Adaptive Cancer Therapy in the Age of Generative Artificial Intelligence

Abstract

Therapeutic resistance is a major challenge facing the design of effective cancer treatments. Adaptive cancer therapy is in principle the most viable approach to manage cancer’s adaptive dynamics through drug combinations with dose timing and modulation. However, there are numerous open issues facing the clinical success of adaptive therapy. Chief among these issues is the feasibility of real-time predictions of treatment response which represent a bedrock requirement of adaptive therapy. Generative artificial intelligence has the potential to learn prediction models of treatment response from clinical, molecular, and radiomics data about patients and their treatments. The article explores this potential through a proposed integration model of Generative Pre-Trained Transformers (GPTs) in a closed loop with adaptive treatments to predict the trajectories of disease progression. The conceptual model and the challenges facing its realization are discussed in the broader context of artificial intelligence integration in oncology.

Introduction

The space of treatment options available to cancer patients is ever expanding, fueling the drive for treatment optimizations that involve combining multiple types of drugs either sequentially or concurrently to overcome acquired resistance and hedge against minimal residual disease.^1,2 There are more than 27, 000 clinical trials currently registered in CinicalTrials.org for combination cancer therapies, highlighting the magnitude of the challenge facing the design of optimal cancer treatments. The lack of effective biomarkers, the limited number of available patients compared to the number of possible drug combinations, and the cumulative toxicity resulting from drug combinations are some of the major barriers to improving patient outcomes. Although drug combinations being explored through clinical trials may not always have clear rational underpinnings, ¹ even drug combinations that are based on rationales such as targeting reactivated signaling pathway, synthetic lethality, inhibition of driver pathways, and enhancing immune response, will have to reckon with the evolutionary dynamics of tumor growth that often leads to therapeutic resistance. Adaptive therapy (AT) is a treatment paradigm conceived with the objective to manage the disease through an explicit consideration of cancer evolutionary dynamics. ³ Treatment adaptation is applied through dose scheduling (ie, dose timing and/or dose modulation) to maintain a tumor burden with a sufficient proportion of therapy-sensitive cancer cells to suppress the proliferative growth of therapy-resistant ones.^4,5 A pilot clinical trial of adaptive therapy (NCT02415621) was conducted for metastatic castrate-resistant prostate cancer (mCRPC) using abiraterone monotherapy, and showed a significant improvement in time to progression (TTP) and overall survival (OS) compared to standard of care (SOC).^6,7 Adaptive therapy has since been the object of other clinical trials, including those for melanoma (NCT03543969) and mCRPC (NCT05393791). These and other planned clinical trials of adaptive therapy for different cancer types are expected to yield crucial clinical data for the parameterization of mathematical models of tumor response, which are critical to the design of adaptive treatment strategies. However, there remains many open questions that need to be addressed towards paving the way for adaptive therapy to become a mainstream protocol of cancer treatments. Many of these open questions have been raised in the context of mathematical and computational modelling as a catalyst for clinical translation of adaptive therapy. ⁸ The questions were binned into 3 broad categories ⁸ : (1) mathematical modeling, (2) design, and (3) clinical translation of adaptive therapy. The first category focuses on the elements that would need to be included in mathematical models to support experimental and clinical investigations of adaptive therapy. Open questions in this category revolve around the components that mathematical models should consider, including phenotypic plasticity, drug-induced mutations, normal tissue homeostasis, inter-tumor heterogeneity, the tumor microenvironment, and the strength of competition between sensitive and resistant phenotypes. These factors highlight the complex adaptive nature of cancer, ⁹ manifested through a time-varying, nonlinear response of tumors to therapeutic interventions, which may as a result require a near real-time prediction of tumor growth dynamics to enable an effective adaptive therapy. On the other hand, the design and validation of adaptive treatments will also require the investigation of numerous open issues, including the optimization of dose administration, the combination of multiple drugs and the stratification of patients to distinguish those for whom adaptive therapy is more advantageous to use for the management of the disease instead of opting for a cure through a current standard of care strategy. ⁸ The clinical translation of adaptive therapy faces another set of open questions and related challenges. In particular, adaptive therapy is fundamentally dependent on a timely disease state feedback to predict disease progression trajectory under therapy, which is essential to treatment decision-making. However, there are numerous challenges facing the feasibility of real-time predictions of treatment response, including the lack of clinically calibrated mathematical models, the inherent uncertainty associated with any predictive mathematical model, and the lack of timely and quality clinical data about disease progression for patients under therapy. ⁸

