Deep learning using tumor HLA peptide mass spectrometry datasets improves neoantigen identification

題目：利用腫瘤白細(xì)胞抗原的質(zhì)譜數(shù)據(jù)集進(jìn)行深度學(xué)習(xí)來提高癌癥抗原的識(shí)別

作者及單位：

Brendan Bulik-Sullivan艘刚， Jennifer Busby偷崩， […]勋乾， Roman Yelensky

• Gritstone Oncology, Inc., Emeryville, California and Cambridge, Massachusetts, USA.

發(fā)表刊物及時(shí)間：

Nature Biotechnology Published: 17 December 2018

摘要

==Neoantigens==（新生抗原）, which are expressed on tumor cells, are one of the main targets of an effective antitumor T-cell response. Cancer immunotherapies to target neoantigens are of growing interest and are in ==early human trials==（人類臨床實(shí)驗(yàn)早期）, but methods to identify neoantigens either require invasive or difficult-to-obtain clinical specimens, require the screening of hundreds to thousands of synthetic peptides or tandem minigenes, or are only relevant to specific human leukocyte antigen (HLA) alleles. We apply deep learning to a large (N = 74 patients) HLA peptide and genomic dataset from various human tumors to create a computational model of antigen presentation for neoantigen prediction. We show that our model, named EDGE, increases the positive predictive value of HLA antigen prediction by up to ninefold. We apply EDGE to enable identification of neoantigens and ==neoantigen-reactive T cells== （新抗原反應(yīng)性T細(xì)胞 ） using routine clinical specimens and small numbers of synthetic peptides for most common HLA alleles. EDGE could enable an improved ability to develop neoantigen-targeted immunotherapies for cancer patients.

在腫瘤細(xì)胞上表達(dá)的新抗原是有效抗腫瘤 T 細(xì)胞應(yīng)答的主要靶點(diǎn)之一。 新抗原靶向癌癥免疫療法越來越受到人們的關(guān)注书幕， 并且應(yīng)用在人類臨床實(shí)驗(yàn)早期。 但是鑒定新抗原的方法需要侵入感染或難以獲得的臨床樣本， 需要篩選成百上千個(gè)合成肽或串聯(lián)小基因口糕，或者只能夠與特定的人類白細(xì)胞抗原（HLA）等位基因相關(guān)搞挣。我們將深度學(xué)習(xí)應(yīng)用于來自不同人類腫瘤的大型（N=74 個(gè)患者） HLA 肽和基因組數(shù)據(jù)集带迟，以創(chuàng)建用于新抗原預(yù)測(cè)的抗原呈現(xiàn)的計(jì)算模型。我們發(fā)現(xiàn)囱桨，我們的模型——EDGE仓犬，將 HLA 抗原預(yù)測(cè)的陽性預(yù)測(cè)值提高了 9 倍。 我們應(yīng)用 EDGE 鑒定新抗原和新抗原反應(yīng)性 T 細(xì)胞舍肠，使用常規(guī)臨床標(biāo)本和大多數(shù)常見 HLA 等位基因的少量合成肽搀继。 EDGE 可以提高癌癥患者開發(fā)新抗原靶向免疫療法的能力

==Neoantigen==能被免疫細(xì)胞所識(shí)別、能被免疫系統(tǒng)所攻擊的貌夕、由于癌細(xì)胞基因突變所導(dǎo)致的律歼、正常細(xì)胞所沒有的異常蛋白質(zhì)，就是我們所要討論的 Neoantigen啡专。

圖表選析：

image.png

Figure 1 : Tissue samples and data for model training.

