The combination of radiation therapy and immunotherapy presents a dynamic field with ongoing research to optimize treatment strategies and enhance patient outcomes. Despite limitations in available preclinical data, we have leveraged our AI-based approach to maximize the potential for accurate modeling and prediction. Using the LSTM-RNN model, we successfully demonstrated that: 1) The LSTM-RNN model can effectively simulate the process of tumor growth and control in a preclinical model, considering the combined therapy and its dosage over a period of two weeks or longer; 2) The LSTM-RNN model can integrate various parameters, including pulse interval, radiation dose, drug dose, and timing, to predict more effective combinations. This AI-based predictive model has the potential to help guide the design of dynamic PULSAR trials, offering increased flexibility while reducing costs. It is also able to provide insights into the underlying biological processes in different cancer cell models. Additionally, the model can support in-silico studies, generating essential data to address other relevant research questions in the field of radio-immunotherapy. We are optimistic to continue to improve the model by addressing the limitations in future studies.
AI modeling and interpretation. The LSTM-RNN model has demonstrated its ability to effectively model the process of tumor growth and control in combined therapy over a period of two or three weeks. Each group in our study corresponds to a specific input configuration, consisting of radiation dose and α-PD-L1 drug dose. It is worth noting that the inclusion of fifty samples per group serves to introduce noise, thereby mitigating the risk of overfitting and improving the model’s generalization capacity. The training phase of our study validated the presence of a direct one-to-one relationship between the input parameters and the corresponding mean output for each group, although these results were not included in the present analysis. In essence, the role of the LSTM-RNN model was to ascertain the mapping between inputs and outputs for all 26 groups. This task can be likened to curve fitting, yet it is a much more formidable task, considering factors such as the number of time steps, as well as the intricate and unknown mechanisms of interaction involved in the dataset.27–29
The LSTM-RNN model capitalizes on its inherent ability to process sequential information, which greatly aids us in achieving our objective. An additional challenge we encountered was the presence of many zeros in the input sequence, representing days without either radiation or anti-PD-L1 treatment, as well as missing data in the output sequence due to unmeasured days. We conducted comparative tests and found that a feed forward network (FNN) performed less effectively than the LSTM-RNN model. The key advantage of the LSTM-RNN model lies in its recurrent structure, where the output from the previous time step is fed as input to the subsequent time step. This property makes the LSTM-RNN model well-suited for predicting the next tumor volume based on the knowledge of previous volumes. Initially, we attempted to employ a one-to-one structure within the LSTM-RNN model, aiming to predict responses for all time points, including those with missing measurements. However, the results obtained from this approach were inferior compared to the many-to-one structure employed in this manuscript. We attribute this performance discrepancy to the fact that when a sequence contains numerous missing data points, establishing a consistent temporal correlation becomes a more challenging task for the AI model.
By examining the patterns depicted in Figs. 3 and 4, we can gain insight into the level of complexity arising from the interplay of numerous physical and biological processes, disregarding the considerable variance observed within each group. The hidden states of the LSTM-RNN model capture historical information from the sequence up to the current time step. This recurrent computation property is advantageous as it prevents the growth of the number of model parameters as the number of time steps increases, which is particularly beneficial when dealing with small datasets. We believe that any further speculation beyond what is presented in this manuscript exceeds the capability of the AI modeling, given the limitations of the available experimental data. Another noteworthy aspect of our approach is the utilization of hidden unit values to identify temporal correlations between the respective responses associated with radiation or immunotherapy. This aspect allows us to gain insights into the “causal relationship” in a quantitative or semi-quantitative manner, further enhancing our understanding of the underlying dynamics when combining PULSAR with immunotherapy.
The control of tumor growth is influenced by various factors, including the dose per pulse, pulse structure, α-PD-L1 administration, and relative timing. The significance of relative timing is clearly demonstrated by the predictive capability of our model, as depicted in Fig. 5c. For the LLC tumor model, the most effective tumor control is achieved when a single pulse of 40 Gy dose is administered, accompanied by four doses of PD-L1 given after a six- or seven-day interval. These findings offer valuable insights into determining the optimal timing sequence for a given tumor model, characterized as either "hot" or "cold", in terms of the sequencing of radiation pulses and administration of α-PD-L1. The timing between radiation and immunotherapy plays a critical role in achieving synergistic effects at different time points. If the spacing between treatments is too short, radiation may eliminate newly recruited T cells entering the tumor microenvironment from lymph nodes. Conversely, if the spacing is too long, the maximum therapeutic benefit from immunotherapy may not be fully realized. While our study has provided valuable insights, we acknowledge the limitations of the current animal experiments. Therefore, we have not exhaustively explored all possible sequences to identify the optimal delivery conditions that yield the maximum AVC. This task, along with further experimental validation, will be pursued in the next phase of our study.