The challenges facing the clinical success of adaptive cancer therapy mirror in their categorization the steps of the general problem-solving process, ie, problem understanding/analysis, solution design, and solution implementation. The fundamental dependence of adaptive cancer therapy on a timely disease state feedback and predictions of tumor progression trajectory, makes it imperative to develop prediction models of treatment response that are sufficiently accurate to support adaptive cancer treatments. While eco-evolutionary mathematical models based on the perspective of drug-resistant cancer cells competing with drug-sensitive ones has yielded initial clinical success, the open questions and challenges facing the paradigm of adaptive cancer treatment require further advances in expanding our actionable understanding of cancer. In this respect, artificial intelligence models that are trained using the vast amount of accumulating multimodal cancer data (clinical, molecular and radiomic) hold the promise of closing the gap in human knowledge about cancer, which may in turn enable the development of effective cancer treatments for all patients. As an exploration of this perspective, the article offers a conceptual framework for the use of pre-trained generative artificial intelligence (GenAI) models ¹⁰ as predictors of treatment response, integrated in adaptive treatment decision-making systems to provide timely disease state estimation feedback.

Artificial Intelligence in Oncology

The anticipated results of AT clinical trials, that are either planned or in progress for different cancer types, are expected to be translated into personalized therapeutic strategies whose effective realizations would need the leverage of mathematical modeling and artificial intelligence (AI). The utility of mathematics and AI span treatment-related aspects such as modeling disease progression, predictions of treatment outcomes and toxicity, and treatment adaptation, all deemed essential for treatment decision-making. ¹¹ These treatment-related dimensions are in turn dependent on a myriad of specific tasks for which AI is increasingly explored as a feasible enabler, including cancer subtyping, patient stratification, prognosis prediction, treatment selection, and treatment response prediction.^12-17 In particular, the integration of clinical, radiomics, histopathology and molecular data has the potential to advance precision oncology, ¹² and may be an essential ingredient towards the development of clinically effective AI-driven models of cancer subtyping and patient stratification. Catalyzing the convergence towards this objective is the increasing maturity of digital pathology^18,19 and radiomics^20,21 combined with advances in liquid biopsy^22,23 and next-generation sequencing (NGS). ²⁴ AI has also been explored for optimizing cancer treatments through various avenues such as the matching of cancer patients to drug combinations, ²⁵ and the conception of “cancer patient digital twins” as a foundation for predictive oncology. ²⁶ Provided their availability for a sufficiently large patient population, clinical, genomic and radiomic data can in theory be harnessed to train AI systems which would then be used to stratify and predict treatment responses towards more effective cancer tratements.^13,16,27 For instance, machine learning (ML) driven by clinicogenomic, radiologic and histopathologic data have shown promising results with respect to risk stratification of high-grade serous ovarian cancer. ²⁸ Recently, deep learning models trained using data from a large cohort of patient where shown to predict the occurrence of pancreatic cancer within a 3 years future window. ²⁹ The use of ML to plan radiation therapy (RT) for prostate cancer patients, ³⁰ and to identify patients that are likely to need acute care visit during RT or chemoradiation ³¹ are among the few instances of clinically integrated applications of AI in oncology. These studies are part of intensified efforts to explore the utility of AI in cancer diagnosis, prognosis, prediction and treatments.^13,27,32-35 These efforts are progressing along with clinical trials^33,36-38 and an expanding list of AI-based Software as Medical Devices (SaMDs) approved by the Federal Drug Administration (FDA). ³⁹ The largest proportion of these approved devices targets cancer radiology and pathology, ³⁹ representing advances in line with AI pattern recognition and classification capabilities for cancer diagnostics.^40-44

