Mikhail Karasikov

Mikhail Karasikov

ML Engineer, PhD

kaiko.ai

About me

I am a Machine Learning researcher and engineer at Kaiko.AI, where I train vision transformers (ViTs) and build foundation models (FMs) for computational pathology and other data modalities. This includes large-scale pretraining on diverse datasets, fine-tuning and optimizing the models for various downstream tasks, and making these models robust, interpretable, and useful in clinical practice.

I completed my PhD at ETH Zurich, where I designed novel algorithms and compressed data structures for indexing huge collections of biological sequences (up to Petabytes) and developed methods scalable to the entire Sequence Read Archive. These methods finally made this trove of data accessible for search by sequence at scale and demonstrated the feasibility of making all of life's code easily searchable.

Before that, I studied Math, Physics, CS, and Machine Learning at MIPT, Skoltech, and YSDA and worked on a few problems of computational structural biology at Inria Grenoble-Rhône-Alpes.

Interests
  • Machine Learning
  • Bioinformatics
  • Computational Biology
  • Compressed Data Structures
Free time
Education

  • Ph.D. in Computer Science, 2023

    ETH Zurich, Zurich, Switzerland

  • M.Sc. in Math. and Computer Science, 2017

    Skoltech, Moscow, Russia

  • M.Sc. in Applied Math. and Physics, 2017

    MIPT, Moscow, Russia

  • PG Dip. in Computer Science, 2016

    Yandex School of Data Analysis, Moscow, Russia

  • B.Sc. in Applied Math. and Physics, 2015

    MIPT, Moscow, Russia

Recent & Upcoming Talks

Efficient training of state-of-the-art pathology foundation models on orders of magnitude fewer WSIs
ECDP 2025 (Barcelona, Spain).
Jun 25, 2025 — Jun 28, 2025
Project
PhD defense
Doctoral examination (ETH Zurich, Zurich, Switzerland).
Jul 10, 2023 20:30 — 22:30
Cite Slides Thesis
Searching in nucleotide archives at Petabase scale with MetaGraph
Zurich Seminars in Bioinformatics 2022 (UZH Irchel Y55-l-06/08, Zurich, Switzerland).
Nov 24, 2022 12:15
Project Slides
Searching in nucleotide archives at Petabase scale with MetaGraph (CSHL)
Biological Data Science - CSHL Meeting 2022 (Cold Spring Harbor, New York, USA).
Nov 9, 2022 — Nov 12, 2022
Project Poster
Scalable Indexing of Sequence Data in Annotated Genome Graphs
Invited talk at JOBIM 2022 (Rennes, France).
Jul 7, 2022 15:00
Project Slides
Scalable Indexing of Sequence Data in Annotated Genome Graphs
Multi-genome Representations with MetaGraph
Talk at IGGSy 2022 (Ascona, Switzerland).
Jul 4, 2022 11:15
Project Slides
Multi-genome Representations with MetaGraph
Lossless Indexing with Counting de Bruijn Graphs
Talk at RECOMB 2022 (La Jolla, California, USA).
May 24, 2022 14:30
Project Slides Paper
Lossless Indexing with Counting de Bruijn Graphs
Topology-based Sparsification of Graph Annotations
Proceedings Presentation at ISMB/ECCB 2021, HiTSeq COSI track (online).
Jul 25, 2021 11:20 — 11:40
Project Slides Video Paper
Topology-based Sparsification of Graph Annotations

