cool hit counter The appeal of artificial intelligence algorithms: determining whether cancer patients can benefit from immunotherapy_Intefrankly

The appeal of artificial intelligence algorithms: determining whether cancer patients can benefit from immunotherapy

Immunotherapy continues to take over in a world where cancer rules.

Cancer patients see hope for life, but unfortunately it is hard for patients to know if they are actually the ones who are immunotherapy zapped.

Current clinical statistics show that PD-1 inhibitors only work in 20-50% of patients with advanced solid tumors [1]. Who exactly are the beneficiary patients? Current means of locating these patients are limited.

Therefore, finding a reliable biomarker to guide immunotherapy is imminent!

Recently, Dr. Eric Deutsch's team from France has given us new hope with the power of artificial intelligence. They trained the AI with CT images of cancer patients to get an AI platform that can accurately predict the effect of PD-1 inhibitor treatment from the patient's CT images.

Dr. Eric Deutsch is first from the left

This AI platform differentiates the extent to which patients respond to immunotherapy, and the median survival of those deemed effective (24.3 months) more than doubles (by more than a year) the median survival of those predicted to be ineffective (11.5 months), which is quite significant! The results were published in the latest issue of The Lancet Oncology [2].

You may be wondering, isn't immunotherapy at the molecular level? And CT imaging is a macroscopic level visible to the naked eye, so how do the two fit together?

That's the beauty of artificial intelligence, so let Singularity Cake tell you about it.

First, let's look at exactly what kind of tumors immunotherapy will work on.

When it comes to immunotherapy, we all know that it attacks tumors with the help of our own immune system. That tumor has to have plenty of immune cells in or near it, right.

It has been previously reported that the effectiveness of immunotherapy is related to whether the tumor is infiltrated by immune cells [ 3, 4, 5, 6]. These tumors tend to respond to immunotherapy if there is an abundance of CD8 cells (also called cytotoxic T cells) in the tumor tissue and if the tumor cells also express checkpoint markers such as PD-L1 in large numbers, as well as a large number of genetic mutations [7, 8, 9].

So yeah, there are ways now to analyze PD-L1 expression levels in tumor tissue, for example, or to analyze the mutational load of tumor tissue. It is possible to predict to some extent the extent to which patients will respond to immunotherapy, but current research suggests that these two tools alone are ah not enough.

And one of the more obvious shortcomings of both prediction methods is that both require a tumor tissue biopsy.

Dr. Deutsch came up with the idea of CT imaging.

High-dimensional medical imaging is now able to visualize tumors at a macroscopic level. But the scientists have more ambitions than that, they want to use these eyes to see directly into the tumor tissue at the cellular and molecular level to completely identify the nature of the tumor.

That is, Dr. Deutsch's team wanted to see the level of immune cells inside the tumor directly from CT images of the patient's tumor tissue, and then predict how well the patient would respond to immunotherapy.

So Dr. Deutsch had a bold vision - to use medical imaging to guide immunotherapy with the help of artificial intelligence.

When all was said and done, the scientists quickly put the idea into action.

They selected a cohort called MOSCATO [10], a cohort of 135 patients with advanced solid tumors, and their CT images as well as transcriptome sequencing data of the tumor tissue were preserved. Among other things, these transcriptome sequencing data can be used to calculate the number of CD8 cells in the tumor.

The data from 135 patients and transcriptomic data is a huge amount of information, which is impossible to analyze manually.

So, the researchers gave a stream of this data to a machine learning platform and let the machine find the pattern on its own. Without disappointing the researchers, machine learning helped them find the pattern and develop an algorithm that could predict the number of immune cells in tumor tissue.

Artificial Intelligence Training Process

But whether this algorithm is reliable needs to be further verified.

The researchers used another cohort called TCGA [11, 12], including 119 patients, all of whom also had CT images and transcriptome sequencing data. The number of CD8 cells in this cohort calculated using the algorithm agrees with the number calculated from the corresponding transcriptome data, indicating that the algorithm is reliable.

However, this only shows that the immune cell information read by this algorithm from CT images can be consistent with the tumor cell information of the tumor tissue transcriptome response.

Whether it is accurate or not will have to be tested by practice.

So, the researchers found a third cohort. There were 100 patients in this cohort, and this cohort was unique in that all patients' tumor tissue was accurately typed. The implication is that the immune cell profile of the tumor tissue in these patients is known.

1 cohort training algorithm, 3 cohort validation

The tumor tissue of these 100 patients was divided into 3 types: immune infiltrative, immune rejection, and immune desert [13]. Tumors in the immune-infiltrating type are heavily infiltrated by immune cells; the immune-rejecting type prevents T-cell infiltration; and tumors in the immunodesert type have little or no T-cell infiltration.

The analysis showed that the images were scored by the algorithm and the immunophenotypes predicted by the scores could correspond well to the known results. This hurdle is also passed.

Since we can determine the immune phenotype, can this phenotype predict the effectiveness of immunotherapy? The algorithm has to go through the ultimate ultimate test.

The final test cohort included 137 patients, all of whom had received immunotherapy and had undergone follow-up after treatment, with a median follow-up value of 16.5 months [14, 15].

A study analyzing follow-up records using a time-tested artificial intelligence platform found that in the first three months of treatment, patients with high scores (23%) responded to treatment, while those with low scores (77%) did not, although the difference was not significant. But in the sixth month, that gap became very apparent.

More importantly, patients with high scores had significantly higher median survival (24.3 months) than those with low scores (11.5 months), with a median survival improvement of more than 1 year, a very significant effect, and the ultimate test passed!

Median survival of patients treated with immunotherapy

So far, we can say that this imaging signal, obtained using artificial intelligence, brilliantly predicts the effectiveness of immunotherapy!

Of course, to really move towards practical application, clinical trials will need to be conducted on the basis of this retrospective trial. In fact, that day won't be long in coming, as 27 trials using imaging data to guide clinical oncology treatment have been registered to date (

Moreover, there are some limitations to this study, such as the existence of additional immune subtypes of tumors [16], and imaging signals need to be more finely differentiated in order to make more accurate guidance for immunotherapy.

Nonetheless, we are still excited about AI-enabled imaging histology. Compared to tissue biopsies, CT scans are non-invasive and not harmful to the body, making it an undoubtedly better option for those patients who are unsuitable or unwilling to undergo a tissue biopsy. In addition, the fact that imaging is also cheaper relative to sequencing is a not insignificant advantage.

The increasing use of artificial intelligence in disease treatment is believed to overcome more difficult diseases in the future. And now, a new era of cancer immunotherapy testing is coming!

1、GIF fighting device
2、Chen Gang Chinas population will urbanize faster compared to the world
3、Iterator for python classes an article aimed at the n00b
4、test machinery etc
5、The King Next Door and the Blockchain POS and DPOS

    已推荐到看一看 和朋友分享想法
    最多200字,当前共 发送