DeltaMSI: Artificial Intelligence-based modeling of microsatellite instability scoring by next- generation sequencing

doi:10.21203/rs.3.rs-1645344/v1

Download PDF

Research Article

DeltaMSI: Artificial Intelligence-based modeling of microsatellite instability scoring by next- generation sequencing

https://doi.org/10.21203/rs.3.rs-1645344/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 01 Mar, 2023

Read the published version in BMC Bioinformatics →

You are reading this latest preprint version

Background: DNA mismatch repair deficiency (dMMR) testing is crucial for detection of microsatellite unstable (MSI) tumors eligible for immunotherapy. MSI is detected by aberrant indel length distributions of microsatellite marker loci, either by operator-dependent visual inspection of PCR-fragment length profiles or by automated scoring using bioinformatic scripts on next-generation sequencing (NGS) data. The former is time consuming and low-throughput while the latter provides simplified binary scoring of a single parameter, typically the number of discrete integrated peaks of the indel distribution. The aim of this study was to leverage the multi-parametric visual scoring of complex indel distribution panels and high-throughput of automated NGS data analysis into a script trained by artificial intelligence and validate it on a large real-world data set of solid tumors.

Results: A script (DeltaMSI) was developed that integrates the full complexity of indel distribution (number of peaks, relative distribution and area under the curve). It was trained on the indel distribution of 36 widely used microsatellite marker loci in a dataset of 133 MMR proficient (pMMR) and 46 dMMR tumor samples, taking loss of MLH1/MSH2/PMS2/MSH6 protein expression measured by immunohistochemistry as independent reference technique. After selection of the optimal modelling and marker panel, the two top-performing models per locus (logistic regression and support vector machine) were then integrated into DeltaMSI for combined prediction of MSI status on 28 marker loci at sample level. A bootstrapping method of 50 simulations was used for training and the final model was evaluated on an independent validation set. Diagnostic performance of DeltaMSI was compared to that of mSINGS, a widely used script for NGS-based MSI detection, with the same bootstrapping method. Finally, we tested the robustness of the optimal model on a large set (N=1072) of unselected, consecutive solid tumor samples in a real-world setting and fine-tuned the thresholding for routine clinical use.

Conclusions: DeltaMSI achieved higher robustness at equal diagnostic power (AUC=0.950; 95% CI 0.910-0.976) as compared to mSINGS (AUC=0.876;95%CI 0.823-0.918). Its sensitivity of 90% at 98% specificity indicated its clinical potential for high-throughput MSI screening in all tumor types.

Clinical Trial Number/IRB: B1172020000040, Ethical Committee, AZ Delta General Hospital

Microsatellite instability

indel length distribution analysis

DNA mismatch repair deficiency

targeted resequencing

immunotherapy

logistic regression

support vector machine

machine learning

Microsatellites are DNA elements composed of short repetitive motifs that are prone to misalignment and frameshift mutations during cell division. In healthy cells, the ensuing small indels or single-base mispairs are corrected by heterodimer enzyme complexes of the DNA mismatch repair (MMR) system encoded by the key MMR genes MLH1, MSH2, PMS2 and MSH6. DNA mismatch repair deficiency (dMMR) results in the progressive accumulation of genetic alterations, potentially dysregulating many oncogenes or tumor suppressor genes. The molecular hallmark of dMMR is microsatellite instability (MSI), with expansions or contractions in the number of tandem repeats throughout the genome. This phenomenon is observed in a considerable proportion of colorectal, endometrial, gastric, pancreatic, brain, biliary tract, urinary tract and ovarian tumors [1].

As dMMR is caused by loss of function variants in the mismatch repair genes, immunohistochemistry (IHC) for MLH1, MSH2, MSH6 and PMS2 loss of expression is widely used as pre-screening for MMR status [1-6]. An alternative approach is to look for microsatellite instabilities by fragment length analysis of fluorescent PCR products of microsatellite loci [3, 4, 6, 7]. MSI is typically detected molecularly by aberrant indel length distributions of a panel of microsatellite marker loci, either by operator-dependent visual inspection of PCR-fragment length analysis profiles or by automated scoring using bioinformatics scripts on next-generation sequencing (NGS) data. The former is time consuming and low-throughput while the latter typically provides simplified binary scoring using a single parameter. Commonly used scripts such as mSINGS [8], MSIsensor [9] and MSI-Finder [10] all use the number of discrete, integrated indel lengths within each marker locus as single parameter. Such simplification can result in inaccurate stability calling at locus level in case of left-shifted distributions or minimal microsatellite shifts (Figure S1 and [11]). A script that can take into account the full complexity of indel distribution could potentially overcome such inaccuracies.

