alex rubinsteyn

Genomics + immunology + ML = personalized cancer immunotherapy | |

[alex rubinsteyn @iskander]

The thing I feel most about coding agents is tremendous gratitude. Every half-finished project that has haunted me for a decade gets to actually come to life

[alex rubinsteyn @iskander]

@RuxandraTeslo @BenjaminGVincen @WorksInProgMag I’ll give you the update if you want

I’m like 30% less skeptical now than before

[alex rubinsteyn @iskander]

@sokrypton @design_proteins I think this is what both folks did but I’ll dig into methods details later this week

But you agree that 2% seems low?

[alex rubinsteyn @iskander]

Got some preliminary results back from Adaptyv and...it seems like de novo binder design is still hard?

[alex rubinsteyn @iskander]

@sokrypton @design_proteins We weren’t looking to validate a new method though, just try a few different ways to get binders on demand. I’m surprised so few worked! Do you have a better/recommended approach?

[alex rubinsteyn @iskander]

@sokrypton @design_proteins An awful lot, maybe 10K+ filtered in different ways.

Going to get the two designers to tag the designs with methods so I can make sense of why even a handful worked (hopefully there’s some pattern)

Or maybe ecCD45RO is just hard?

[alex rubinsteyn @iskander]

@SynBio1 @design_proteins Yes! Seen lots of claims of 20%+ from newer tools and pipelines

[alex rubinsteyn @iskander]

@joe_fenrir @DrSamuelBHume I’m sure you could with great effort create placebo infusion schedules to blind people taking an oral drug vs mostly infusion control, but it adds a ton of complexity to a trial so not sure anyone ever does it (esp when the control arm has different treatments for every patient)

[alex rubinsteyn @iskander]

@joe_fenrir @DrSamuelBHume Bc they’d be criticized for any specific choice of control arm, they are all known to be bad at this point in the disease trajectory. So it’s common to sweep that choice into “maybe the treating oncologists will have better ideas than us”

[alex rubinsteyn @iskander]

@design_proteins High in silico metrics (eg ipSAE mostly in the 0.75-1.0 range)

[alex rubinsteyn @iskander]

@Ed_Harris_ Got two different people to make mixed libraries from a few different methods, we’ll do a writeup blog post once the full Adaptyv data is finished

[alex rubinsteyn @iskander]

Any interesting ASCO results in TNBC?

[alex rubinsteyn @iskander]

@design_proteins CD45RO (shortest tail isoform) but some of the ~150 designs targeted shared pan-CD45 regions

Only 3 seemed to bind! ~20 still pending

[alex rubinsteyn @iskander]

@RuxandraTeslo @OmicsOmicsBlog @ScienceMagazine We see a suspicious number of germline structural variants in the BL/6 mice we use vs a reference from the same line

[Heng Li @lh3lh3]

Jeremy Wang developed rammap, a minimap2 rewrite in Rust. It achieves comparable or better performance than minimap2 and produces identical output to minimap2. During rewrite, Jeremy found two long-existing bugs in minimap2 which are fixed in v2.31. https://t.co/9FNiwgbgqW

[Sergey Ovchinnikov @sokrypton]

Updating my diffusion slides to better explain binder design for the general audience. 🐶🐱

[alex rubinsteyn @iskander]

@BenjaminGVincen Ha! I’m working very hard on seeing the utility, original motivation, or (at worst) the explanatory incentive structure in all the weird corners of bureaucratized biomedicine.

[owl @owl_posting]

if you enjoyed this article + know anyone who works at a zombie biotech liquidator fund (Tang Capital/XOMA/etc), i would love love love love an introduction and will buy you dinner in exchange

[Zeming Lin @ebetica]

The ESMC embeddings are incredible, especially for non-protein biomolecules where MSAs do not help much. The single sequence folding there is state of the art even amongst MSA conditioned models

[Matthias Mann Lab @labs_mann]

4/5
Cell type resolution exposes biology bulk profiling misses. We find cancer-testis antigens, proteins normally restricted to immune-privileged sites and actively targeted by immunotherapies, in oocytes of the female germline. Worth noting for CTA-directed therapy safety.

[Sergey Ovchinnikov @sokrypton]

@anshulkundaje So is esmfold2 a "world" model? On one 🫱, it's uses evolution stats from the pLM, so no. But on the other 🫲 the PPI results from inference-time-scaling (multiple seeds, select w/ confidence) suggests it did learn an energy function and limited search in the "folding" module? 🤔