1 of 1

Availability of adaptive potential shifts favor from nonmutators to mutators. Area plots show non-vs. mutator fixation probabilities. As available beneficial mutations are increased, mutators gain favor in progressively larger population sizes. Shaded bands show bootstrapped 95% confidence intervals.

throughput stable scaling

per-PE workload

 1 2 4 6 8 10 12 14 16

Wafer-Scale Simulation of Mutator Allele Dynamics

in Large Asexual Populations

  • agent-based model based on Raynes et al., 2018
  • using Wafer-Scale Engine (WSE) enabled experiments with large agent populations
  • implemented custom kernel via Cerebras SDK
  • on WSE, each processor hosts an independent subpopulation; msg send/recvs migrate agents
  • model parameters:
    • Pbeneficial mutation = 1/100k
    • Pdeleterious mutation = 1/1k
    • fitness = nben muts - ndel muts

hypermutators

favored

normomutators favored

  • mutator traits can play role in cancer and antibiotic resistance

  • existing work finds increased population size can favor selection for hypermutator traits

  • research question: under what conditions do very large asexual populations favor normomutators?

Matthew Andres

Moreno

(UMich)

Emily Dolson (MSU)

Luis

Zaman

(UMich)

morenoma@umich.edu

normomutator”

hypermutator”

  • effective DNA repair genes � normal mutation rate

  • evolutionary trade-off:�offspring suffer fewer common harmful mutations
  • flawed DNA repair genes� elevated mutation rate

  • evolutionary trade-off: offspring may find new rare beneficial mutations

Raynes et al., 2018:

🔎

Cerebras

CS-2

Wafer-Scale

Engine

NVIDIA A100 GPU

850k processors

MSU Computer Science & Engineering

Ecology, Evolution, & Behavior

hyper�mutator: 100x mut probs

  1. when few beneficial mutations are available, normomutators regain advantage at very large population sizes
  2. normomutators persist through greater available beneficial mutations when hypermutators must arise de novo
  3. time series data indicate hypermutators can transiently peak >99% prevalence in scenarios where normomutators persist
  4. WSE allows pop size >1.5 billion, 294× speedup vs. GPU

parent

offspring

vs

fixation probability

Results

Introduction

Simulation Design

Discussion

Buitrago, P. A. & Nystrom, N. A. (2021). Neocortex and Bridges-2: A High Performance AI+HPC Ecosystem for Science, Discovery, and Societal Good, 205–219. Carla 2020. https://doi.org/10.1007/978-3-030-68035-0_15

Brashear, W., Chakravorty, D., He, Z., O’Connor, D., Siegmann, E., Buitrago, P. A., & Sanielevici, S. (2025). ByteBoost: An advanced cybertraining program designed to enhance research on testbed systems. In Practice and Experience in Advanced Research Computing 2025: The Power of Collaboration (pp. 1–5). PEARC ’25: Practice and Experience in Advanced Research Computing. ACM. https://doi.org/10.1145/3708035.3736082

John C. Gunther (2014). Algorithm 938: Compressing circular buffers. ACM Trans. Math. Softw. 40, 2, Article 17 (February 2014), 12 pages. https://doi.org/10.1145/2559995

Moreno M. A., Zaman L., & Dolson E. (2024). Structured Downsampling for Fast, Memory-efficient Curation of Online Data Streams. arXiv preprint arXiv:2409.06199. https://doi.org/10.48550/arXiv.2409.06199

Y. Raynes, C.S. Wylie, P.D. Sniegowski, & D.M. Weinreich (2018), Sign of selection on mutation rate modifiers depends on population size, Proc. Natl. Acad. Sci. USA 115 (13) 3422-3427, https://doi.org/10.1073/pnas.1715996115.

Olivier Tenaillon, Bruno Toupance, Hervé Le Nagard, François Taddei, Bernard Godelle (1999), Mutators, Population Size, Adaptive Landscape and the Adaptation of Asexual Populations of Bacteria, Genetics, Volume 152, Issue 2, 1 June, Pages 485–493, https://doi.org/10.1093/genetics/152.2.485

294× speedup on WSE vs GPU

  • explore fitness landscape ruggedness as mechanism limiting adaptive potential, using more sophisticated genome models
  • implement agent migration via point-to-point messaging, to explore effect of population mixing on evolutionary dynamics
  • extend scope of wafer-scale simulations to applications in evo-epidemiology and evolutionary transitions in individuality

Future Work

Performance Benchmark

References

UMich Ecology & Evolutionary Biology

Complex Systems

MIDAS

poster abstract @ hopth.ru/fm

This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research (ASCR), under Award Number DE-SC0025634. This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.

This material is also based upon work supported by the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, a Schmidt Sciences program. Computational resources were provided by MSU iCER and by Pittsburgh Supercomputing Center Neocortex under NSF ACCESS Innovative Projects Allocation BIO240102 (Buitrago & Nystrom, 2021). This work was made possible by the ByteBoost 2024 workshop (Brashear et al., 2025). Thank you also to M. Jacquelin and L. Wilson at Cerebras.

Acknowledgement

On-device Data Management

Cerebras Software Language (CSL) simulation code at https://github.com/mmore500/

wse-sync-ga

(MIT License).

Spatiotemporal composition of simulated populations from Wafer-Scale Engine experiments. The snapshots depict populations of 191 million agents with a 2D spatial structure, modeled using a site-explicit genome framework configured with an adaptive potential of 14 beneficial mutations. Mutators are introduced de novo. Raster values are binary: white pixels represent sampled non-mutators, while black pixels represent sampled mutators. Panels show replicates in which mutators either fail to reach fixation (a) or succeed in reaching fixation (b).

  • challenge: balancing simulation content vs. data in limited on-device memory
    • 48kb memory per PE
  • best-effort data collection can meet constraints of hardware accelerators
    • bio analysis: coarsening/sampling often ok
  • some experiments required time series measurements of mutator prevalence
    • duration (i.e., to fixation) unknown a priori
  • used generalized ring buffer data structure (Gunther, 2014; Moreno, 2024)
    • flat, fixed-size buffer (static allocation)
    • store successive local samples as single bits
    • recording density dynamically coarsens via overwrites once storage capacity is reached
    • keeps systematic temporal cross-sample
    • no bookkeeping overhead or unused storage

Initial supply of mutators influences nonmutator resilience to available adaptive potential. When initially rare, substantially more adaptive potential is required to fix mutator alleles. Error bands provide bootstrapped 95% confidence intervals.

Restricted adaptive potential favors nonmutators in large populations. When adaptive potential is unlimited (left column), mutators are strongly favored in large populations. Right column shows fixation outcomes across population sizes with adaptive potential restricted to one beneficial mutation. Under the adaptation-restricted regime, nonmutators regain favor at very large population sizes. Top row shows results with 32 agents per PE (pop size size up to 23.9 million). Bottom row shows results with 2,048 agents per PE (population size up to 1.5 billion). Error bands indicate bootstrapped 95% confidence intervals.

256 agents per PE

2048 agents per PE

50/50 Treatment

hypermuts fix

–or–

normomuts fix

N

H

mix normo/hyper

H

N

H

N

N

H

H

H

H

H

H

H

H

H

H

H

N

N

N

N

N

N

N

N

N

N

N

De Novo Treatment

1 in 100k

all normomuts

N

H

N

N

N

N

N

N

–or–

normomuts fix

H

H

H

H

H

H

H

H

H

H

N

N

N

N

N

N

N

N

N

N

N

hypermuts fix

N