Bioinformatic tools for MSI detection on NGS data are available in different flavours: A) using indel distributions of microsatellite loci within a tumor-normal pair (e.g. Mantis, USCI-msi) [12, 13], B) using indel distributions of microsatellite loci versus a fixed reference set (e.g. mSINGS, MSIFinder) [8, 10] or C) analysing the occurrence of small variants (single nucleotide variants or microindels) (e.g. MSIseq) [14]. In clinical settings, systematically sequencing pairs of normal and tumor samples as in A, or sequencing large genomic areas as required by C is not cost-effective. A script with minimal need for optimization of a reference set and a high robustness would thus be a useful advance.

All available tools for MSI screening on NGS data report high diagnostic performance, based on comparisons with MSI-PCR as reference. However, as shown by Dedeurwaerdere et al [11], despite their concordance, all dMMR detection techniques based solely on molecular methods (MSI-NGS and MSI-PCR) fail to detect up to 5%-10% of dMMR tumors with demonstrated loss of MLH1/MSH2/PMS2 or MSH6 immunohistochemical (IHC) staining for biological reasons (minimal MSI shifts) (mostly endometrial carcinoma) or sensitivity issues (low tumor cell percentage). To assess real-world clinical performance, training and evaluation of molecular methods versus IHC is therefore important.

The aim of this study was to leverage the multi-parametric visual scoring of complex indel distribution panels and high-throughput of automated NGS data analysis into a single script trained by artificial intelligence on a set of solid tumors, taking MMR status measured by IHC as imperfect independent reference technique, and validate this on a large, real-world data set of solid tumors.

Study specimens

All specimens were obtained from cancer patients as part of standard clinical care. The complete NGS dataset of N=1072 consecutive solid tumor samples (large variety of origin tissues) consists of all unselected, consecutive patient samples processed for diagnostic NGS at AZ Delta General Hospital from Jan 1, 2019 to Dec 31, 2020, and included N=215 patients with an additional IHC testing results.

Wet lab pre-processing

Immunohistochemistry for MSH2, MSH6, PMS2 and MLH1 was performed on 5-µm thick sections of formalin-fixed, paraffin-embedded (FFPE) tumor tissue on a Ventana, Benchmark Ultra device (Ventana Medical Systems, Arizona, USA) as previously described [11]. Tumors were classified as dMMR if no nuclear staining or nuclear staining in less than 10% of invasive tumor cells for 1 or several markers was seen in the presence of a positive internal control (inflammatory and stromal cells). Tumors with nuclear staining for all for markers in at least 10% of invasive tumor cells are considered to be pMMR. DNA was extracted from 10-µm thick sections of the same FFPE tumor tissue blocks as used for immunohistochemistry, using the Cobas DNA Sample Preparation Kit (Roche, Basel, Switzerland) with macrodissection guided by haematoxylin eosin staining where needed, with elution in 10 mM Tris–HCl pH8.0.NGS was performed on an Illumina MiSeq (PE150) using a 138 kb customized hybridization capture-based (NimbleGen SeqCap EZ HyperPlus, Roche) gene panel that included 36 microsatellite loci [11] (Supplementary Table 1).

Data pre-processing

Reads were aligned to the human reference genome (hg19) with BWA-mem2 (2.2.1) [15], sorting and duplicate marking were performed using elPrep (5.0.2) [16] and SAMtools (1.13) [17]. Starting from this output the DeltaMSI tool was developed. In parallel mSINGS v4.0 was used to create a comparison with the same baseline sample sets and regions. The NGS panel design included 36 published microsatellite regions: 15 proposed by Salipante [8], 7 from the Idylla assay (Biocartis, Belgium) [18, 19], 10 proposed by Hause[20] and 4 regions from the updated Bethesda guideline [21]. Bam files were read by DeltaMSI, and additional read filtering was performed. Besides the removal of duplicates, reads were filtered on a mapping quality of at least 20. Only reads overlapping the complete microsatellite regions with a flanking region of 5 bases on each side were used. Regions within samples were filtered with a minimum depth of 30x (similar as mSINGS). The length distribution graphs, obtained by counting the observed fragment lengths, were cleaned by removing lengths with a depth of only 1 read (to decrease the complexity of the regions). As raw data for modeling we used normalized read depth per position for each marker locus, obtained by dividing the absolute read number per position by the maximally observed depth of all lenghts within that locus, obtaining a value between 0 and 1 for each possible length within that region.

