Sign in

Dhuvi Karthikeyan

@dkarthikey1.bsky.social
774 followers 341 following 39 posts

Shape rotator and vibes curator at UNC’s Personalized Immunotherapy Research Lab. Generative modeling and representation learning in biology. 👨‍🏫👨‍🔬

PostsRepliesMedia
Reposted by Dhuvi Karthikeyan
Eric Kernfeld @ekernf01.bsky.social · 12/10/2025
This is so friggin cool. You don't have to rewrite all that FORTRAN to do autodiff on it. "Enzyme"
041
Reposted by Dhuvi Karthikeyan
Kilian Schober @kischober.bsky.social · 08/10/2025
When a T cell carries both its natural TCR and an engineered CAR, do these two receptors cooperate - or compete? Our new preprint on bioRxiv shows that TCR activation can enhance CAR-T activation but impairs CAR-mediated killing when targets are on separate cells. www.biorxiv.org/content/10.1...
biorxiv.org
TCR activation impairs CAR-T cytotoxicity against separate target cells
Chimeric antigen receptor T cells (CAR-T) are effective therapeutics against cancer and autoimmunity, but whether the endogenous T cell receptor (TCR) is beneficial, detrimental or irrelevant for CAR ...
2145
Reposted by Dhuvi Karthikeyan
bioRxiv Microbiology @biorxiv-microbiol.bsky.social · 10/09/2025
No Distinct Gut Microbiome Signature in Alzheimer's Disease: A Reanalysis with Modern Tools www.biorxiv.org/content/10.1101/202…
063
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 09/09/2025
Thanks Professor!
000
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 09/09/2025
Now out in @natmachintell.nature.com TCRT5 is a rapid generator of target-conditioned CDR3b, leads SoTA, and yields the first AI-designed self-tolerant binder to an OOD non-viral epitope (w val) 📑: www.nature.com/articles/s42... 🤗: huggingface.co/dkarthikeyan1 👨‍💻: github.com/pirl-unc/tcr_translate
142
Reposted by Dhuvi Karthikeyan
Alex Rubinsteyn @alexr.bsky.social · 09/09/2025
tl;dr T5 style translation model to generate T cell receptors from targets actually kinda works Caveats: not full TCR (CDR3b), strong signal that public data biased towards “polyspecific” TCRs — lo and behold, generated TCR-T is polyspecific in vitro. Open weights, go wild.
193
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 09/09/2025
This story is nothing without its reviewers 🩵
010
Reposted by Dhuvi Karthikeyan
Alex Rubinsteyn @alexr.bsky.social · 09/09/2025
With some key contributions by Sarah Bennet in actually turning predicted TCRs into TCR-Ts and testing them! Hopefully this raises the bar on TCR generation evals, I think we the best in silico metrics and 1 in 20 TCRs just kinda works.
041
Reposted by Dhuvi Karthikeyan
Jamie Heather @jamieheather.bsky.social · 08/09/2025
This is a cool story that I've watched with great interest - making big steps in the field of de novo TCR generation! Great work @dkarthikey1.bsky.social / @benjamingvincent.bsky.social / @alexr.bsky.social et al
2104
Reposted by Dhuvi Karthikeyan
AlphaFold Unofficial @alphafold.bsky.social · 19/05/2025
Today’s news about Joe Biden makes it almost difficult not to reflect on the unrealistic promises from AI leaders who claim AI could end disease and usher in radical abundance. Disease will persist. AI is a tool, not a cure-all. People deserve truth, not hype. www.cbsnews.com/amp/news/art...
cbsnews.com
Artificial intelligence could end disease, lead to "radical abundance," Google DeepMind CEO Demis Hassabis says
At Google DeepMind, researchers are chasing what's called artificial general intelligence: a silicon intellect as versatile as a human's, but with superhuman speed and knowledge.
0194
Reposted by Dhuvi Karthikeyan
Dustin Rubenstein @dustinrubenstein.com · 10/05/2025
NSF to ONLY fund research in five areas: artificial intelligence, quantum information science, biotechnology, nuclear energy, and translational science. Can this be happening? www.science.org/content/arti...
science.org
NSF slashes number of ‘rotators’ and well-paid managers as part of restructuring
Smaller future budgets will require fewer people, NSF official tells staff
23257145
Reposted by Dhuvi Karthikeyan
kyunghyuncho.bsky.social @kyunghyuncho.bsky.social · 08/05/2025
what do you teach at the first-year graduate course on machine learning, in this era of LLM and large-scale compute? here's my experiment on answering this question: let's teach everything that admits SGD and that is not LLM, and ask students to read old papers.
1423
Reposted by Dhuvi Karthikeyan
George Takei @georgetakei.bsky.social · 03/05/2025
We hadn’t won World War II by May 8th. I know because my family and I were still in an internment camp in Tule Lake, California.