Synergies between radiotherapy and immunotherapy. It is widely acknowledged that α-PD-L1 provides additional benefits when administered concurrently with radiation.30–34 However, the dynamic interplay between these two treatment modalities over time has received limited attention. Figures 3 and 4 indicate the existence of an equilibrium period between cancer cell proliferation and the cytotoxic activity of CD8 + T cells. This equilibrium phase typically occurs two or three days after the initial radiation pulse and is most pronounced following a 20 Gy dose, lasting approximately 10 days (as seen in the plateau regions of Figs. 3c, 3d, 4a, and 4b). Subsequent alterations in either PD-L1 or PD-1 levels can tilt this equilibrium towards more effective tumor control. Furthermore, we speculate that repeated radiation may longitudinally amplify the regulatory effects of Treg cells, suggesting that α-PD-L1 administration should be correspondingly enhanced. A study has reported that intra-tumoral T cells, especially Treg cells, can mediate tumor control without the influx of newly infiltrating T cells and exhibit greater resistance to radiation compared to circulating or lymphoid tissue T cells, similar to memory T cells.31 This was achieved by selectively eliminating circulating/peripheral T cells through a dose of 8 Gy whole-body irradiation (WBI), while protecting the tumors in the window chambers using lead shielding to preserve the intratumoral T cells. As the proportion of surviving T cells increases and more interferon gamma is produced through repetitive radiation, the impact of PD-L1 antibody may be amplified. However, due to the rapid tumor growth and the short duration of our study (28 days in Fig. 2b), this speculation has yet to be validated.
Another aspect of particular interest is the timing of radiation pulses. It is important to consider that when the first radiation pulse recruits circulating CD8 + T cells from the blood into the tumor microenvironment, these newly arrived CD8 + T cells should not coincide with the second radiation pulse. Additionally, having the two pulses too close together may not be optimal as it can stimulate tumor growth. For example, a study using similar preclinical models demonstrated that tumor growth was accelerated and survival decreased when subsequent 3-Gy doses were administered daily following ablative radiation therapy.32, 33 This study also mechanistically showed that repeated daily doses of 3 Gy abrogated CD8 + T-cell infiltration into tumors, which is an important predictor of immunotherapy response in humans.
Additionally, the infiltration of CD8 + T cells into the tumor microenvironment also plays a critical role in the context of PULSAR (Fig. 1c). In "hot" tumors, there is a robust immune response characterized by significant immune cell infiltration and activation of immune pathways. Conversely, "cold" tumors exhibit limited immune response, with minimal immune cell infiltration and reduced immune activity. Our previous study already found that the "cold" LLC tumor model responded best to radiation doses spaced ten days apart, with only the second PD-L1 dose required for maximizing therapeutic effect. On the other hand, in "hot" tumor models such as mouse colon carcinoma (MC38), a single radiation dose is optimal for synergy with immunotherapy, as initial priming of the immune response has already occurred due to preexisting tumor immunity.4, 34 In immune-resistant tumors, similar to the LLC murine model and many human tumors, an initial priming dose of radiation is necessary, and its immune-stimulatory effects manifest after several days. The investigation of how T cell infiltration is temporally influenced by radiation pulses in terms of dose and spacing remains an open question. Understanding this behavior will have significant implications for treatment strategies, allowing PULSAR to effectively target the specific immune characteristics of different tumor types.
