ML in Gene Therapy

Ravi Kalia

Impactful Sequences

CHADOX1

CHADOX1 (30k seq)

  • Viral vector vaccine: for Covid-19
  • Adenovirus-based
  • Delivers spike protein gene to muscle cells
  • Spike protein processed into antigens
  • B cells make antibodies; T cells target infected cells
  • Antibodies and T cells attack infected lung cells

Sarah Gilbert et al. (2021)

J07BN01

J07BN01 (4k seq)

  • Lipid nanoparticles: from cellular liposomes
  • mRNA: encodes spike protein
  • Delivery: crosses cell membrane
  • Immune response: processed by APCs
  • Outcome: infected cells destroyed

Ozlem Tureci et al. (2021)

M09AX15

M09AX15 (5k seq)

  • Duchenne muscular dystrophy is a severe progressive disease
  • Caused by deletions/mutations to 79 exons of a gene
  • Elevidys rAAV Gene Therapy

Hwu (2024)

L01FF02

L01FF02 (1k seq)

  • Cancer: uncontrolled spread
  • Keytruda: humanized antibody
  • Source: mouse antibodies
  • Action: blocks immune checkpoints, enables T-cells
  • Use: melanoma, including brain (e.g., Jimmy Carter)

Emancipator (2020)

Gene Therapy*

Gene Therapy

  • Gene Therapy: fixes faulty genes
  • Goal: cure genetic disorders
  • Methods:
    • add genes
    • modify genes
    • silence genes
  • Uses: hereditary diseases, cancer

!!!VIRUSES ARE GREAT TRANSPORTS TO HUMAN CELLS!!!

Sadelin et al. (2018)

Adeno Associated Virus (rAAV) Therapy

What is a Virus?

A sequence of DNA/RNA housed inside a protein shell

  • Definition: infectious agents requiring host cells
  • Size: 1kb to 30kb
  • Reproduction: inside host cells
  • Types: DNA, RNA
  • Transmission: various routes
  • Impact: disease-causing
  • Evolution: mutations, recombination

A system defined by its information & replication dynamics

Adenovirus

  • Common cold virus
  • GOOD vaccines:
    • efficient delivery
    • strong immune response
    • proven, like covid-19
  • BAD gene therapy:
    • limited capacity
    • pre-existing immunity
    • potential inflammation

Tollefson et al. (1999)

Adeno Associated Virus

  • AAV Genome: single-stranded DNA virus
  • Components: two genes (rep and cap)
  • Rep Gene: viral replication and packaging
  • Cap Gene: capsid proteins
  • Payload: up to 5kb of therapeutic genes
  • Advantages: low immunogenicity, long-term expression

Linden et al. (2011)

AAV Approved Medicines

Listgarten et al. (2024)

Improving On Nature

Viral Vector Gene Therapy: Problems

  • Immunogenicity: immune response can reduce efficacy
  • Limited payload: small capacity for therapeutic genes
  • Pre-existing immunity: common viruses can be recognized and attacked
  • Insertional mutagenesis: risk of disrupting host genes
  • Production costs: expensive and complex manufacturing processes

Segurado (2024)

Solver: Machine Directed Evolution

Arnold et al. (2009)

Central Dogma of Molecular Biology

Crick (1970)

AAV Sequencing Workflow

  • fastqc: Check quality of raw sequencing data
  • trimmomatic: Trim adapters and low-quality sequences
  • bwa mem: Align reads to reference genome
  • samtools view -bS: Convert SAM to BAM format
  • bedtools coverage: Calculate read coverage over regions
  • bcftools call: Call variants from BAM files
  • grep, awk, cut: Extract and manipulate data from files

GENOME SEQUENCE => FITNESS

Church et al. (2019)

How?