The potential utility of AI to treatment decision-making has been premised on the assumption that AI models trained on clinical and omics data of many patients would assist in the design of effective treatment strategies for new patients. This AI approach of “learn from many to treat one” (LMTO) has to account for cancer inter-patient and intra-tumor heterogeneity as fundamental factors that affects the generalizability of AI predictive models. First, while tumors from distinct patients may share histological and molecular features, they are ultimately distinct time-varying stochastic nonlinear systems whose trajectories of growth will inevitably diverge under treatment. Second, tumors of 1 cm³ weighting no more than 1g may contain up to 10⁹ dynamically interacting cells yielding a near-infinite number of diverse time-varying nonlinear treatment response trajectories driven by evolutionary dynamics. Third, the nonlinear dynamics of tumor growth are bound to, and dependent on the specific health state trajectory of the patient, making it imperative for therapeutic interventions to be adaptive to the specific patient’s health state dynamics rather than guided by the treatment recommendations of ML models, which would be driven by population data. Furthermore, while AI integration in real-world clinical workflows is burgeoning,^30,31 there is emerging evidence highlighting the generalizability shortcoming faced by the application of predictive AI models in medicine.^45,46 Notwithstanding the generalizability question and the challenges of integrating AI in real-world clinical setting,^32,38,47-49 data-driven AI may in the long run be pivotal to optimizing cancer treatments. Indeed, while adaptive therapy is conceived to address the challenges of cancer evolutionary dynamics and the ensuing time-varying, nonlinear dynamics of tumor response to therapy, its performance depends on an accurate monitoring and prediction of treatment response. Given the accumulating big data about treatments and their outcomes, GenAI could in principle be trained to learn a prediction model of the treatment-response causal relationship, providing hence the disease state feedback required to adapt cancer therapy.

Monitoring Treatment Response

The time-varying, nonlinear dynamics of cancer response to therapy makes it imperative to monitor patient disease state at sufficient frequency to achieve effective cancer management or cure. Tumor burden, tumor clonal composition, as well as immune response and immune cell metabolism in the tumor microenvironment (TME) are among the many observable phenotypic features that should be monitored to gauge evolving disease state dynamics. Advances in liquid biopsy (LB),^50-52 NGS, radiomics^21,53 and radiogenomics ⁵⁴ are expected to provide the necessary means for non-invasive, repeated observations of treatment response^55,56 and the estimation of the overall state of disease progression.^23,57 Conventional disease monitoring relies on the RECIST (response evaluation criteria in solid tumors) assessment of treatment response. ⁵⁸ The RECIST states (ie, stable disease, partial response, complete response, and progressive disease) need, however, to be expanded into fine-grain disease states to achieve a sufficiently adequate observational resolution for an effective therapy adaptation to the time-varying, nonlinear dynamics of tumor treatment response. ¹¹ These complex dynamics of cancer are driven by genetic,^59-61 eco-evolutionary^62-64 and immunological^65-69 causal dimensions, which must therefore be repeatedly monitored in order to construct a reliable estimate of disease state. The monitoring of eco-evolutionary determinants of cancer would entail identifying measures of temporal and spatial intratumoural heterogeneity, TME metabolic resources and TME immune cell infiltration. The conception of these measures could be guided by the eco-evolutionary tumor classification framework and corresponding Evo-index and Eco-index proposed to quantify neoplastic cell diversity in a tumor and its changes over time as well as the components defining the ecology of the TME, respectively. ⁷⁰ Moreover, the classification involves an integrated consideration of the genetic and immunological dimensions shaping treatment response, including spatial and temporal intratumoural genetic, epigenetic and phenotypic heterogeneity, genetic instability, TME metabolic heterogeneity vis-à-vis nutrients, vasculature and hypoxia, and TME immune heterogeneity regrading immune cell abundance, their activation states and proximity to cancer cells. The classification yields 16 tumor classes, covering the combinations of high/low levels of evolutionary factors (D: diversity, $dD/dt$ : change of D in time) and ecological factors (H: hazard, R: resources). ⁷⁰ Advances in LB and radiomics would enable repeated monitoring of diversity (eg, Shannon index ⁷¹ ), TME immune conditions (eg, immunoscore,^72-74 immune profile^67,75), tumor metabolism,^76-78 and tumor burden.^79,80 Assuming tumor burden and the 4 eco-evolutionary factors to be the sole determinants of treatment response, quantizing their monitored values using $n$ levels, with n > 2, would lead to $n^5$ possible states that the tumor/disease would span under therapy. For n = 5, representing levels such as Very Low, Low, Medium, High and Very High, the disease state-space would cover 3125 possible states. The consideration of the 4 RECIST states would bring the number of possible disease states to 12 500, which would provide an adequate resolution for the prediction of treatment response trajectories. Starting from an initial tumor state $s_0$ at the time of diagnosis, there is a large number of possible state trajectories the tumor may take under therapeutic intervention. Often, the prognosis for a chosen conventional (non-adaptive) therapeutic strategy would be made based on disease biomarkers sampled at the time of diagnosis. This would ultimately make up the basis for therapy selection. However, a one-time prediction of the ultimate treatment outcome would not be sufficient to support adaptive therapy, which requires repeated monitoring and predictions of treatment response.

