Algorithms for reconstructing within-patient tumor histories to support precision oncology

Alex Gavryushkin

DKFZ (The German Cancer Research Center), March 15, 2024, Heidelberg

Disclaimers

Precision oncology

Phylogenetic trees on tumor cells

Problem

This result is not consistent across datasets — why?

Finding important genes

Model-based approaches to genotype-phenotype data

$ f(x_1,\, \ldots,\, x_p) = \beta_0 + \sum_{i} x_i \beta_i + \sum_{i \le j} x_i x_j \beta_{i,j} $ $ + \sum_{i \le j \le k} x_i x_j x_k \beta_{i,j,k} + \ldots $

State of the art in 2018


Kieran Elmes et al. Learning epistatic gene interactions from perturbation screens. 2021

Problem #1: The hierarchy assumptions


Sarah Howles et al. Genetic variants of calcium and vitamin D metabolism in kidney stone disease. 2019

Problem #2: Scalability


Kieran Elmes et al. Learning epistatic gene interactions from perturbation screens. 2021
Hierarchy and scalability are largely solved (until you sequence more)
Kieran Elmes et al. A fast lasso-based method for inferring higher-order interactions

Problem #3 (unsolved!):

Model assumptions

$ f(x_1,\, \ldots,\, x_p) = \beta_0 + \sum_{i} x_i \beta_i + \sum_{i \le j} x_i x_j \beta_{i,j} $ $ + \sum_{i \le j \le k} x_i x_j x_k \beta_{i,j,k} + \ldots $
Kieran Elmes et al. A fast lasso-based method for inferring higher-order interactions

SNVformer


Kieran Elmes et al. SNVformer: an attention-based deep neural network for GWAS data

"eQTL" simulations (work in progress)

  • Multi-Layer Perceptron [ AI ]
  • Logistic Regression [ :) ]
  • Transformer-encoder [ AI ]
  • LightGBM (Random Forests) [ AI? ]
  • Differentiable Logic [ AI ]
  • Support Vector Machines [ AI? ]

Why don't you just do AI?


What AI?

bioDS lab @UCNZ

Thank you!