Tissues of origin and numbers of tumor and normal samples used (image credit: Andrii Bezvershenko/Bigstock.com), with the key generated data types: HLA peptide sequences, HLA types and tissue transcriptome measurements. Frozen tissue specimens were pulverized and lysed, with lysis product subjected to HLA immunoprecipitation with antibody W6/32 and peptide sequencing, along with mRNA extraction and transcriptome sequencing. We obtained HLA types from exome or targeted sequencing of matched normal tissues and used the integrated dataset to train a deep learning model to predict HLA epitope presentation. MS, mass spectrometry; ==NGS, next-generation sequencing==; Comet, an open source MS/MS sequence database search tool; Percolator, semisupervised learning for peptide identification from shotgun proteomics datasets; OptiType, HLA typing from NGS data; STAR-RSEM, ==bioinformatics pipeline==（生物信息學(xué)流程） for estimating gene expression levels from RNA-seq data; MiSeq, NGS platform for ==targeted sequencing==（靶向測(cè)序）; HiSeq, NGS platform for high-throughput sequencing.

圖 1： 用于模型訓(xùn)練的組織樣本和數(shù)據(jù)险毁。腫瘤及正常樣本來源及數(shù)量 (圖片來源 :Andrii Bezvershenko/Bigstock.com)，關(guān)鍵生成數(shù)據(jù)類型:HLA 肽序列们童、HLA 類型及組織轉(zhuǎn)錄組測(cè)量畔况。 冷凍組織標(biāo)本粉碎、裂解慧库，裂解產(chǎn)物用抗體 w6/32 經(jīng) HLA 免疫沉淀跷跪，并進(jìn)行肽段測(cè)序， 以 及 mRNA 提取和轉(zhuǎn)錄組測(cè)序齐板。我們從外顯子組或匹配正常組織的靶向測(cè)序中獲得 HLA 類型吵瞻， 利用整合數(shù)據(jù)集訓(xùn)練深度學(xué)習(xí)模型預(yù)測(cè) HLA 表位表現(xiàn)葛菇。MS：質(zhì)譜； NGS:新一代測(cè)序:Comet： 一個(gè)開源的 MS/MS 序列數(shù)據(jù)庫搜索工具橡羞； Percolator： 基于散彈蛋白質(zhì)組學(xué)的多肽識(shí)別的半監(jiān)督學(xué)習(xí)數(shù)據(jù)集;OptiType,：從 NGS 數(shù)據(jù)獲取的 HLA 分型眯停； STAR-RSEM： 用于估計(jì) RNAseq Miseq 基因表達(dá)水平的生物信息學(xué)流程；MiSeq卿泽，靶向測(cè)序的NGS平臺(tái) HiSeq： NGS 高通量測(cè)序平臺(tái)莺债。

image.png

Figure 2: Overview of the tumor peptidomics dataset.

(a) Peptide counts at various q-value thresholds from the 74 tumor mass spectrometry peptidomics samples. (b) Length distribution of presented peptides (FDR < 0.1). (c) Proportion of presented peptides with predicted binding affinities below various thresholds from 1 to 1,000 nM. Each blue line represents one of the 13 training set samples for which MHCflurry 1.2.0 binding affinity predictions were available for all HLA class I alleles; the red line shows the median across 13 samples. The dashed vertical lines show the common 500 nM and 50 nM 'binder' and 'strong binder' thresholds, respectively. (d) Relationship between peptide presentation and the RNA expression of the source gene measured in TPM for peptide lengths between 8 and 11. For each peptide length k, all possible peptides from all genes were grouped into 20 bins by TPM of the source gene and the proportion of k-mer peptides in each TPM bin that were detected via mass spectrometry is shown. (e) Genes with average expression across the 69 training set samples within narrow, 0.1 log10(TPM)-wide windows around 5, 10 and 100 TPM were selected. For each gene, the average prevalence of presentation of peptides of lengths 8–11 (i.e., the proportion of all possible peptides from that gene detected by mass spectrometry) from that gene across all training samples was computed. A histogram of this per-gene prevalence for all genes within each window is shown.