Machine learning modelling by train-validation-test split approach

The model set (N=179, N=133 pMMR and N=46 dMMR deficient samples) was randomly split in a training set, validation set and test set of equal size (graphical study design in Fig.1)

The training set was first used to select the loci with acceptable raw data quality based on acceptable coverage within sample and in 75% of all training set samples: 29 of 36 loci were retained (Fig. 2). We then trained 7 machine learning models on the training set (scikit-learn) [22] using normalized read depth per position as feature versus binary dMMR status measured by IHC as target. The three outlier methods (isolation forest, local outlier factor and one class support vector machine) were trained with pMMR samples only, the other methods (logistic regression, random forest, naive bayes and support vector classifier (SVC, aka support vector machine)) were trained with an unbalanced set (45 pMMR, 16 dMMR). For each method hyperparameter tuning was performed using a grid search with a kfold of 3. Each model used the same kfold split of the samples, regardless of the number of filtered samples for that region within that set.

Next, the validation set was used to select the best performing models based on the AUC score, additionally curate the markers for diagnostic power and set thresholds. The highest AUC were obtained with logistic regression, random forest and SVC (Fig. 2). Random forest was removed due to strong overfitting of each region. 28 of 29 marker loci achieved AUC of 0.60 or higher in logistic regression and support vector classifier and were retained in the final modeling. We additionally combined both models into a combined voting model that scores regions as unstable if both models predict the region as unstable. For each model, the percentage of loci scored as unstable was calculated (sample score) for final dMMR/pMMR classification at sample level.

ROC analysis was done by MedCalc (version 19.2.1, MedCalc software Ltd, Mariakerke, Belgium) and Python scikit-learn[22].

Clinical validation real world data

We first tested the diagnostic power of logistic regression and SVC and the combined voting model in the test set, and compared it to another widely used script, mSINGS (v4.0) ¹³. We implemented mSINGS first on all 28 marker regions also used by DeltaMSI (mSINGS all regions) and then separately, using a highly curated set of 10 informative loci (mSINGS10) and a clinically validated fixed set of MSS baseline samples, exactly as previously reported [11]. Receiver Operating Characteric (ROC) curve plotting (Fig. 3A) indicates strong diagnostic power of DeltaMSI, with AUC of logistic regression, support vector classifier and the combined voting model of 0.92, 0.91 and 0.94, respectively, similar to the optimized mSINGS10 (AUC=0.93). The test set was also used to optimize cutoffs for subsequent clinical verification on real-world data. We first opted for dual thresholding with a lower cutoff below which samples are pMMR with 95% specificity, separated by a gray zone interval from a higher cutoff, above which samples are dMMR with 95% specificity (Fig.3B).

DeltaMSI was then applied on n=1072 consecutive clinical samples of solid tumors of various tissue types, including colorectal (20%), lung (27%), breast (14%), pancreas (6%), ovarium (5%), melanoma (5%), endometrium (4%), brain (4%), prostate (2%), GIST (1%) and stomach (1%) (Supplementary table 2). The combined voting model of DeltaMSI was compared to mSINGS on 28 marker loci, using 0.2-0.3 as gray zone for mSINGS: the overall concordance of MSS/MSI samples on colorectal cancers (N=214) was 72% with an estimated prevalence of MSI of 21% (DeltaMSI) and 18% (mSINGS) that was on the high range epidemiologically (normally estimated around 10%-15%). In the total set of 1072 samples, DeltaMSI and mSINGS classified 11% and 8% of samples as MSI, respectively, with gray zone results in 17% and 24% of samples, respectively, suggesting suboptimal thresholding with too many gray zone results for both scripts. Therefore, we selected an additional set of n=36 samples with gray zone for independent verification using IHC, and aggregated these into our initial model set for iterated ROC analysis on a total of n=215 samples (Table 1): all DeltaMSI models achieved outstanding AUC of 0.95 (95%CI 0.91-0.98), similar to the optimized mSINGS10 (AUC=0.93, 95%CI 0.89-0.96) but statistically superior to mSINGS on all 28 marker regions (AUC=0.876, 95%CI 0.82-0.92, P<0.05). A single binary threshold could be defined (>0.26) above which microsatellite instability was predicted at >98% specificity and likelihood ratios above 70 (illustrated in Supplementary Table 3, Supplementary information paragraph 4). This corresponded to a sensitivity of 90% (95%CI 76-96). Further decreasing binary cutoff had no impact on exclusion of MSI (negative likelihood ratio) and thus no beneficial impact on screening rate: in total 5 cases were missed of which 2 for obvious technical reasons (lower tumor cell percentage <40%).

