Pinned Tweet
After 12 years at @PacBio, I'm proud to announce that I've been promoted to Fellow - moving back to a full-time individual contributor role to work with the great folks across the company!
I'm excited to finally have the space to explore the hard problems and even crazier ideas
Bioinformatics is fast again! 🚀
I might have just set a record: 1 min 46 sec to analyze a 30x @PacBio HiFi human genome (90 GB), from reads to SNVs, indels and structural variants. On a single server with one GPU.
That's an estimated 26 cents in the cloud.🧵
Preview builds, benchmarks, and the commands behind every number except the SV caller: pblab, our experimental playground at PacBio.
github.com/PacificBioscience…
Some of the optimizations went upstream too. The agents made hifiasm 2.9x and minimap2 1.5x faster on a whole genome, with identical output. Thanks to Haoyu Cheng and @lh3lh3 for building these incredible tools.
github.com/chhylp123/hifiasm…
github.com/lh3/minimap2/pull…
With a minimal CLAUDE. md and guards to validate outputs, I can talk to Opus like to a team of the best SWEs in my field. It will outperform any of us. Still needs plenty of steering that I'm willing to provide.
Max Q! I slowed down the pace to focus on memory usage:
- 40 seconds for unaligned CRAM (-10s)
- 7 GB RAM for unaligned BAM (-11GB)
Comparing against pbmm2 and the last checkpoint.
Pushing the frontier of bioinformatics is addictive. 2026Q4 shapes up to be banger!
Lift-off!
Aligning 30x HiFi HG002 in less than 60 seconds on 64 cores.
- 128c: 33s/59s un/sorted CRAM
- CRAM output reduces output size by 2.5x
- Better SNP, INDEL, SV F1s
- Lower peak memory
It's a PoC for now!
@lh3lh3 I'll start porting some of it to mm2 soon.
Developing new software in an established field requires constant output testing. Those gates must be FAST to not slow down the development cycle.
DeepVariant 40s, hap.py 3s, Truvari 3.5 minutes?
No longer! How about 19 seconds with identical results.
Lift-off!
Aligning 30x HiFi HG002 in less than 60 seconds on 64 cores.
- 128c: 33s/59s un/sorted CRAM
- CRAM output reduces output size by 2.5x
- Better SNP, INDEL, SV F1s
- Lower peak memory
It's a PoC for now!
@lh3lh3 I'll start porting some of it to mm2 soon.
And another experiment, re-implemented hap.py in C++ to get it from 16 min to 3 seconds. Identical results. What's next?!
How much faster can we map and align accurate long reads?
I benchmarked pbmm2 on 30× HiFi HG002 against GRCh38 using 64 CPU cores: ~17 minutes.
Can we get below 2 minutes while maintaining SNV, INDEL, and SV calling metrics?
mm2 was a jump. Is it time to leap?
How far we've come in bioinformatics!
Google won Truth Challenge in May 2016. Then Parabricks (later Nvidia) made it significantly faster. Now it's a commodity.
Cheers to everyone maintaining it for the past decade!
Now that it's so damn fast, can we make it better?