圖2. 腫瘤多肽組數(shù)據(jù)集總覽。(a) 74個(gè)腫瘤質(zhì)譜多肽樣本q值低于閾值的計(jì)數(shù)签夭。 (b)所表達(dá)的多肽的長度分別齐邦。 （FDR<0.1） (c)在1~1000nM的不同閾值下， 預(yù)測(cè)結(jié)合親和力低于閾值的多肽所占 比例第租。每一條藍(lán)線代表13個(gè)訓(xùn)練集中的一個(gè)措拇，它們是由MHCflurry 1.2.0預(yù)測(cè)的所有HLA 等位基因的親和力預(yù)測(cè)值。紅線代表13個(gè)樣本的中值煌妈。 兩條虛豎線分別代表500nM 普通結(jié)合肽與50nM強(qiáng)力結(jié)合肽儡羔。 (d) 肽段長度在8-11之間，多肽表達(dá)和以TPM度 量的源基因RNA表達(dá)的關(guān)系璧诵。 對(duì)于每一個(gè)肽段長度k汰蜘，所有可能的、來自所有基因的多肽依據(jù)源基因RNA表達(dá)的TPM值被分為20組之宿，在不同TPM值的組別中k-mer多肽的比例通過質(zhì)譜分析來呈現(xiàn)族操。 (e)根據(jù)69個(gè)訓(xùn)練集樣本的基因表達(dá)均值， 我們選擇的基因值在5比被、 10和100TPM附近的0.1 log10(TPM)范圍色难。 對(duì)于每個(gè)基因，計(jì)算出該基因 在所有訓(xùn)練樣本中出現(xiàn)長度為8 - 11的多肽的平均流行率等缀。 （即質(zhì)譜法檢測(cè)到的該基因中所 有可能多肽的比例）枷莉。每個(gè)圖中顯示了每個(gè)基因流行度的直方圖。

image.png

Figure 3: Architecture and features of the model.

(a) The architecture of our neural network (NN), with the subcomponents of the network active in a single patient with six HLA alleles. Pr, probability. (b) The learned dependence of HLA presentation on each sequence position for peptides of lengths 8–11 for two common HLA alleles. See Supplementary Figure 3a, b, c for learned motifs for all alleles. (c) Observed (dark blue) values are the proportion all detected peptides in the test samples found at each peptide length. Predicted (light blue) values are the sum of probabilities of all proteome peptides of length k over the total sum of probabilities of all peptides of lengths 8–11 (i.e., the expected proportion of presented peptides of each length). (d) Observed (dark blue) values are the proportion all detected peptides in the test samples found from genes at each mRNA expression TPM level. Predicted (light blue) values are the sum of probabilities assigned to all proteome peptides at the TPM level over the total sum of probabilities of all peptides. (e) Test set prevalence of detected peptides binned by learned per-gene propensity of presentation (xaxis) and RNA expression (y-axis) of the source genes.

圖3. 模型的體系結(jié)構(gòu)及特征

(a) 我們神經(jīng)網(wǎng)絡(luò)（NN）的體系結(jié)構(gòu)尺迂，在含有六個(gè) HLA 等位基因的單個(gè)患者中笤妙， NN 子組件的活躍情況。Pr 表示概率噪裕。(b) 對(duì)于兩個(gè)常見的 HLA 等位基因蹲盘， 長度為 8-11 的肽的 HLA 呈現(xiàn)對(duì)每個(gè)序列位置的學(xué)習(xí)依賴性。所有等位基因的學(xué)習(xí)模塊見補(bǔ)充圖3a, b, c膳音。(c) 觀察值(深藍(lán)色)是在每個(gè)肽長度處發(fā)現(xiàn)的測(cè)試樣品中所有檢測(cè)到的肽的比例召衔。 預(yù)測(cè)值（淺藍(lán)色）是在總的所有長度為8-11的肽段中，長度為k的所有蛋白 質(zhì)組肽概率的總和（即每個(gè)長度的呈遞肽的期望比例）祭陷。 (d) 觀察值（深藍(lán)色）是測(cè)試樣品中所有檢測(cè)到的肽在每個(gè)mRNA表達(dá)TPM水平的比例苍凛。預(yù)測(cè)值（淺藍(lán)色）是在TPM水平上分配給所有蛋白質(zhì)組肽 的概率與所有肽的概率總和的總和趣席。 (e) 通過學(xué)習(xí)每個(gè)基因的呈遞偏好（x軸）和RNA表達(dá)（y- 軸）將檢測(cè)肽的數(shù)據(jù)分箱，得到的測(cè)試集普遍性毫深。

image.png

Figure 5: Identification of neoantigen-reactive T cells from patients with non-small-cell lung cancer.

