The Johns Hopkins Kimmel Comprehensive Cancer Center and its Bloomberg~Kimmel Institute for Cancer Immunotherapy reveal that they effectively trained a machine learning system to predict which melanoma patients would react to treatment and which would not.
DeepTCR, an open-source programme, was useful as a predictive clinical tool, but it also taught the researchers about the biochemical mechanisms behind patients’ immunotherapy responses.
The team published its study “Deep learning reveals predictive sequence concepts within immune repertoires to immunotherapy” in Science Advances.
“TCR sequencing has been utilised to define the immune response to malignancy.” “However, most analyses have been limited to quantitative metrics such as clonality, which do not make use of the complementarity-determining region 3 (CDR3) sequence,” the researchers write.
“We apply DeepTCR, a framework of deep learning algorithms, to uncover sequence ideas that predict immunotherapy response.”
We show that DeepTCR can predict response and utilise the model to deduce the antigenic specificities of the predictive signature as well as their distinct dynamics during therapy.
Nonresponse is linked with a high frequency of TCRs anticipated to recognise tumour-specific antigens, and these tumour-specific TCRs experience a greater degree of dynamic alterations on treatment in nonresponders vs responders.
“These findings are consistent with a biological model in which the hallmark of nonresponders is an accumulation of tumour-specific T cells that undergo turnover on therapy, possibly due to the nonresponders’ dysfunctional state of these T cells.”
Developing more advanced models using this technology
DeepTCR was created by Sidhom while he was an MD/PhD student at the Johns Hopkins University School of Medicine. Deep learning is used to detect patterns in enormous amounts of data. The data in this example are the amino acid sequences of T-cell receptors or TCRs. When a TCR is triggered by a protein from cancer, germs, or viruses, the T cell releases chemicals to attack the threat while also cloning to strengthen the response.
Unfortunately, some tumour cells evolve mechanisms to prevent T cells from responding, even when the TCRs are active. Checkpoint inhibitors, which are now used in immunotherapy, are proteins that block this capacity in malignancies, forcing T cells to react to malignancy. However, these medications only benefit a small percentage of patients.
Sidhom, who is now a resident, used materials from the CheckMate 038 clinical trial, which compared the efficacy of one immunotherapy drug (nivolumab) versus a combination of two (nivolumab and ipilimumab) for 43 patients with inoperable melanoma. Before and during treatment, biopsies of the tumours were taken, which contained an array of infiltrating T cells.
There were no significant differences in participants treated with a single treatment vs a two-drug combination in the CheckMate study. Some patients from each group reacted, while others did not.
Sidhom employed genetic sequencing to determine the kind and amount of TCRs in each biopsy to find the TCR repertoire around each tumour. He then entered the data into the DeepTCR software, telling it which data sets belonged to responders and which belonged to nonresponders. The programme then looked for patterns.
The researchers initially wondered if there were any variations in the TCR repertoires of immunotherapy responders and nonresponders prior to treatment. The algorithm-identified differences were equally predictive of patient response as established biomarkers—molecular properties of malignancies used to guide therapy. However, before the algorithm can be used to guide therapy in the clinic, the researchers must validate their findings in a broader patient group.