The response rate to first-line EGFR-TKIs approaches 70% which is much higher than chemotherapy. However, it still means that almost 30% of patients cannot benefit from EGFR-TKIs. It has clinical implications to predict the efficacy and screening the less effective patients. There are still many challenges developing a practical clinical tool to assess the efficacy of EGFR-TKIs. Therapeutic resistance is a complex and multi-factorial participating process. At the present stage, models based on imaging have demonstrated the ability to predict the efficacy of cancer treatment[14]. These models often lack generalization and are difficult to transduce. So, we attempted to develop models based on clinical characteristics, genotypes and artificial intelligence.
Many clinical characteristics are predictive for EGFR-TKIs. Larger tumors are always thought to be associated with poor prognosis[15]. Patients with concomitant liver metastases, brain metastases and uncommon mutations are also reported to have bad outcomes[16]. Clinical characteristics could determine a degree of the efficacy for EGFR-TKIs but not sufficient.
Co-mutations played an important role in the drug resistant of EGFR-TKIs[17]. TP53 was the most frequent and impactful concurrent mutation in lung cancer with EGFR mutation. The co-mutation rate of TP53 ranged from 54.6–64.6%[18]. The co-mutation of TP53 influenced the natural history of patients with EGFR-mutant and allowed for the diversification of subclone. It prompted the therapeutic resistance. A large number of studies also identified that TP53 can be a negative prognostic marker for the outcomes following EGFR-TKIs[19–21]. As an important early genetic event of EGFR-mutant LUADs, the inactivation of RB1 often harbors TP53 co-alterations and controls the cell cycle with mutation rate of approximately 10%[18, 22]. TP53 and RB1 co-mutations in EGFR-mutant LUAD also increase the risk
of small-cell transformation[23]. ATM alterations, IDH1 mutations and PTEN mutations were also reported to be associated with shorter PFS and OS in patients receiving first-line EGFR-TKIs[24]. LKB1/AMPK pathway was shown to reduce sensitivity of EGFR-TKIs in vitro[25]. After analyzing genomic changes, TP53, RB1 and PI3Ka were selected as the final model. Genotypes of patients were demonstrated to increase the stability and accuracy of models. NGS tests also increase the accessibility of models. With the constantly mature technology and gradually decreasing costs of NGS tests, genotypes of most NSCLC patients can be definitive prior to first-line treatment. In conclusion, Genotypes well assisted in predicting the efficacy for EGFR-TKIs.
In the last two decades, ML models have been widely applied in medical and health sciences[26]. It is evident that ML models have improved the understanding of cancer progression[27]. It is the first model to predict the efficacy for EGFR-TKIs based on ANN. ANN can handle classification problems satisfactorily and even serve as a gold standard in some tasks. Hidden layers represent the neural connections in mathematical process (Fig. 1C). However, it is also an important drawback it suffered from. It is difficult to explain the classification process as a “black-box” technology. Additionally, the layered structure could be time-consuming and lead to poor performance for some models. So, epoch of convergence is also an important parameter for our model. The model based on dataset2 showed best performance in this regard.
The risk of progression with different generation EGFR-TKIs is an important research direction. The efficacy of first-line osimertinib is much better than previous EGFR-TKIs[28]. In FLAURA study, the median PFS of osimertinib was up to 18.9 months which was prominently longer than gefitinib or erlotinib. Given the excellent efficacy of osimertinib, the number of untreated patients with osimertinib who have completed the follow-up are limited. We decided to exclude patients with osimertinib from our model. With the large-scale use of Osimertinib in untreated EGFR-mutated NSCLC, the comparison between first-generation EGFR-TKIs or second-generation EGFR-TKIs and third-generation EGFR-TKIs will also be included in next improved model after completing the follow-up. Further work will be dedicated to addressing the selection of more specific individualized treatment including the combinations with anti-angiogenic therapies or chemotherapies and the difference between first-generation EGFR-TKIs or second-generation EGFR-TKIs and third-generation EGFR-TKIs.
There are still some limits in our study. Only first-generation EGFR-TKIs and second-generation EGFR-TKIs are involved into the model. Our new model needs validation for more third-generation EGFR-TKIs. The sample size should also be enlarged in the future to make the model more stable.