(a) Detection of T-cell responses to neoantigen peptide pools. In vitro–expanded patient PBMCs were stimulated with controls or patient-specific neoantigen peptide pools in IFN-γ ELISpot. Data are presented as spot-forming units (SFU) per 105 plated cells with background (corresponding DMSO controls) subtracted. Background measurements are shown in Supplementary Figure 11. For each patient, predicted neoantigens were combined into two pools of ten peptides each according to model ranking and any sequence homologies (homologous peptides were separated into different pools). ==In vitro==（在體外） T cell responses in single wells (1-038-001, CU02, CU03 and 1-050-001) or duplicates (all other samples) against cognate peptide pools 1 and 2 are shown for patients 1-038-001, 1-050-001, 1-001-002, CU04, 1-024-001, 1-024-002 and CU05. For patients CU02 and CU03, cell numbers allowed testing against specific peptide pool 1 only. Patients with in vitro T cell responses are represented in shades of blue; those without are represented in shades of orange and red. (b) Detection of T-cell responses to individual neoantigen peptides. In vitro–expanded patient PBMCs were stimulated in IFN-γ ELISpot with controls or patient-specific individual neoantigen peptides. In vitro T cell responses against cognate peptides are shown for patients whose cells showed positive responses against peptide pools (shown in shades of blue in a), along with, where cell numbers permitted, ==deconvolution==（解卷積） to individual peptides. Patients 1-038-001 and 1-024-001: data are presented as spot forming units (SFU) per 105plated cells for one visit, with background (corresponding DMSO controls) subtracted. Patients 1-024-002 and CU04: data are shown as cumulative (added) SFU for responses from three visits (CU04) or two visits (1-024-002). See also Supplementary Figure 10b. (c) Representative example of ELISpot wells from patient CU04 from data shown in a and b. Data were confirmed in an independent culture repeat (Supplementary Fig. 10c).

