸² and databases of regulatory requirements such as FDALabel. ¹⁸³ This knowledge may be indexed into a heterogenous knowledge index, ¹⁸⁴ whereby treatment recommendations, monitored treatment responses, disease state estimations, and patient characteristics would be parsed to extract the keys for RAG’s real-time retrieval and selection of the most relevant knowledge that would support LLMs’ generation of treatment recommendations. However, best practices and guidelines for assessment of LLMs’ output need also to be developed by the oncology community to support enforcing a backstop for unsafe recommendations using clinician-in-the-loop frameworks. This concern may require more attention for the case of healthcare systems where the scarcity of trained personnel may lead to a critical dependence on GenAI, whose vulnerabilities may detrimentally impact the safety of cancer care.

Regulatory compliance, acceptance and adoption of GenAI-assisted adaptive therapy face unique challenges associated with explainability, bias, performance drift as well as opacity and hallucination vulnerabilities of their underlying LLMs and their impacts on aspects of patient treatments, including safety, accountability, liability, patient-clinician communication, trust, and ethical concerns.^185-188 These issues would have to be addressed within the regulatory approval process and need further mitigation throughout the life cycle of AI-assisted therapy systems in the clinic. While the proposed evaluation roadmap would be instrumental in removing many of the hurdles to regulatory approval, more research is needed in strengthening quality assurance and bioethics frameworks used in clinical institutions to facilitate regulatory approval and integration of AI-assisted adaptive therapy in clinical practice. Ultimately, the successful integration of GenAI in oncology will depend on healthcare delivery environments, their cultures of technology adoption and capacity for innovation, and the extent to which clinical institutions can overcome resource constraints and build the necessary computing and data infrastructures. World-wide, the hurdles to GenAI adoption in the clinic may be more challenging to overcome, especially in LMICS (low to medium income countries) where resources and trained personnel are scarce. ¹⁸⁹ However, irrespective of the socio-economy context of healthcare delivery, the long-term potential for GenAI to reduce the societal burden of cancer through significant improvements of treatments while addressing cost and personnel challenges may provide the incentive for investments in training and development of computing and data infrastructure necessary to realize the potential of GenAI in cancer care.

Conclusions

The persisting impotence of current standard of cancer care to decisively thwart therapeutic resistance and improve patient outcomes is a compelling incentive to leverage the data-driven learning capabilities of generative artificial intelligence to support the realization of adaptive cancer therapy by providing disease state predictions and treatment recommendations. In this respect, cancer treatment is framed as an adaptive control problem whose solution entails addressing its planning, prediction and control components. Multiple models of GenAI-assisted adaptive cancer therapy are explored, reflecting the many possible integration degrees of pre-trained LLMs. This accounts for the diversity of healthcare delivery environments and their varying cultures of technology adoption and capacity for innovation. The clinical success of GenAI-assisted adaptive cancer therapy will ultimately depend on its ability to improve patient outcomes and thwart therapeutic resistance as the most pressing challenge of cancer treatments.

Supplemental Material

Supplemental Material - Generative AI - Assisted Adaptive Cancer Therapy

Supplemental Material for Generative AI - Assisted Adaptive Cancer Therapy by Youcef Derbal in Cancer Control