Tables

Table 1

AUC of DeltaMSI models and mSINGS versus IHC on n=215 samples Table shows the AUC (95% CI) of DeltaMSI models (logistic regression, SVC and combined voting) as compared to mSINGS on all 28 marker regions (mSINGS_all regions) and mSINGS on 10 top-performing regions (mSINGS10) in N=215 samples confirmed by IHC (179 of the original model set plus 36 gray zone result samples that were additionally tested by IHC). P value versus mSINGS_all regions. Table also indicates the proposed binary cut-off for clinical use and the associated sensitivity/specificity and positive (+LR) and negative (-LR) likelihood ratios to predict or rule out dMMR status.

Variable	AUC	95% CI ^b	P	Cut-off	Sensitivity	95% CI	Specificity	95% CI	+LR	-LR
Logistic regression	0.949	0.909 to 0.975	0.0216	0.26	88	75.7 - 95.5	98.79	95.7 - 99.9	72.6	0.12
SVC	0.950	0.911 to 0.976	0.0198	0.26	90	78.2 - 96.7	98.79	95.7 - 99.9	74.25	0.1
Combined Voting	0.950	0.910 to 0.975	0.0224	0.26	88	75.7 - 95.5	100	97.8 - 100.0	∞	0.12
mSINGS10	0.931	0.887 to 0.962	0.0295	0.26	90	78.2 - 96.7	98.79	95.7 - 99.9	74.25	0.1
mSINGS_allregions	0.876	0.823 to 0.918		0.29	68	53.3 - 80.5	98.17	94.7 - 99.6	37.17	0.33

Robustness of DeltaMSI

DeltaMSI was developed on a well powered model set which is not always available in routine diagnostic settings. To evaluate the impact of sample size on AUC, we performed bootstrapping on training/validation sets of decreasing size. 50 simulations were run on 5 sample sizes: 20-10:20-10; 20-10:10-5; 20-20:10-10; 40-20:20-10; 45-16:43-15 (train pMMR-dMMR: validation pMMR-dMMR). Decreasing sample size down to the smallest sample size (20-10:10-5) did not negatively affect AUC (AUC mean: 0.904, median: 0.902, 95% CI: 0.873-0.947) as compared to the original sample size (45-16:43-15) (AUC mean: 0.893, median: 0.904, 95% CI: 0.782-0.966) for the combined voting model (Fig. 4A) (P=0.25). Comparison of DeltaMSI models versus mSINGS on all 28 regions, in 50 bootstrapping simulations using the smallest 20-10:10-5 train-validation size, confirmed the trend towards superior AUC of the combined voting model (AUC mean 0.904, median: 0.902, 95% CI: 0.873-0.947) as compared to mSINGS (AUC mean: 0.872, median: 0.872, 95% CI: 0.831-0.915) (P<0.001) (Fig.4B) and also confirmed robustness of the 28 selected marker loci (not shown). Finally, we evaluated robustness of mSINGS versus DeltaMSI (combined voting model) on the complete real-world data set (n=1072), using minimal sample size (training 20-10: validation 10-5 pMMR-dMMR): over 10 bootstrapping simulations, DeltaMSI had a more stable performance with less variation across simulations in the percentage of samples classified as MSI/gray zone/MSS classification (Fig.4C), and thus indicating a lower dependency on the optimal training/validation set.