圖 5：鑒定來自非小細(xì)胞肺癌患者的新抗原反應(yīng)性T細(xì)胞

（a） T 細(xì)胞對(duì)新抗原肽庫反應(yīng)的檢測(cè)吩坝。 體外擴(kuò)增的患者外周血單核細(xì)胞（PBMCs）被 IFN-γ ELISpot 中對(duì)照組或病人特定的新抗原 肽庫刺激毒姨。 數(shù)據(jù) 被展示為每 10 5 個(gè)鋪板細(xì)胞的斑點(diǎn)形成單位（SFU）哑蔫，其中減去 背景（相應(yīng)的 DMSO 對(duì)照）。 背景的測(cè)量結(jié)果在增補(bǔ)的圖 11 中顯示弧呐。對(duì)于每個(gè)病人闸迷， 根據(jù)模型排序和任意 序列同源性（同源的多肽被分成不同庫）預(yù)測(cè)的新抗原被合并成 2 個(gè)各包含 10 個(gè)多肽的庫。 在單孔（1-038-001俘枫， CU02腥沽， CU03 and 1-050-001） 或?qū)ν措膸斓闹貜?fù)（所有其他樣本） 的 體 外 T 細(xì) 胞 反 應(yīng) 中 ， 1 和 2 展 示 了 病 人 1-038-001,1-050-001,1-001-002 鸠蚪， CU04,1-024-001,1-02-002 和 CU05. 對(duì)于 CU02 和 CU03 患者今阳， 細(xì)胞數(shù)量只允許針對(duì)特定的肽 庫 1 進(jìn)行檢測(cè)。體外 T 細(xì)胞反應(yīng)患者用藍(lán)色圖形展示茅信；沒有 T 細(xì)胞反應(yīng)的則用橙色和紅色展 示盾舌。（b） T 細(xì)胞對(duì)單個(gè)新抗原肽的反應(yīng)的檢測(cè)。在 IFN-γELISpot 中蘸鲸， 體外擴(kuò)增的患者 PBMC 被對(duì)照或患者特異性個(gè)體腫瘤抗原肽刺激妖谴。體外，患者（其 細(xì)胞對(duì)肽庫有陽性反應(yīng)（在 a 中以藍(lán)色陰影顯示））的同源肽的 T 細(xì)胞應(yīng)答被展示酌摇，以及在允許細(xì)胞數(shù)量的情況下膝舅，對(duì)單個(gè)肽進(jìn) 行解卷積∫ざ啵患者 1-038-001 和 1-024-001：數(shù)據(jù)以一次訪問每 10⁵ 個(gè)層疊的細(xì)胞中的 SFU 并去掉背景（相應(yīng)的 DMSO 控 制） 展示仍稀。患者 1-024-002 和 CU04：數(shù)據(jù)顯示為三次訪問（CU04） 或兩次訪問（1-024-002） 中反應(yīng)累積 SFU埂息。在圖 10b 中也可以看到技潘。 （c）在 a 和 b 中展示的數(shù)據(jù)中，患者 CU04 的 ELISpot 井的代表例子耿芹。數(shù)據(jù)在一個(gè)獨(dú)立培養(yǎng)重復(fù)（圖 10c） 中被證實(shí)崭篡。

討論：

With the progress of cancer immunotherapy, identification of neoantigens and neoantigenrecognizing T cells has become a central challenge in assessing tumor responses2,33, examining tumor evolution34 and designing the next generation of personalized therapies3. Current neoantigen identification techniques are either time-consuming and laborious7,20 or insufficiently precise10,14,15,16. Although it has recently been demonstrated that neoantigen-recognizing T cells are a major component of TILs7,20,35 and circulate in the peripheral blood of cancer patients7, current methods for identifying neoantigen-reactive T cells have some combination of the following three limitations: they rely on difficult-to-obtain clinical specimens such as TILs20,21 or leukaphereses7, they require screening impractically large libraries of peptides19, or they rely on MHC multimers, which may practically be available for only a small number of MHC alleles.

隨著腫瘤免疫治療的進(jìn)展，新抗原和新抗原識(shí)別T細(xì)胞已成為評(píng)估腫瘤應(yīng)答吧秕、檢測(cè)腫瘤進(jìn)化及設(shè)計(jì)下一 代個(gè)性化治療的主要挑戰(zhàn)琉闪。目前的新抗原識(shí)別技術(shù)要么費(fèi)時(shí)費(fèi)力，要么耗費(fèi)人力砸彬，或不夠精確颠毙。盡管最近已經(jīng)證明斯入，新抗原識(shí)別的t細(xì)胞是 tumor-infiltrating lymphocytes(腫瘤浸潤淋巴細(xì)胞)的主要組成部 分，并在癌癥患者的外周血中循環(huán)蛀蜜，但目前識(shí)別新抗原反應(yīng)性t細(xì)胞的方法有以下三個(gè)限制：它們依賴于 難以獲得的臨床標(biāo)本刻两，如腫瘤浸潤淋巴細(xì)胞或白細(xì)胞，但它們需要篩選不實(shí)際的大型肽庫滴某“跄。或者，它們 依賴于MHC多個(gè)等位基因霎奢，而MHC等位基因?qū)嶋H上可能只提供給少數(shù)幾個(gè)MHC等位基因户誓。