Can Generative Pre-Trained Transformers Predict Treatment Response?

In adaptive therapy, a treatment cycle consists of a sequence of therapeutic actions (drug administration, radiation, etc.) $u_i$ , $i\ge 0,$ at instances of time, identified by the integer $i$ . Given the tumor’s initial state $s_0$ (ie, prior to treatment commencement), applying the therapeutic actions (controls) ${\hspace{0.17em}u}_1,u_2\dots ,u_{i-1},u_i$ would lead to the treatment responses ${\hspace{0.17em}s}_1,{\hspace{0.17em}s}_2,\dots ,{{\hspace{0.17em}s}_{i-1},s}_i$ , respectively. The synthesis of the control $u_i$ that would be needed at instant $i$ to yield a desired treatment response ${\hspace{0.17em}\overline{s}}_i,$ requires an estimate of the expected treatment response ${\hspace{0.17em}\widehat{s}}_i.$ The difference between the desired and estimated treatment response would be used to modulate the controls using an adaptive scheme of choice. ⁸¹ The treatment response $s_i$ may be estimated as a nonlinear function $f\left(U_{i-1},S_{i-1}\right)$ of applied controls and past treatment responses, represented by $U_{i-1}=\left[u_0,{\hspace{0.17em}u}_1,\dots ,u_{i-2},u_{i-1}\right]$ and ${\hspace{0.17em}{S}_{i-1}=[s_0,s}_1,{\hspace{0.17em}s}_2,\dots ,{\hspace{0.17em}s}_{i-2},{\hspace{0.17em}s}_{i-1}]$ , respectively. Consequently, an estimate ${\widehat{s}}_i$ of treatment response may be achieved by learning the nonlinear mapping $f.$ The mathematically established result that multilayer feedforward neural networks can be used as universal approximators, with an arbitrary precision, of any nonlinear function $f$ ^82-85, provides the rationale for predicting treatment response using GPTs ¹⁰ since these are based on neural networks with multiple hidden layers, ie, deep neural networks (DNNs). In the case of treatment response prediction, the attention mechanism of GPTs captures the respective auto-correlations of treatments and treatment responses on one hand and the cross-correlation between treatments and treatment responses on the other. These correlations are embodied by the weights of the DNNs underlying the GPTs, which are learned through training. ⁸⁶ Once trained using a sufficiently large number of sequences of treatments and corresponding sequences of treatment responses, the GPT, that we will refer to as OncoGPT to highlight the target application, would be used to provide a one-step ahead prediction of treatment response as part of a closed-loop adaptive therapy, ⁸¹ as illustrated in Figure 1.

Figure 1.

One-step ahead prediction of treatment response using OncoGPT. The prediction or “generation” of the expected next tumor state (treatment response) is made in the context of the N most recent controls and the corresponding N tumor responses, where N is a positive integer representing the length of the historical context being considered. The delays indicate that at instant $i$ the values of controls and treatment responses are known only up to the last previously monitored values.