DeltaMSI achieved a clinically acceptable sensitivity around 90% at excellent specificity for MSI screening of all common cancer types. Its diagnostic performance was comparable to that of another widely used script, mSINGS. As compared to mSINGS, however, it offered a higher robustness during implementation, with a diagnostic performance that was less influenced by the composition of the training/baseline samples sets, and thus guaranteeing a more streamlined implementation with less validation efforts. Our study shows that machine learning on complex raw data (indel distribution at locus level) can achieve higher robustness at sample level than scripts running on monoparametric decomplexified data.

AUC: area under the curve of Receiver Operating Characteristics Curve; dMMR: DNA MisMatch Repair Deficiency, IHC: immunohistochemistry; Logres: logistic regression; MSI: microsatellite instability; NGS: next-generation sequencing: SVC: support vector classifier

Acknowledgements

The authors thank Joke Breyne for clinical data curation.

Authors contributions

Data collection and curation by K.S., F.D. Scripting by K.S. Manuscript writing by G.M. and K.S. Study design by G.M., K.S. and F.D. Manuscript editing by all co-authors. Advice on statistics and modeling by D.D. and P.D.J. All authors read and approved the final manuscript. G.M. is guarantor of the data.

Funding

The study was supported by unconditional grant by Roche Diagnostics International, Switzerland to GM, and by an unconditional investigator-initiated study grant by AstraZeneca (Cambridge, UK) to GM (ESR-18-13805). The sponsors had no influence on data collection and were not involved in data analysis and manuscript preparation.

Availability of data and materials

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Ethics approval and consent to participate

The study was conducted in accordance with the Declaration of Helsinki. The study was approved by the AZ Delta Ethical Committee (Clinical Trial Number/IRB B1172020000040, study 20126, approved on November 23, 2020) with a waiver of informed consent since the study relied only on secondary use of biomaterials and data that were previously obtained as part of standard medical care.

Consent for publication

Not applicable

Competing interests

The authors declare that they have no competing interests.

Availability and requirements

Project name: DeltaMSI
Project home page: https://github.com/RADar-AZDelta/DeltaMSI
Operating system(s): Platform independent
Programming language: Python
Other requirements: available in Pipfile
License: GPL v3.0
Any restrictions to use by non-academics: See License