Here we demonstrate that all of these challenges can be addressed by improving the specificity of HLA epitope prediction algorithms by training models on mass spectrometry data instead of in vitro HLA–peptide binding affinity data. We generated the largest dataset of tumor HLA peptides, and HLA types reported to date, to our knowledge, and used these data to train a deep learning model of HLA peptide presentation. Using held-out mass spectrometry test data and retrospective neoantigen immune monitoring data, we demonstrate that our model, EDGE, outperforms state-of-the-art predictors trained on binding affinity and early predictors based on mass spectrometry peptide data by up to an order of magnitude, and show that the full scope of the predictive improvement is only achievable with the combination of several key modeling techniques. Finally, we show that by prioritizing peptides with prediction, it is possible to reliably identify neoantigen-specific T cells using a clinically practical process that requires only limited volumes of patient peripheral blood, screening a few peptides per patient, and does not rely on MHC multimers.

在這里，我們證明幕侠，所有這些挑戰(zhàn)都可以通過提高Hla表位預(yù)測(cè)算法的特異性來解決帝美，方法是在質(zhì)譜數(shù) 據(jù)上訓(xùn)練模型，而不是在體外Hla-肽結(jié)合親和力數(shù)據(jù)上進(jìn)行訓(xùn)練晤硕。據(jù)我們所知悼潭，我們生成了最大的腫瘤 Hla肽數(shù)據(jù)集和迄今報(bào)道的Hla類型，并利用這些數(shù)據(jù)訓(xùn)練了Hla肽表達(dá)的深度學(xué)習(xí)模型舞箍。利用保留的質(zhì)譜 測(cè)試數(shù)據(jù)和回顧性新抗原免疫監(jiān)測(cè)數(shù)據(jù)舰褪，我們證明，我們的模型创译， EDGE抵知，在結(jié)合親和力和基于質(zhì)譜肽 數(shù)據(jù)的早期預(yù)測(cè)器方面訓(xùn)練的先進(jìn)水平優(yōu)于基于質(zhì)譜肽數(shù)據(jù)的預(yù)測(cè)器，并表明只有將幾種關(guān)鍵建模技術(shù) 結(jié)合起來软族，才能實(shí)現(xiàn)預(yù)測(cè)改進(jìn)的全部范圍刷喜。最后，我們表明立砸，通過對(duì)多肽進(jìn)行預(yù)測(cè)排序掖疮，可以使用一種 臨床實(shí)用的方法可靠地識(shí)別新抗原特異性T細(xì)胞，這一過程只需要有限的患者外周血颗祝，每名患者只需篩 選幾個(gè)多肽浊闪，而不依賴MHC多倍體。

Critically, this improved performance for neoantigen identification is achieved by training a predictor based on data acquired by standard data-dependent acquisition mass spectrometry, despite this technique having insufficient sensitivity to detect all neoantigens directly14. Although targeted mass spectrometry approaches36 may ultimately improve sensitivity for direct neoantigen identification when sufficient tissue is available, our results highlight the synergy possible between analytical and computational tools.

關(guān)鍵的是螺戳，這種改進(jìn)的新抗原識(shí)別性能是通過訓(xùn)練一個(gè)基于標(biāo)準(zhǔn)數(shù)據(jù)依賴的獲取質(zhì)譜數(shù)據(jù)的預(yù)測(cè)器來實(shí) 現(xiàn)的搁宾，盡管這種技術(shù)沒有足夠的靈敏度直接檢測(cè)所有的新抗原。雖然靶向質(zhì)譜方法可能最終會(huì)提高直接 新抗原識(shí)別的敏感性倔幼，當(dāng)有足夠的組織可用時(shí)盖腿，我們的結(jié)果突出了分析工具和計(jì)算工具之間的協(xié)同作 用。

A clear limitation of our work is that it does not currently address prediction of HLA class II binding epitopes presented to CD4+ cells, which provide help for CD8+ T cell responses and can also exert antitumor activity directly. We chose to focus our study first on class I CD8+ T-cell epitopes because class I HLA expression is substantially more abundant than class II expression on solid tumor cells37,38 and class I–only antigen targeted immunotherapy has been shown to result in durable solid tumor regression with adoptive cell therapy39. Furthermore, CD8+ T cell responses to neoantigens have now been linked to clinical efficacy of immune checkpoint inhibition1,2 Nevertheless, we expect that our modeling approach is applicable to the prediction of class II epitopes. To demonstrate this possibility, we trained and successfully tested our prediction model using a (currently limited) publicly available class II HLA peptide dataset (Supplementary Fig. 12).