The success of large language models (LLMs) illustrated by the demonstrated utility of ChatGPT, Gemini, and other LLMs for various cognitive tasks such as writing, computer programming and natural language (NL) translation, embodies a strong rationale for exploring the potential utility of transformers in predicting cancer treatment responses. Framing the prediction of treatment response as an inference problem for a generative pre-trained transformer is a logical extension of GPT’s application to natural language translation. The controls (therapeutic actions) and disease states (treatment responses) are analogous to words in the vocabularies of natural languages of interest in a translation task. A cycle of treatment, which consists of a sequence of controls, is analogous to a phrase, while the resulting sequence of disease states (sequence of treatment responses) corresponds to a translation of the phrase to another natural language. The overall GPT architecture is adopted as initially conceived. ¹⁰ However, the inputs of the encoder part of the OncoGPT consists of the sequences of controls (ie, treatment cycle) rather than natural language sentences while the sequences of disease states (ie, treatment responses) serve as the inputs to the decoder. The input vocabulary is made up of the set of distinct controls, while the output vocabulary is the set of distinct disease states. The set of “words”, ie, the set of controls or disease states, that make up these two vocabularies, are the tokens whose embeddings will be processed by OncoGPT. ¹⁰ Once trained, as will be detailed in the next section, OncoGPT would be used to infer the expected sequence of treatment responses, given a treatment (ie, a sequence of controls). OncoGPT may be used to explore different possible treatments and help select the best possible treatment to be applied without any adaptation based on the inferred sequences of treatment responses. Such inferencing may also be used to guide the design of clinical trials. In the context of adaptive cancer therapy, OncoGPT would be part of the adaptive treatment loop (see Figure 1) with the objective of predicting the next treatment response given a sequence of past controls. As an example, consider an adaptive treatment cycle of 2 weeks, where the treatment response is predicted daily by OncoGPT and fed to the adaptive controller which decides about the next therapeutic action. Initially, ie, at instant $i=0$ , the encoder part of OncoGPT is fed the sequence $u_0{\hspace{0.17em}u}_0\dots u_o\hspace{0.17em}{u}_0$ of cycle length, ie, 14, as input u₀. At this point in time, the input to OncoGPT decoder is a single token that consists of the initial disease state $s_0$ asserted based on the initial diagnosis of the disease. Running OncoGPT for 1 step would yield the first predicted disease state ${\widehat{s}}_1$ which will be used by the adaptive therapeutic strategy to yield the control $u_1$ . The input of the encoder will then be updated to $u_0{\hspace{0.17em}u}_0\dots u_o\hspace{0.17em}{u}_1$ . Once the control $u_1$ is applied, the monitored (ie, actual) treatment response $s_1$ would then be known and can therefore be used to update the input sequence of the decoder to $s_0$ $s_1$ . Figure 2 illustrates the traces of the encoder and decoder inputs for the 14 steps required to complete the example 14-day treatment cycle.

Figure 2.

Traces of OncoGPT inputs and its output predictions for a 14-day cycle adaptive treatment.

The proposed OncoGPT includes a notable deviation from the original GPT architecture, ¹⁰ whereby the decoder’s input is the sequence of past actual disease states instead of the past predictions as it would be for other problems such as language translation. The use of the actual history of tumor’s treatment response instead of the predicted one provides the most relevant inferencing context. As such, it is plausible to hypothesize that the approach would yield a higher accuracy of inferencing the next treatment response given a sequence of past controls and corresponding treatment responses.

Training OncoGPT to Predict Cancer Treatment Response

OncoGPT training requires a dataset that consists of treatment records comprised of sequences of therapeutic controls and corresponding responses. Treatment response is represented by a discrete tumor state defined based on tumor burden, RECIST assessment and eco-evolutionary variables. Likewise, therapeutic controls are discretized into finite sets of distinct therapeutic actions, each defined by the drug name or therapy (eg, radiation), drug/therapy dose and the interval of time before applying the next therapeutic action. Therapeutic doses would be represented by their values normalized based on the body surface area, and quantized along a finite number of possible dose levels. Likewise, the delay time allowed before the next therapeutic action would also be quantized to take an integer number of base periods of time. The latter may be set to the minimum onset time across all cancer therapies being used. The discretization of therapeutic controls and treatment responses enables the training of OncoGPT using sequences of controls and disease states, respectively. In the previous section, the representation of treatment response was estimated to require a vocabulary of 12 500 “words”, ie, possible distinct disease states (treatment responses). Considering an estimated 280 cancer drugs^87,88 that were FDA approved by 2022, and assuming 10 quantization levels for both doses and the time interval before the administration of the next dose, there would be 28 000 possible distinct controls. The sizes of these control and response vocabularies are well below the size of the English vocabulary in common use, ⁸⁹ suggesting that the size of treatment data that would be required to train OncoGPT would be below the 3 trillion tokens used to train the French-English model CroissantLLM. ⁹⁰ In particular, given that one instance of therapy administration (ie, control) and the corresponding instance of treatment monitoring (ie, disease state) would be 2 tokens, the daily monitoring of a hypothetical 6-month long treatment would be about 360 (2 x180) control-response data points, or 360 tokens, requiring hence the monitoring of more than 8 billion patients/treatments to curate a training dataset in the 3 trillion size. Collecting, annotating and curating a quality clinical dataset of this size is clearly challenging. However, simulation data generated from clinically validated mathematical models of tumors under treatment may be used to augment accumulating clinical data being collected and curate the necessary dataset to train OncoGPT for treatment response prediction.

