Moon Kim, Ph.D.

Founder & CEO, Karesis: AI for cancer genomics. Indel accuracy (F1): Karesis 0.99 vs Google DeepSomatic 0.88. Public benchmark, research use only. Ph.D. EE.

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard. We exceed it on some of the hardest cancer mutations to call: 0.99 vs 0.88 on small insertions and deletions (public benchmark, 5 cancer cell lines, higher on all 5).

Karesis is an AI company. Cancer genomics is where we started.

In 3 minutes: one AI engine for Illumina, PacBio, Oxford Nanopore and MGI, and the AI behind it: self-improving, agentic (an assistant built on Claude), auditable, and small enough to run on one notebook.

Labs and partners: try it on your own samples. Investors in AI × bio: we're raising our seed round. [email protected]

Research use only; not FDA cleared. Manuscript under review.

#AI #Genomics

[Moon Kim, Ph.D. @MoonKimPhD]

@PurdueRCAC Useful session. For labs running several sequencers, one thing we found: the same variant caller can serve Illumina, PacBio, Oxford Nanopore and MGI, so you don't maintain a pipeline per platform. Research software, research use only.

[Moon Kim, Ph.D. @MoonKimPhD]

@OmicsLogic Good topic. Small insertions and deletions are the hardest calls on long reads. On Oxford Nanopore data from a public benchmark of five cancer cell lines our indel F1 was 0.86 vs 0.83 for Google's DeepSomatic. Research use only.

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard. We exceed its performance: 0.99 vs 0.88 on small insertions and deletions.

For CROs and sequencing services: a plug-in analysis layer instead of one pipeline per sequencer 👇

[Moon Kim, Ph.D. @MoonKimPhD]

@eclecticV @boardyai Thanks @eclecticV for the kind words and for looping in @boardyai. I really appreciate it. If any sequencing teams want to try Karesis on their own data, my DMs are open.

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard.

On Oxford Nanopore data, we exceed its performance: indel F1 0.86 vs 0.83, ahead on 4 of 5 cancer genomes.

Long reads, explained in 90 seconds 👇

[Moon Kim, Ph.D. @MoonKimPhD]

Long reads can span a whole repeat, but a run of the same letter is still easy to miscount, so small insertions and deletions are the hardest changes to call.

On Oxford Nanopore data from a public benchmark of five cancer cell lines, our indel F1 is 0.86 vs 0.83 for DeepSomatic, ahead on 4 of 5 (only 1 is a confirmed held-out sample). PacBio HiFi: tied. Research use only; manuscript under review.

90-second explainer: https://t.co/j0F33bf1KJ

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard.

On Oxford Nanopore data, we exceed its performance: indel F1 0.86 vs 0.83, ahead on 4 of 5 cancer genomes.

Long reads, explained in 90 seconds 👇

[Moon Kim, Ph.D. @MoonKimPhD]

Sequencing platform teams: we're open to collaboration.

Try it on your own data: [email protected]

We're also raising our seed round. Investors and strategic partners: DMs are open.

Research use only; not FDA cleared.

[Moon Kim, Ph.D. @MoonKimPhD]

The numbers, stated plainly:
• Nanopore indel F1 0.8623 vs DeepSomatic 0.8308 (5-sample average, public CASTLE benchmark)
• Only 1 of 5 samples is a confirmed held-out sample
• PacBio HiFi: tied
• Methylation in the same model: early stage

[Moon Kim, Ph.D. @MoonKimPhD]

Try it on your own data: [email protected]

We're also raising our seed round. Investors and strategic partners: DMs are open.

Research use only; not FDA cleared.

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard. Same test, same cancer genomes: we exceed its performance (0.99 vs 0.88 on small insertions and deletions).

Now meet the platform behind that number 👇

[Moon Kim, Ph.D. @MoonKimPhD]

Honest scope:
• 11 model domains; 6 of 11 validated on real matched data, the rest in validation
• DeepSomatic is strong on tissue but doesn't read blood samples
• Guardant, Natera and Tempus sell clinical tests; GC-Platform is research software

[Moon Kim, Ph.D. @MoonKimPhD]

@eladgil One structural reason: in clinical genomics, proving a better method takes years of new samples and review, so tools improve slowly even when models jump. Open benchmarks plus fast validation on held-out samples would shorten that loop more than any single breakthrough.

[Moon Kim, Ph.D. @MoonKimPhD]

@patricksmalone Same shift in genomics: sequencing got cheap, so the bottleneck moved to calling variants correctly and validating them on samples the model never saw. The valuable data is expert-reviewed truth sets, not raw reads, and whoever owns that feedback loop keeps the margin.

[Moon Kim, Ph.D. @MoonKimPhD]

@dr_alphalyrae Agree on the closed loop. In genomics the bottleneck isn't model size, it's feedback: getting expert-reviewed calls back into training fast, and only promoting a new model when it beats the old one on held-out data. Make that loop cheap and biology speeds up.

[Moon Kim, Ph.D. @MoonKimPhD]

For the Biology & AI community: we compared our AI variant caller with Google's DeepSomatic on a public cancer benchmark. Indel F1 0.99 vs 0.88 (research use only; manuscript under review).

60-second explainer below. Questions and pushback welcome.

https://t.co/dXL6C8fnQw

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard for finding cancer mutations.

On small insertions and deletions, some of the hardest to call, we exceed its performance: 0.99 vs 0.88 on a public benchmark of 5 cancer genomes.

What that means, in 60 seconds 👇

[Moon Kim, Ph.D. @MoonKimPhD]

Try it on your own data: [email protected]
We're also raising our seed round. Investors and strategic partners: DMs are open.
Research use only; not FDA cleared.

[Moon Kim, Ph.D. @MoonKimPhD]

Google's DeepSomatic sets the standard for finding cancer mutations.

On small insertions and deletions, some of the hardest to call, we exceed its performance: 0.99 vs 0.88 on a public benchmark of 5 cancer genomes.

What that means, in 60 seconds 👇

[Moon Kim, Ph.D. @MoonKimPhD]

The numbers, stated plainly:
• Indel F1: Karesis 0.9902 vs DeepSomatic 0.8774 (5-sample average, public CASTLE benchmark, Illumina)
• Single-letter changes: on par
• Only 1 of 5 samples is a confirmed held-out sample
• Manuscript under review