我們的工作的一個(gè)明顯的局限性是，它目前沒有解決預(yù)測(cè)HlaⅡ類結(jié)合表位呈現(xiàn)到CD4細(xì)胞翩腐，這提供了幫 助CD8 t細(xì)胞的反應(yīng)鸟款，也可以直接發(fā)揮抗腫瘤活性。我們選擇首先研究Ⅰ類CD8-T細(xì)胞表位茂卦，因?yàn)棰耦恏la 的表達(dá)比II類在實(shí)體腫瘤細(xì)胞上的表達(dá)要豐富得多和I類單一抗原靶向免疫治療已被證明是持久的實(shí)體腫 瘤退行性變與過繼性細(xì)胞治療何什。此外， CD8 T細(xì)胞對(duì)新抗原的反應(yīng)與免疫檢查點(diǎn)抑制的臨床療效有關(guān)等龙。 然而处渣，我們期望我們的建模方法適用于II類表位的預(yù)測(cè)。為了證明這種可能性而咆，我們訓(xùn)練并成功地測(cè)試 了我們的預(yù)測(cè)模型霍比，使用一個(gè)(目前有限的)公開的II類HLA肽數(shù)據(jù)集(補(bǔ)充圖)。 12)

Another limitation of our modeling approach is that it does not incorporate TCR binding or the availability of T-cell precursors. For example, it has been proposed that some peptide sequences may have biophysical properties that hinder TCR recognition in general40, or that some neoantigens are more self-similar than others, such that the T-cell clones that recognize them are deleted by central tolerance41. Addressing TCR binding is an important direction for future research in neoepitope prediction; however, the predictive performance of our model on the TIL neoepitope dataset and the prospective neoantigen-reactive T cell identification task demonstrate that although modeling TCR binding may provide additional benefit42,43, it is now possible to obtain therapeutically useful neoepitope predictions by modeling only HLA processing and presentation. In summary, this work offers practical neoantigen identification from routine patient samples and should be useful for the design and evaluation of future cancer immunotherapies.

我們建模方法的另一個(gè)限制是它不包含TCR結(jié)合或T細(xì)胞前體的可用性暴备。例如，有人提出们豌，某些肽序列 可能具有生物物理性質(zhì)涯捻，在一般情況下阻礙TCR識(shí)別，或者某些新抗原比其他新抗原更自我相似望迎，因此 識(shí)別它們的t細(xì)胞克隆被中央耐受性所刪除障癌。探討TCR結(jié)合是未來新表位預(yù)測(cè)研究的一個(gè)重要方向；然 而辩尊，我們的模型在TIL新表位數(shù)據(jù)集上的預(yù)測(cè)性能和預(yù)期的新抗原反應(yīng)T細(xì)胞識(shí)別任務(wù)表明涛浙，盡管模擬 TCR結(jié)合可以提供額外的效益，但現(xiàn)在僅通過對(duì)HLA處理和表示建模摄欲，就可以獲得治療有用的新表位預(yù) 測(cè)轿亮。總之胸墙，這項(xiàng)工作提供了實(shí)用的腫瘤抗原鑒定從常規(guī)的病人樣本我注，應(yīng)該是有用的設(shè)計(jì)和評(píng)估未來的癌 癥免疫療法

翻譯小組：

葉名琛、陳凱星迟隅、王俊豪但骨、鄧峻瑋、倪昊辰智袭、常彥琪奔缠、黃敬潼、李碧琪吼野、黃子亮校哎、陳志榮、鄭凌伶