The notion of learning the causal relationship between treatment and treatment response using OncoGPT is aligned with the vision of rapid-learning cancer care systems,^91,92 whereby rapidly accumulating data, be it molecular, radiomics or clinical, about cancer and its treatment could be leveraged to inform treatment decision-making. Records of treatment cases are usually maintained in health information systems (HIS) of clinical and research institutions. The collection of these records and the results of clinical trials represent a treatment knowledge base that may be streamlined for use by the oncology community using learning platforms.^93,94 Although monitoring cancer progression using regularly sampled radiogenomics and LB data is clinically feasible, the accuracy and reliability of corresponding treatment response assessments are still in need of greater advancement. In particular, treatment response biomarkers are currently limited in reflecting intra-tumor heterogeneity, limiting the accuracy of collected disease state data. Furthermore, collection of clinical data is often not timely or performed with sufficient frequency to yield high resolution and relevant observations of disease progression. In order to mitigate the effect of these limitations and account not only for cancer inherent diversity but also the fact that a tumor is composed of spatially distinct regions where each may be undergoing different evolutionary trajectory under therapy, it is more appropriate to consider OncoGPT models that are dedicated to each patient sub-population, identified through patient stratification using clinically validated biomarkers. Furthermore, the unique evolutionary trajectory of each tumor makes it necessary to further fine-tune these sub-population OncoGPT models into patient-personalized OncoGPT treatment response models. This three-step training of OncoGPTs is akin to the three-step fine-tuning approach used for ChatGPT 3.5. ⁹⁵ However, different subsets of the training dataset would be used for training, fine-tuning and personalization, respectively. In the first step, OncoGPT would be trained using data for all cancer types combined, followed by tuning using data specific to various cancer types, leading to multiple models each dedicated to one cancer type. Finally, each one of these models would be further personalized for a narrow class of patients with high phenotypic similarity (see Figure 3). Intuitively, the first training step addresses the similarity of cancer growth dynamics across all cancer types. In the second training step, the pre-trained model is tuned into instances dedicated to distinct cancer types by restricting the training data to that of patients with the same cancer type. However, inter-patient and intra-tumor heterogeneity combined with the evolutionary dynamics of tumor growth under therapy represent intrinsic challenges to the capability of these models to yield accurate predictions of treatment response, motivating the need for a third fine-tuning step to personalize these models to classes of patients that are similar with respect to clinical, molecular and radiomic features. This personalization step of the training process would yield OncoGPT models that are fine-tuned for narrow classes of phenotypically similar patients.

Figure 3.

Three-step OncoGPT training. In the first step, OncoGPT is trained as a general model applicable to all cancer types. In a second step, further tuning is applied to yield distinct models for each cancer type. Finally, additional fine-tuning is applied to obtain distinct models that are personalized to narrow classes of phenotypically similar patients.

The appropriate training data subset for the personalization step may be identified through the classification of patient records using metrics of phenotypic similarity. ⁹⁶ Taking the example of Lung cancer as the current leading cause of cancer deaths, the classification would be based on factors such as the alteration status of EGFR, MET, ALK, HER2, ROS1, KRAS, BRAF and RET, which harbor actionable mutations. ⁹⁷ Other factors would include PD-L1 expression, tumor mutational burden (TMB), aneuploidy, TILs (Tumor infiltrating lymphocytes) abundance, cancer-immune set point, ⁶⁷ and immune evasion capacity, ⁹⁸ HLA (human leukocyte antigens) loss of heterozygosity (LOH) and antigen-processing defects. The alteration status of KRAS, TP53 and STK11/LKB1 may also be used to inform about T-cell exclusion or inflammation of the TME, and resistance to immune PD-1 inhibitors, ⁹⁹ as well as provide a correlate for PD-L1 expression.^99,100 The profile of lung and gut microbiota would equally be essential in defining patient phenotype given its role in carcinogenesis^101,102 and influence on therapy efficacy.^103,104