AAV Genome Modules

Volckov et al. (2023)

Capsid

Protein shell protecting the genome

  • Structure: 60 capsomers (VP1, VP2, VP3 proteins)
  • Function: genome protection, cell entry
  • Variability: different serotypes have unique proteins
  • Engineering: modified for targeting and immune suppressing

!!!60 Points of Symmetry!!! Icosahedral shape, with internal 5 and 3 symmetry points

Kalia and Rabinowitz (2023)

Promoter

  • Regulates gene expression
  • DNA region that initiates transcription
  • Binds RNA polymerase and transcription factors
  • Interaction: TFBMs in promoters control gene expression

Kundaje et al. (2023)

Inverted Terminal Repeats

Repeated sequences at both ends of the AAV genome, with secondary structure symmetry

  • Essential for
    • packaging
    • genome encapsidation
    • replication

but 2-mers ‘CG’ are inflammatory

Logan et al. (2019)

Infeasible: Search Space

“Biochemistry disrupts Machine Learning” — Jennifer Listgarten

  • AAV: 4.5 kb long
  • Protein sequence: 20 amino acids per position
    • 60-length protein: 20**60 possibilities
    • >> atoms in the universe (~10**80)
  • Classifier/regressor can’t label space
  • Not enough lifetime/compute

Ihara et al. (2023)

Discriminative vs Generative Learning

Discriminative Learning (DL)

  • Focus: models decision boundaries
  • Objective: estimates \(P(Y \mid X)\)
  • Examples: SVMs, CNN, NN, RF

Generative Learning (GL)

  • Focus: models data distributions
  • Objective: estimates \(P(X \mid Y)\) and \(P(Y)\)
  • Examples: Naive Bayes, VAEs, Transformers

!!!GL can work better when search space infeasible!!!

Ng and Jordan (2001)

Generative Models with High Dimensional Data

  • GANs
  • Transformers
  • Normalizing Flows
  • VAEs
  • Diffusion Processes
  • Graph Neural Networks
  • Energy Based Models

Foster (2022)

Machine Learning Ritual

Steps:

  • Process raw data (Experimental data, NGS)
  • Enrich with secondary data (Open source databases)
  • Perform alignment
  • Encode metadata, sequences, DNA counts, transcript counts
  • Select candidate model
  • Train on (input, output) pairs
  • Generate novelty and/or predictions
  • Validate predictions experimentally

Alpaydin (2020)

Capsid Engineering

Objective: Optimize capsids for gene delivery and therapeutic effect

  • Improve multiple properties:
    • Target specific tissues
    • Evade immune response
    • Improve stability
    • Appropriate gene expression

Capsid Engineering: Shuffling Wild Type

  • Only 9 wild type capsids
  • Different tissue & cell specificity
  • Shuffling and expert opinion used in past to generate novel capsids
  • ML difficult on 9 examples

Grimm et al. (2018)

Capsid Engineering: Machine Learning

  • Apply ML (EBM, RF, VAE, Transformers) for novel designs
  • Generate multi-output predictions
  • Single point mutations found amazing plethora of relationships including new gene

FDA Guidance: less interest in novel capsids

Schaffer et al. (2019) and Kelsic et al. (2021)

Promoter Engineering

  • Input promoter sequences
  • Identify potential TFBMs using motif discovery tools
  • Use ML models to predict/generate functional TFBMs
  • Optimize sequences for improved binding
  • Approaches being applied include clustering, density estimation, classification with Neural and Bayesian learning
  • Transcription factors are used to compose promoters from open source and cell specific databases

Tools

  • HOMER
  • HMMR
  • bpNET

FDA more inclined to permit DNA sequences of promoters, not proteins being injected into organs.

Roberts (2011)

ITR Engineering

  • Strong belief that ‘CG’ is immunogenic and should be depleted
  • Different problem from promoter and capsid
  • ITR must maintain the symmetries expected by different arms
  • Need to generate ITRs that are different, but plausible compared to wild type
  • Different, but not too different
  • Looking for fitness gaps that nature missed
  • lookup table and modifications that are distance preserving

Duan et al. (2022)

Considerations

  • Viruses can be engineered good
  • Gene therapy is exciting and important technology
  • Machine learning is a powerful tool for this
  • Size of search space necessitates new approaches (Generative AI)
  • Maybe use tools like Topological Data Analysis

Questions