🚨🚨🚨Be part of the SC&IS Silver Jubilee Symposium at JNU on 27–28 Nov 2026! An excellent line of speakers (invited and alumni) & posters!🚨🚨🚨 Submit abstracts by 15 Nov and register by 20 Nov. (tinyurl.com/scis25jnu) Please share with your students! #SCIS25#JNU
1/2
Lo and behold the website has been launched for the symposium!!!
silver-jubilee.github.io/sci…
Thanks to SC&IS student Sougata Jana for designing it!
Deadline for abstract submission: 18th Nov 2026
🚨🚨🚨Be part of the SC&IS Silver Jubilee Symposium at JNU on 27–28 Nov 2026! An excellent line of speakers (invited and alumni) & posters!🚨🚨🚨 Submit abstracts by 15 Nov and register by 20 Nov. (tinyurl.com/scis25jnu) Please share with your students! #SCIS25#JNU
1/2
🚨🚨🚨Be part of the SC&IS Silver Jubilee Symposium at JNU on 27–28 Nov 2026! An excellent line of speakers (invited and alumni) & posters!🚨🚨🚨 Submit abstracts by 15 Nov and register by 20 Nov. (tinyurl.com/scis25jnu) Please share with your students! #SCIS25#JNU
1/2
Announcing LigParGen 3.0: OPLS-AA parameters for organic molecules, now a free app for Mac, Windows and Linux.
Paste a SMILES or draw a molecule, and get files for OpenMM, GROMACS, AMBER, CHARMM, LAMMPS and more.
jorgensenresearch.com/ligpar…
Our takeaway is NOT that anomaly detection has no place in AML. Rather, how we define an anomaly matters.
📄 Preprint: arxiv.org/abs/2610.39447
Comments and feedback welcome! 8/9
Why?
Blast and non-blast cells substantially overlap in the measured feature space.
RF also learned biologically meaningful signals: CD117 and CD34 emerged as the strongest contributors in classification.
So the model was not exploiting an arbitrary statistical pattern. 7/9
The real challenge appeared when blasts became rare.
In an external follow-up sample containing only ~1% blasts:
RF: F1 = 0.587
Anomaly detectors: F1 = 0.0006–0.194
Being rare, it turns out, not necessarily makes a biological population an easily detectable “anomaly.” 6/9
The result surprised us.
The supervised Random Forest clearly outperformed the anomaly-detection approaches.
Cross-validation: F1 = 0.911 AUROC = 0.994
External diagnostic sample (~41% blasts): F1 = 0.965 AUROC = 0.994
5/9
We tested:
🔹 Isolation Forest
🔹 One-Class SVM
🔹 Local Outlier Factor
🔹 Half-Space Trees
🔹 Random Forest
🔹 UMAP + HDBSCAN
Models were evaluated using patient-grouped nested cross-validation and then tested on independent samples. 4/9
This raises an interesting computational question:
If leukemic blasts are unusual cells within a much larger population of normal cells, can we simply treat them as anomalies?
We compared three different ML strategies on the same blast-identification problem. 3/9
Acute myeloid leukemia (AML) diagnosis and monitoring rely heavily on flow cytometry.
But identifying blast populations still involves expert manual gating, a process that can be time-consuming, subjective, and particularly challenging when blasts become rare. 2/9
🚨🚨New preprint alert! 🚨🚨
Can machine learning identify leukemic blasts from flow cytometry data?
And, since blasts are abnormal/rare populations, is anomaly detection actually better suited for the task than conventional supervised learning?
We put this idea to the test 1/9