Given the multiplicity of causal dimensions (eg, genomic, metabolic, immunologic and eco-evolutionary) underlying cancer response to therapy, patient multimodal data need to be integrated into phenotype models that reflect disease state dynamics. Given that cell signaling pathways and their dynamic coupling with metabolism underlie cancer pathophysiology and the progression trajectory of cancer cell phenotypes, ¹⁰⁵ then as the regulators of cell functions (ie, growth, death, proliferation, survival, metabolism, etc.) and the ultimate objects of therapeutic interventions, cell signaling pathways represent the most relevant networks to integrate patient data into phenotype models for patient classification in the training dataset, in a fashion similar to the proposed use of deep phenotyping to stratify patients for personalized care. ¹⁰⁶

Discussion

The capacity of multi-layer neural networks to approximate nonlinear functions with an arbitrary precision^82-85 provides a strong rationale for the expectation that GPTs, which are built using DNNs, can learn a model of the nonlinear cancer treatment-response mapping. The proposed GenAI model for treatment response prediction assumes that OncoGPT sees controls and treatment responses as tokens just as GPTs see words in natural language processing (NLP). Indeed, like words, controls and responses may be considered as discrete entities or tokens from vocabularies of finite sizes. Ultimately, it is their numerical vector representations (embeddings) that are processed by the transformer. These embeddings give meaning to the tokens through training using data about patients and their treatments. For NLP tasks such as translation, ChatGPT constructs a semantic space where words are placed in accordance with their semantic similarity ¹⁰⁷ which is also the basis of its semantic activation. ¹⁰⁸ Repurposing LLMs to predict treatment response raises the question about the space being constructed by OncoGPT when trained using cancer treatment data. It is also important to recall that the success of ChatGPT and other LLMs in completing the myriad of linguistic, cognitive and artistic tasks are only supported by empirical evidence and that there are no scientific theories that predict their unexpected and surprising performance. Hence, it is essential to explore the question at hand in an effort to cultivate the necessary confidence in the use of OncoGPT as a predictor of cancer treatment response. First, the inputs to the encoder part of OncoGPT are sequences or time series of controls, while the inputs to its decoder are time series of treatment responses (disease states). Given a treatment cycle made up of a sequence of controls, each control in the sequence would be represented by its embeddings, ie, a vector of numerical values that represents its “essence” as well as its position in the treatment cycle. The embeddings will ultimately be multiplied by learnable weights in the computation of self-attention, which is a key component of the transformer. Second, self-attention quantifies the extent to which each control relates to each one of the other controls in the treatment cycle, in addition to capturing information about the essence of the control and its position in the treatment cycle. The self-attention mechanism is also applied to the sequences of treatment responses. However, in this case it is referred to as a masked self-attention in order to enforce model causality whereby its computation for a given treatment response should only depend on the previous treatment responses in the sequence. As a result, the masked self-attention quantifies the extent to which each treatment response relates to past treatment responses within a specified historic window of the disease state trajectory, as well as holds information about its essence and its position in such trajectory. Third, the “encoder-decoder” layers of OncoGPT are also self-attention blocks, however in this case the objective is for treatment responses to attend to the controls, hence providing a quantification of the extent to which controls relate to responses. For training, a loss function that is defined based on the errors between predictions and observations of treatment responses is back propagated to tune the weights of OncoGPT. Through the use of attention applied to treatment data, as succinctly explained, OncoGPT learns a model that captures an associative map between controls and treatment responses with the consideration of the contexts of treatment cycle and past treatment responses. It may therefore be plausible to hypothesize that OncoGPT inferencing would be operating in a control-response space where predictions of treatment responses would be more similar for phenotypically similar patients subjected to the same treatments. In this context, phenotypic similarity may be understood as the similarity between control-response trajectories, which may be defined in the control-response space using one of the many possible trajectory similarity measures. ¹⁰⁹ Provided that further research supports the notion that OncoGPT constructs a phenotypic space where control-response pairs for different patients are placed in accordance with patient phenotypic similarity, this would suggest that OncoGPT would yield similar predictions of treatment response trajectories for phenotypically similar patients subject to the same treatment. Such property would be an essential ingredient towards the clinical validation of the model with respect to reliability and explainability.