Lynch HT, Snyder CL, Shaw TG, Heinen CD, Hitchins MP: Milestones of Lynch syndrome: 1895-2015. Nat Rev Cancer 2015, 15(3):181-194.
Hampel H, Frankel WL, Martin E, Arnold M, Khanduja K, Kuebler P, Nakagawa H, Sotamaa K, Prior TW, Westman J et al: Screening for the Lynch syndrome (hereditary nonpolyposis colorectal cancer). N Engl J Med 2005, 352(18):1851-1860.
Lindor NM, Burgart LJ, Leontovich O, Goldberg RM, Cunningham JM, Sargent DJ, Walsh-Vockley C, Petersen GM, Walsh MD, Leggett BA et al: Immunohistochemistry versus microsatellite instability testing in phenotyping colorectal tumors. J Clin Oncol 2002, 20(4):1043-1048.
Shia J: Immunohistochemistry versus microsatellite instability testing for screening colorectal cancer patients at risk for hereditary nonpolyposis colorectal cancer syndrome. Part I. The utility of immunohistochemistry. J Mol Diagn 2008, 10(4):293-300.
Shia J, Tang LH, Vakiani E, Guillem JG, Stadler ZK, Soslow RA, Katabi N, Weiser MR, Paty PB, Temple LK et al: Immunohistochemistry as first-line screening for detecting colorectal cancer patients at risk for hereditary nonpolyposis colorectal cancer syndrome: a 2-antibody panel may be as predictive as a 4-antibody panel. Am J Surg Pathol 2009, 33(11):1639-1645.
Zhang L: Immunohistochemistry versus microsatellite instability testing for screening colorectal cancer patients at risk for hereditary nonpolyposis colorectal cancer syndrome. Part II. The utility of microsatellite instability testing. J Mol Diagn 2008, 10(4):301-307.
Boland CR, Thibodeau SN, Hamilton SR, Sidransky D, Eshleman JR, Burt RW, Meltzer SJ, Rodriguez-Bigas MA, Fodde R, Ranzani GN et al: A National Cancer Institute Workshop on Microsatellite Instability for cancer detection and familial predisposition: development of international criteria for the determination of microsatellite instability in colorectal cancer. Cancer Res 1998, 58(22):5248-5257.
Salipante SJ, Scroggins SM, Hampel HL, Turner EH, Pritchard CC: Microsatellite instability detection by next generation sequencing. Clin Chem 2014, 60(9):1192-1199.
Niu B, Ye K, Zhang Q, Lu C, Xie M, McLellan MD, Wendl MC, Ding L: MSIsensor: microsatellite instability detection using paired tumor-normal sequence data. Bioinformatics 2014, 30(7):1015-1016.
Zhou T, Chen L, Guo J, Zhang M, Zhang Y, Cao S, Lou F, Wang H: MSIFinder: a python package for detecting MSI status using random forest classifier. BMC Bioinformatics 2021, 22(1):185.
Dedeurwaerdere F, Claes KB, Van Dorpe J, Rottiers I, Van der Meulen J, Breyne J, Swaerts K, Martens G: Comparison of microsatellite instability detection by immunohistochemistry and molecular techniques in colorectal and endometrial cancer. Scientific reports 2021, 11(1):12880.
Kautto EA, Bonneville R, Miya J, Yu L, Krook MA, Reeser JW, Roychowdhury S: Performance evaluation for rapid detection of pan-cancer microsatellite instability with MANTIS. Oncotarget 2017, 8(5):7452-7463.
Zheng K, Wan H, Zhang J, Shan G, Chai N, Li D, Fang N, Liu L, Zhang J, Du R et al: A novel NGS-based microsatellite instability (MSI) status classifier with 9 loci for colorectal cancer patients. J Transl Med 2020, 18(1):215.
Huang MN, McPherson JR, Cutcutache I, Teh BT, Tan P, Rozen SG: MSIseq: Software for Assessing Microsatellite Instability from Catalogs of Somatic Mutations. Scientific reports 2015, 5:13321.
Vasimuddin Md SM, Heng Li, Srinivas Aluru: Efficient Architecture-Aware Acceleration of BWA-MEM for Multicore Systems. arXiv:190712931 2019.
Herzeel C, Costanza P, Decap D, Fostier J, Wuyts R, Verachtert W: Multithreaded variant calling in elPrep 5. PloS one 2021, 16(2):e0244471.
Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, Marth G, Abecasis G, Durbin R, Genome Project Data Processing S: The Sequence Alignment/Map format and SAMtools. Bioinformatics 2009, 25(16):2078-2079.
Samaison L, Grall M, Staroz F, Uguen A: Microsatellite instability diagnosis using the fully automated Idylla platform: feasibility study of an in-house rapid molecular testing ancillary to immunohistochemistry in pathology laboratories. J Clin Pathol 2019, 72(12):830-835.
Pecriaux A, Favre L, Calderaro J, Charpy C, Derman J, Pujals A: Detection of microsatellite instability in a panel of solid tumours with the Idylla MSI Test using extracted DNA. J Clin Pathol 2020.
Hause RJ, Pritchard CC, Shendure J, Salipante SJ: Classification and characterization of microsatellite instability across 18 cancer types. Nat Med 2016, 22(11):1342-1350.
Umar A, Boland CR, Terdiman JP, Syngal S, de la Chapelle A, Ruschoff J, Fishel R, Lindor NM, Burgart LJ, Hamelin R et al: Revised Bethesda Guidelines for hereditary nonpolyposis colorectal cancer (Lynch syndrome) and microsatellite instability. J Natl Cancer Inst 2004, 96(4):261-268.
Pedregosa F: Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 2011, 12:2825-2830.

No competing interests reported.

Download PDF

Journal Publication

published 01 Mar, 2023

Read the published version in BMC Bioinformatics →

Editorial decision: Major revision
28 Sep, 2022
Reviews received at journal
03 Aug, 2022
Reviewers agreed at journal
30 Jul, 2022
Reviewers agreed at journal
13 Jul, 2022
Reviewers invited by journal
15 Jun, 2022
Editor assigned by journal
07 Jun, 2022
Editor invited by journal
26 May, 2022
Submission checks completed at journal
26 May, 2022
First submitted to journal
11 May, 2022

You are reading this latest preprint version

DeltaMSI: Artificial Intelligence-based modeling of microsatellite instability scoring by next- generation sequencing

Status:

Journal Publication

Version 1

Abstract

Figures

Background

Implementation

Results

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1