Effects of radiation pulses on T reg . We propose that the regulation of Treg cells is the most likely explanation for the observed temporal response in our study. Treg cells, known for their role in regulating anti-tumor immunity both inside and outside the tumor microenvironment, are believed to be a major contributor to immune escape mechanisms in cancer. Following ablative radiotherapy, a rapid and robust cytotoxic CD8 + T-cell response is initiated (Fig. 1a). To prevent autoimmune reactions, Treg cells are recruited to the tumor site and secrete immune-suppressive mediators such as CTLA4 for antigen-presenting cells and macrophages, as well as cytokines IL10 and TGF-β for CD8 + T cells. Simultaneously, some Treg cells infiltrate the tumor bed and release type-I interferon gamma, which leads to the upregulation of PD-L1 on the surface of tumor cells. Among the various subtypes of Treg cells, naturally occurring CD4 + CD25hiFoxp3 + Treg cells are considered the most relevant for tumor immune evasion.35 Previous in-vitro studies have demonstrated the combined effect of Treg regulation and radiation therapy, which aligns with our findings. In one study, the administration of anti-CD25 antibody in a murine prostate tumor model significantly enhanced the efficacy of radiation, resulting in delayed tumor growth.36 In a murine mesothelioma model, CTLA4 blockade using anti-CTLA-4 antibody reduced the proportion of Treg cells relative to effector T cells after radiation, thereby boosting the activation of CD8 + T cells.37 Additionally, the combination of local radiation with CTLA-4 blockade inhibited metastasis in a mouse model of breast cancer.38
The impact of radiation on the production of Treg cells remains a topic of debate, and our AI model may help provide further insights to obtain a definitive answer. For example, one study suggested that low-dose total body irradiation led to a selective decrease in the percentage and absolute count of Treg cells using C57BL/6 mice and 1.25 Gy gamma irradiation.39 However, our results seem to support the opposite trend, indicating that radiation stimulates the production of Treg cells. Our findings are also consistent with another study in which mice were irradiated with 10 Gy gamma radiation to the right leg (prostate C1 cells implantation).40 In this study, Treg cells significantly increased in the spleen, lymph nodes, blood, and lung within two days after exposure, returning to normal levels by ten days. The authors also reported three dose-dependent patterns: 1) Treg cells increase even after a low dose of 2 Gy, 2) mice receiving over 10 Gy irradiation exhibited double the fraction of splenic CD4 + Treg cells compared to the control group, although there was no significant difference between 10 Gy and 20 Gy, 3) Fractionating a larger dose into smaller daily fractions (e.g., 10 Gy vs. 5×2 Gy, 20 Gy vs. 3×8 Gy) slightly reduced the production of Treg cells. It is noteworthy that fractionation optimization is another task that can be easily performed using our proposed AI model. In addition to the fixed interval of ten days chosen in Fig. 5c, we can explore more combinations of dose and fraction numbers for PULSAR treatment. Daily fractions are very likely not to be the optimal choice.
Limitations. The large error bars in Figs. 3 and 4, indicating high variance in volume measurements, can be attributed to the limited number of animals in each group. Our main objective was to identify overall trends and provide plausible interpretations of the interaction between radiation and immunotherapy using the AI model. Consequently, the AI model focused on the mean output of each group to detect differences in temporal response, without accounting for inter-animal variation. To reduce the error bars, we can increase the number of animals in each group, for example, from eight to sixteen. A broader question arises regarding the level of personalization required in combined PULSAR and immunotherapy. In this manuscript, only tumor volume change was measured. If additional biological features could be collected as a function of time, animal-specific predictions might become feasible. To be more specific, the definition of synergy is solely based on tumor size. There is a lack of biological correlatives including experimental data on immune cell infiltration and their activities. One solution would be conducting flow cytometry experiments to collect more information about the temporal behavior of T cells, CD8 + T cells and Treg. Along with the information of tumor volume change, these biomarkers will reveal valuable information about tumor microenvironment and the regulation of immune response. First, how long does it take for CD8 + T-cells to infiltrate the tumor? Second, what are the effects of radiation pulses on Treg? To ensure the reproducibility of tumor volume measurement, we plan to use both cone beam computed tomography (CBCT) and positron emission tomography (PET) imaging, O-(2-18F-fluoroethyl)-L-tyrosine ([18F]-FET) as the radiotracer to measure the biological tumor volume (BTV). This allows us to verify whether there is a moderate correlation between CBCT and PET measurement, and whether PET measurement shows a lower inter-animal variability compared to CBCT. Nevertheless, the challenge of accounting for animal-to-animal variation would persist.
In future studies, it is important to conduct experiments under more controlled irradiation settings, such as using a fine radiation beam, improved conformity, and minimizing leakage dose. More accurate contouring techniques for volume measurement should be employed. It is also necessary to test the LSTM-RNN framework using a more clinically relevant orthotopic tumor model. Finally, additional investigation is needed to determine how the findings in mouse models can translate to clinical settings. On one hand, mouse tumors grow significantly faster than human tumors and are generally more radio-resistant. On the other hand, PD-L1 antibodies have a longer half-life time in humans body (e.g. >30 days). Based on these two consideration, we speculate that in clinical settings, a longer interval between the radiation pulses will provide enhanced therapeutic benefits relative to preclinical models.
One point deserving clarification is the quantification of radiation dose. Biological equivalent dose (BED) is commonly used in CFRT and SBRT, which is calculated by considering factors such as the physical dose per fraction, the overall treatment time, the number of fractionations, as well as the radiation sensitivity of the specific tumor (see Eqs. 1 and 2). However, we do not think BED is relevant in our study, and therefore only the physical dose per pulse was reported. Given our main focus is about the synergy between radiotherapy and immunotherapy, immune equivalent dose (IED) would be a better choice than BED in combined therapy, a topic awaiting further investigation.