Data Science Salon: Machine Learning for Personalized Cancer Vaccines
•
1 like•1,001 views
Presented by Alex Rubinsteyn Next DSS MIA Event - https://datascience.salon/miami/ A short introduction to cancer immunotherapy followed by several machine learning problems which arise from designing personalized cancer vaccines.
Recommended
More Related Content
What's hot
Similar to Data Science Salon: Machine Learning for Personalized Cancer Vaccines
Similar to Data Science Salon: Machine Learning for Personalized Cancer Vaccines (20)
More from Formulatedby
More from Formulatedby (20)
Recently uploaded
Recently uploaded (20)
Data Science Salon: Machine Learning for Personalized Cancer Vaccines
- 1. Algorithmic selection of therapeutic cancer vaccines Alex Rubinsteyn January 31st, 2018 UM CS Pizza Seminar
- 2. OpenVax @ Mount Sinai ● www.openvax.org ● Focus: personalized cancer vaccines ○ Machine learning for immunology ○ Cancer genomics ● Started: January 1st, 2018 ● Enthusiastically translational research ● Open source software: github.com/openvax
- 5. Immune system kills (most) cancer cells Three E’s of cancer immunity, Ian York (2007)
- 6. Immune avoidance a hallmark of cancer Hallmarks of Cancer: The Next Generation (2011) Don’t get eaten by immune cells
- 7. Cancer immunotherapy ● Traditional treatments: focus on killing cancer cells directly ● Immunotherapy: get the immune system to kill the cancer ● Why is the immune system allowing cancer to spread? ○ Cancer cells inhibiting immune cells ■ Block the inhibitory signals! ○ Immune cells unable to recognize cancer as non-self ■ Teach the immune system what to kill
- 8. Flavors of cancer immunotherapy Checkpoint blockade Cellular therapies Vaccines Disinhibit CD8+ T-cells, antigens responsible for tumor clearance unknown. Success stories: ● CTLA-4 (ipi) ● PD-1 (pembro, nivo) ● PD-L1 ( atezo) Ex-vivo expansion of patient T-cells after receptor engineering and/or selection. Success stories: ● CD19 CAR T-cells for B-cell malignancies Therapeutic vaccines against tumor antigens. Significant interest in personalized “neo-antigen” vaccines. Success stories: ● ??? ● Hints of efficacy in neoantigen vaccine trials
- 9. Immunotherapy vs. Chemotherapy Nivolumab in Previously Untreated Melanoma… (Robert NEJM 2015)
- 11. What’s in a therapeutic cancer vaccine? ● Tumor antigen ○ What should immune system look for? ● Adjuvant ○ Something the immune system already responds to as dangerous ○ Examples: double-stranded RNA, mineral oil, dead bacteria ● Objective: get the immune system to learn that the antigen is bad and cells which have it should be killed
- 12. Tumor-specific antigens ● Don’t occur in normal cells ○ most commonly: mutated proteins ● Unlikely to be shared between patients ● Called “neo-antigens” Getting Personal with Neoantigen-Based Therapeutic Cancer Vaccines
- 13. Typical personalized cancer vaccine pipeline ● Sequence DNA from tumor and patient ○ Identify tumor-specific mutations ● Sequence RNA from tumor ○ Which mutations are being produced into proteins? ● Predict which mutations can be seen by immune system Computational genomics tools for dissecting tumour–immune cell interactions
- 14. Murine Experiment Checkpoint blockade cancer immunotherapy targets tumour-specific mutant antigens, Gubin et al. (2014) ● Taconic 129S6 mice ● MCA-T3 sarcoma cell line ● “mLama4” & “mAlg8” are predicted neoantigens ● Long peptide vaccine + Poly(I:C) adjuvant
- 15. More Mouse Evidence Mutated neo-antigens as targets for individualized cancer immunotherapy (Figure 3.18), Vormehr (2016) ● BALB/c mice ● CT26 colon cell line ● mRNA vaccine ● Two groups of 5 epitopes (#2 works) ○ Individual epitopes don’t work
- 16. Cathy Wu & Pat Ott’s trial @ DFCI ● 6 (stage III & IV) melanoma patients ● Up to 20 mutated peptides per vaccine ● Adjuvant: Poly-ICLC
- 17. DFCI Trial: Tumor Control Of six vaccinated patients, four had no recurrence at 25 months after vaccination, while two with recurrent disease were subsequently treated with anti-PD-1 (anti-programmed cell death-1) therapy and experienced complete tumour regression, with expansion of the repertoire of neoantigen-specific T cells.