Projects

Pathology Foundation Models
We enhance the SSL workflow and train pathology FMs at state-of-the-art level using up to 100-fold less data than current standards. We additionally incorporate high-resolution fine-tuning to further improve the FMs. With 12k TCGA WSIs, we trained a model that is on par with Virchow2, and our best model trained on 92k WSIs with high-resolution post-training showed an average top-1 performance.
Preprint Github
Pathology Foundation Models
Ocean Microbiomics Database
Collaboration with Sunagawa Lab. Provided k-mer based sequence search and Counting de Bruijn graph indexes for the Genome Collection of the Ocean Microbiomics Database. Other contributors: Lucas Paoli, Harun Mustafa, Andre Kahles. (Published in Nature).
Project Paper
Ocean Microbiomics Database
MetaGraph
A C++ framework library for indexing very large collections of DNA/Protein sequences and a tool for sequence search, alignment, and assembly. Although the target use cases of MetaGraph overlap with BLAST, MetaGraph mainly focuses on the scalable indexing of raw sequencing data in annotated de Bruijn graphs with up to $\sim 10^{12}$ nodes and $\sim 10^{7}$ annotation labels. It also provides an online platform MetaGraph Online. Other contributors: Marc Zimmermann, Thomas Zhou, the MetaGraph team.
Preprint Github
MetaGraph
Compressed Hybrid Bit Vector Representations
A C++ library with hybrid schemes for representing bit vectors in compressed space.
Benchmarks 1e7 Benchmarks 1e9 Github
Compressed Hybrid Bit Vector Representations
GeoDNA
A portal for sequence search and geographical positioning based on the metagenomic MetaSUB data. The initial prototype was set up on a weekend but it served well and was also used as a base for the MetaGraph Search platform. Other contributors: Marc Zimmermann, Jiayu Chen, André Kahles, Thomas Zhou. (Published in Cell).
Paper Github News
GeoDNA
De Bruijn Graph Visualizer
A web app visualizing de Bruijn graphs and the BOSS table (Bowe et al.). Developed to interactively illustrate the core data structure used as a k-mer index for graph representation in MetaGraph.
Github
De Bruijn Graph Visualizer
Protein Scoring
A method for single-model coarse-grained protein quality assessment developed during my internship at Inria Grenoble-Rhône-Alpes (with Guillaume Pagès and Sergei Grudinin).
Slides Paper Gitlab
Protein Scoring
Activity Prediction
Classification of time-series data from smartphone accelerometer sensor. Implemented as a practical demonstration of the methods developed in my B.Sc. thesis.
Paper Github
Activity Prediction

Featured Publications

Mikhail Karasikov, Joost van Doorn, Nicolas Känzig, Melis Erdal Cesur, Hugo Mark Horlings, Robert Berke, Fei Tang, Sebastian Otálora (2025). Training state-of-the-art pathology foundation models with orders of magnitude less data. In review.

Cite Project DOI Preprint Code Model

Mikhail Karasikov, Harun Mustafa, Daniel Danciu, Marc Zimmermann, Christopher Barber, Gunnar Rätsch, André Kahles (2024). Indexing All Life's Known Biological Sequences. In review.

Cite Project DOI Preprint Code

kaiko.ai, Nanne Aben, Edwin D. de Jong, Ioannis Gatopoulos, Nicolas Känzig, Mikhail Karasikov, Axel Lagré, Roman Moser, Joost van Doorn, Fei Tang (2024). Towards large-scale training of pathology foundation models. In arXiv.

Cite Project DOI Preprint Model

Mikhail Karasikov (2023). Scalable Annotated Genome Graphs for Representing Sequence Data. In ETH Zurich Research Collection.

Cite DOI Thesis

Mikhail Karasikov, Harun Mustafa, Gunnar Rätsch, André Kahles (2021). Lossless Indexing with Counting de Bruijn Graphs. In RECOMB 2022.

Cite Project DOI Preprint Code

Daniel Danciu, Mikhail Karasikov, Harun Mustafa, André Kahles, Gunnar Rätsch (2021). Topology-based Sparsification of Graph Annotations. In ISMB/ECCB 2021.

PDF Cite Project Slides DOI Code

Mikhail Karasikov, Harun Mustafa, Daniel Danciu, Marc Zimmermann, Christopher Barber, Gunnar Rätsch, André Kahles (2020). MetaGraph: Indexing and Analysing Nucleotide Archives at Petabase-scale. In bioRxiv.

Cite Project DOI Preprint Code

Mikhail Karasikov, Guillaume Pagès, Sergei Grudinin (2018). Smooth Orientation-Dependent Scoring Function for Coarse-Grained Protein Quality Assessment. In Bioinformatics.

PDF Cite Project DOI Code

Mikhail Karasikov, Harun Mustafa, Amir Joudaki, Sara Javadzadeh No, Gunnar Rätsch, André Kahles (2018). Sparse Binary Relation Representations for Genome Graph Annotation. In RECOMB 2019.

PDF Cite Project Slides DOI Code

Teaching

Courses TAed at ETH Zürich, Institute for Machine Learning:

Contact