Pierre-Antoine Bannier (@el_PA_B) header
Pierre-Antoine Bannier (@el_PA_B) profile picture

Research Scientist in Computational Biology and ML. Currently at @spc . @OwkinScience | @Polytechnique | @HECParis

France
Joined April 2012
Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)Photo by Pierre-Antoine Bannier (@el_PA_B)
sam3.cpp - Meta's SAM 3 in pure C++ with @ggerganov's ggml - Supports SAM 3.1, 3, 2.1, 2 and EdgeTAM - FP16, 4-bit quant (EdgeTAM in 15 MB) - Apple Metal GPU, CUDA, CPU - Text-prompted: "peach" → every peach - Single-file C++14 Performance-wise: - 100ms object detection, segmentation - Video object segmentation @ 20FPS on M4 Pro with EdgeTAM github.com/PABannier/sam3.cp…
13
118
6
870
65,537
Pierre-Antoine Bannier retweeted
If we're still in the fun part of the scaling law for protein structure... imagine where we are for structure-free discovery. Lots of low-hanging gains to be made, just need the right data
Five pharma companies just fine-tuned OpenFold3 in a federated setting on 20,000+ structures... ... raising correct ligand-pose predictions from 29% to 47%
1
4
5
1,213
Pierre-Antoine Bannier retweeted
Be more ambitious. Announcing SPC’s $575M Fund IV.
23
30
15
240
44,181
Pierre-Antoine Bannier retweeted
Introducing the world's fastest tokenizer implementation, Gigatoken! Gigatoken is ~500-1000x faster than HuggingFace, and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines. These baselines are already multithreaded Rust implementations! 🧵
84
479
95
4,123
742,509
Pierre-Antoine Bannier retweeted
Congratulations @NeehaZaidi @DanielHaldarMD @DrLizJaffee and team! This is the 1st study to demonstrate the feasibility of mutant KRAS targeted vaccine - not in established #PancreaticCancer - but in HIGH RISK individuals, with persistent T cells ~2years! aacrjournals.org/cancerdisco…
6
62
4
170
33,163
Pierre-Antoine Bannier retweeted
Le fait qu'il faille supprimer les " — " à cause de l'IA est un appauvrissement considérable de la langue française.......
44
176
36
4,033
387,973
Pierre-Antoine Bannier retweeted
One more remarkable PFS curve from #ASCO26 day one 🤯🤯🤯
Strong session for #gynonc at #ASCO26. #RUBY 4-year dMMR data: PFS 57.9% vs 15.7%, HR 0.30 (0.17–0.52). Four new events in 2.5additional years. Cure modeling: 54% curative potential. PFS at 1 year → 91.7%probability still PFS at 3 years. Alive at 1 year → 88.0% probability surviving to 3 years. The plateauis real and it holds. #NRGGY018 dMMR OS HR 0.56, RUBY 0.66. Not a drug difference -LIKELY poststudy ICI use. 93.2% of GY018 control patients received poststudy ICI vs 70.7% in RUBY. Morerescue in the control arm. Both trials say the same thing: first-line is the only clean window indMMR.I am waiting for KEYNOTE-C93 results to answer: does dMMR EC need the chemotherapy at all?And haven't started addressing POLE!Duration equally unresolved. The plateau establishes by month 12. Whether 36 months addsanything over 24 in complete remission is untested. What I do: dostarlimab + CP for dMMR EC.For UCS — dMMR or pMMR — dostarlimab + CP (10% in the RUBY). @NRGonc @ASCO
1
19
74
15,349
Pierre-Antoine Bannier retweeted
7yr CROWN update of 1L lorlatinib 55% progression-free at 7 YEARS. Median PFS STILL not reached. In metastatic solid tumors, this is essentially unheard of. Let that sink in. And hopefully, one day, this becomes the standard (or even better) for ALL metastatic cancers.#ASCO26 – at Chicago, IL
6
74
14
325
230,890
Precision medicine was supposed to kill cytotoxic chemo. Twenty years in, platinum is still the 1L backbone for lung, gastric, esophageal, cervical, ovarian, head and neck, and TNBC. The targeted-therapy revolution is mostly a lung adenocarcinoma story.
1
3
1,084
Platinum is also the original precision medicine. BRCA1/2 and broader HRD tumors are particularly platinum-sensitive: ovarian, prostate, pancreatic, breast. PARPi exploits the same biology. The genomic instability that breaks DNA repair is what makes platinum kill.
1
105
The only near-term displacement is ADCs. EV-302: enfortumab vedotin plus pembro beat gem-platinum in 1L urothelial on PFS and OS. First time platinum got knocked out of 1L in a non-driver tumor. DESTINY-Lung06 is now testing T-DXd+pembro vs platinum+pembro in HER2-OE NSCLC. Source: nejm.org/doi/full/10.1056/NE…
132
Most cancer therapies attack the tumor. PD-1 blockade does something different: it attacks the tumor’s relationship with the immune system. It can produce spectacular responses. But only in some patients. So the central question became: who is actually going to respond?
1
1
296
So the biomarker story of PD-1 blockade is really a story of expanding biological scale. PD-L1: is the brake present? TMB: is the tumor antigenic? MSI-H/dMMR: is the tumor hypermutated? Microbiome: is the host immune system permissive? Each marker explains part of the response. None explains all of it.
1
1
195
The future is not one magic biomarker. It is integrated prediction: tumor genomics antigen presentation PD-L1 / immune checkpoints T cell infiltration and state myeloid suppression microbiome prior treatments tumor type and tissue context
97