- 19. DNA Sequencing NextSeq 500 in Mount Sinai’s Sequencing Core ● Human genome ~= 3 billion nucleotides ○ Longest chromosome ~= 250M nucleotides ● Split DNA into tiny fragments ● Read billions of short sequences
- 20. Alignment to Reference Genome ● Where did each short DNA “read” sequence come from?
- 22. Categories of mutations Mutated neo-antigens as targets for individualized cancer immunotherapy, Mathias Vohrmer ● Easy to detect ○ Substitutions (e.g. C -> G) ○ Small insertions / deletions (~10 nucleotides) ● Hard ○ Larger indels ○ Inversions ○ Gene fusions
- 24. T cell surveillance Yewdell, J.W., Reits, E. & Neefjes, J., 2003. Nature Reviews Immunology ● Proteins cleaved by proteasome ● Some of the resulting peptides loaded onto MHC to be presented on cell surface ● T cells perform surveillance of these peptide/MHC complexes ● Abnormal (non-self) displayed peptides lead to a cytotoxic T cell response 24
- 25. MHC ● Thousands of MHC alleles in human population ● Each allele capable of binding a distinct set of peptides ● Objective: Predict whether an MHC allele will bind a given peptide 25 Holland, C.J., Cole, D.K. & Godkin, A., 2013. Frontiers in Immunology
- 26. Immune Epitope Database (IEDB) ● Public dataset of B and T cell epitopes and related data curated from the literature ● Includes >200,000 in vitro binding affinity measurements of purified MHC/peptides, which is the core training data for MHC ligand prediction tools
- 27. Linear models perform reasonably well ● Binding motifs (Sette 1989) ● Position specific scoring matrices (Parker 1994) ● Ignore dependencies between positions Bjoern Peters
- 28. Neural networks do better: NetMHCpan (2007) ● Standard tool to predict peptide/MHC binding affinity given: peptide sequence and binding groove residues of MHC allele
- 29. PGV001: Safety and Immunogenicity of Personalized Genomic Vaccine (Phase I Clinical Trial at Mount Sinai) Nina Bhardwaj
- 30. PGV-001 Trial ● H&N, NSCLC, Breast, Ovarian, Urothelial, SCC, MM ● Patients w/o evidence of residual or metastatic disease ● Vaccine: ○ 10 peptides (~25 amino acids) ○ Adjuvant: Poly-ICLC ○ 10 intracutaneous injections over 6 months ● Peptide selection: ○ Phase cancer mutations with germline variants using RNA reads ○ Add multiple MHC I binding predictions overlapping same mutation ○ Ranking: expression * MHC I affinity
- 32. Trial Status ● Open & enrolling ● 1 H&N patient treated ● 1 MM patient with manufactured peptides, will begin treatment soon ● 4 patients with DNA/RNA sequencing data
- 33. New Trials in 2018 ● PGV for Glioblastoma ○ + Novocure’s TTfields ○ PI: Adelia Hormigo ● PGV for Bladder Cancer ○ + ⍺PD-1 ○ PI: Matt Galsky
- 35. Open Source Tools Developed for PGV Available at github.com/openvax varcode Python interface for VCFs, variant effect prediction isovar Determine mutant coding sequence from RNA-seq vaxrank Vaccine peptide selection (including manufacturability) epidisco Turn-key workflow to generate vaccine peptide report from FASTQ inputs (runs all bioinformatics tools) ketrew Workflow engine used to run tools on Google Cloud, AWS, and traditional HPC mhctools Standard interface to pMHC binding predictors pyensembl Python interface to Ensembl reference genome annotations
- 36. GGCGACTGTCCGGCTTTGAGCCAGGTGCCTC Intron Isovar: Phasing and Transcript Selection TGTCCGGCT ACTTGTCATGGCGACTGTCCGGCT TGGCGACTGTCCAGCT CGACTGTCCAGCT TGTCATGGCGACTGTCCAGCT Somatic mutation Germline mut. RNA Read 1 RNA Read 2 RNA Read 3 RNA Read 5 RNA Read 4 TTGAGCCAGGAGCCTC TTGAGCCA TTGAGCCAGGAGCCTC TTGTGCCAGGAGCCTC TTGTGCCAGGA Exon 1 Exon 2 Selected coding sequence includes germline mutation:
- 37. Vaxrank: Vaccine Peptide Selection vaxrank --vcf mutect.vcf --vcf strelka.vcf --bam tumor-rna.bam --vaccine-peptide-length 25 --mhc-predictor netmhcpan --mhc-alleles-file alleles.txt
- 39. Funding for Personalized Cancer Vaccines $161M $195M $270M $1.2B
- 40. Thanks!