Notwithstanding the domain-specific target application of OncoGPT, its performance needs to be assessed in light of the limitations of the transformer’s architecture, including the hallucination problem. ¹¹⁰ In the case of OncoGPT, hallucination would consist in yielding a prediction of treatment response that is incompatible with the treatment-response training data and the inputs, ie, controls or therapeutic actions. The risk of hallucination may be mitigated through multiple means involving the training of OncoGPT and the way its output is used in the adaptive control loop. First, the training of OncoGPT on treatment-response data is specifically dedicated to the task of predicting treatment responses. Such in-domain inferencing has been empirically shown to be almost perfect for GPTs, provided that the tasks are of “low compositional complexity”. ¹¹¹ This may be the case for the task of inferencing the next treatment response which may be completed through pattern matching between new input treatment sequences and the treatment sequences in the training dataset. Second, the one-step ahead prediction, for which OncoGPT is used, consists in estimating the next treatment response given a sequence of therapeutic actions and the actual past treatment responses. In other words, highly erroneous predictions are not fed back to OncoGPT, which would otherwise increase the chance of such error to propagate and get amplified into hallucination. Furthermore, additional safeguards may be applied to assert the acceptability of OncoGPT predictions by monitoring the actual prediction error over time. For example, when a drastic or unusual change of the prediction error is observed, independently monitored biomarkers or clinical variables may be analyzed for clues that would either corroborate or refute the plausibility of the next prediction. If the plausibility of such prediction is refuted, the synthesis of the therapeutic action should ignore the prediction and instead rely on the most recently monitored treatment response as a safe alternative. Finally, in the proposed model of GenAI-supported adaptive therapy, treatment recommendations, ie, the controls, are synthesized by the adaptive controller and not the OncoGPT. The control law and output of the adaptive controller would be subject to the oncologist’s criteria of acceptability, asserting the primacy of the oncologist as the final arbiter in the treatment decision-making process.

The feasibility of treatment response prediction using OncoGPT is heavily premised on the availability of sufficiently big, annotated quality training datasets. The curation of large datasets from the monitoring of disease treatment response of a large population of patients presents a significant challenge to realizing the potential of GenAI in oncology. The implementation, clinical validation and deployment of OncoGPT would also be a complex undertaking fraught with numerous challenges that are intrinsic to data-driven AI, including explainability and ethical concerns.^36,112 Ongoing efforts to overcome data-related challenges in cancer reasearch ¹¹³ and the continuous evolution of the regulatory environment to facilitate a safe, ethical and effective deployment of AI models in the clinic, ³⁹ combined with research advances in AI applications for oncology^32,34 and the exploration of effective paths to their implementations and deployments for patient care^114,115 are providing a dynamic environment for addressing the challenges inherent to data-driven AI and are driving the maturation of the AI-assisted cancer care paradigm.

Conclusions

The adaptive evolutionary dynamics of cancer require an equally adaptive treatment to thwart the onset of therapeutic resistance and achieve disease management or cure. Repeated monitoring of disease progression combined with the prediction of treatment responses are critical components of an effective cancer treatment decision-making. The causal relationship between treatments and responses is represented by a nonlinear mapping between sequences of discrete therapeutic controls and sequences of resulting treatment responses. This formulation enables the conception of OncoGPT as a GenAI system proposed to learn a model of treatment-response mapping from patient data. The learned model would provide repeated predictions of treatment response to support treatment adaptation and optimization. Although the success of GenAI in natural language processing provides a robust rationale for the expected performance of OncoGPT in treatment response predictions, its implementation and deployment in oncology will have to address the challenges that are unique to the oncology clinical setting, including clinical validation, data quality and quantity, explainability and ethical concerns.

Ethical Statement

Ethical Approval

This manuscript is a review article and does not involve a research protocol requiring approval by the relevant institutional review board or ethics committee.

ORCID iD

Youcef Derbal https://orcid.org/0000-0003-3317-9417