2179310836395
Reposted by Dhuvi Karthikeyan
Carolien van de Sandt @cvandesandt.bsky.social · 24/04/2025
Your immune system gradually loses it top team players when we get older 👱➡️👵 Our study in @ebiomedicine.bsky.social investigates how long good immune cells stay around – a 🧵⬇️ go.unimelb.edu.au/4sep @katherinekedz.bsky.social @thedohertyinst.bsky.social #Influenza #AgingResearch #Immunity #TCR
12113
Reposted by Dhuvi Karthikeyan
Shobita Parthasarathy @shobitap.org · 21/04/2025
My PhD student's NSF grant was terminated today. Her dissertation investigates why air pollution monitoring, metrics, and regulations don't properly capture the harms experienced by marginalized communities--esp in the most polluted Detroit neighborhoods. She and I are happy to talk with press etc.
nsf.gov
NSF Award Search: Award # 2240343 - Constructing Credible Knowledge and Expertise in Air Pollution Regulation and Monitoring: The Problem of Quantification Lock
452194948
Reposted by Dhuvi Karthikeyan
Amy Pomeroy @amypomeroy.bsky.social · 10/04/2025
I am so excited to share that our new paper is out! A model of intratumor and interpatient heterogeneity explains clinical trials of curative combination therapy for lymphoma Out in Blood Cancer Discovery doi.org/10.1158/2643-3230.BCD-24-0230 #mathonc #lymsm #BloodCancer 1/n
doi.org
A model of intra-tumor and inter-patient heterogeneity explains clinical trials of curative combination therapy for lymphoma
Abstract. Models of tumor drug response have illuminated important concepts in oncology, but there remains a need for theory that combines intra-tumor and inter-patient heterogeneity to explain patien...
174
Reposted by Dhuvi Karthikeyan
Gisela Gabernet @ggabernet.bsky.social · 31/03/2025
Do you analyze large BCR and TCR sequencing datasets? Answer our survey for improving nf-core/airrflow, a Nextflow pipeline to analyze bulk and single-cell AIRRseq data (doi.org/10.1371/jour...). We’re interested in your opinion on useful new features! yalesurvey.ca1.qualtrics.com/jfe/form/SV_...
doi.org
nf-core/airrflow: An adaptive immune receptor repertoire analysis workflow employing the Immcantation framework
Author summary We have created nf-core/airrflow, a workflow to help researchers study the immune system in healthy and disease states, such as infections, autoimmunity, and cancer. The adaptive immune...
0105
Reposted by Dhuvi Karthikeyan
Eric Topol @erictopol.bsky.social · 24/03/2025
HHS's Office of Long COVID Research and Practice has been DOGEd. Closing this week.
37463512906
Reposted by Dhuvi Karthikeyan
Nathan C. Frey @ncfrey.bsky.social · 26/02/2025
We @prescientdesign.bsky.social Genentech pre-printed our "Lab-in-the-loop for therapeutic antibody design." We built a general ML system to accelerate molecule design for challenging, therapeutically relevant targets. www.biorxiv.org/content/10.1...
biorxiv.org
1238
Reposted by Dhuvi Karthikeyan
jp flores (he/him) @jpflores.rbind.io · 25/02/2025
You in the Triangle area and thinking about standing up for science? Come to my beloved 2nd workplace and make a sign 😊🍻 supplies provided! You can leave a sign w me and i can bring it up to DC! @standupforscience.bsky.social
35519
Reposted by Dhuvi Karthikeyan
Jamie Heather @jamieheather.bsky.social · 25/02/2025
Dear TCR researchers of #immunosky, you may be interested to know that our tool Stitchr - for the automated production of full-length T cell receptor sequences - has just had a nice little update: jamieheather.github.io/stitchr/
jamieheather.github.io
stitchr — stitchr documentation
2367
Reposted by Dhuvi Karthikeyan
Darby Saxbe @darbysaxbe.bsky.social · 04/02/2025
🚨BREAKING. From a program officer at the National Science Foundation, a list of keywords that can cause a grant to be pulled. I will be sharing screenshots of these keywords along with a decision tree. Please share widely. This is a crisis for academic freedom & science.
list of banned keywords
12652768615653
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 01/02/2025
Arguably well*
000
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 01/02/2025
Musks method of ruthlessly enforcing efficiency by removing things and adding them back if the system breaks has worked well for pruning complexity of products, pipelines and entire companies. But applying this methodology to govt is dangerous and will result in greater harm to the American people
110
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 27/01/2025
Tagging in @pranam.bsky.social
010
Reposted by Dhuvi Karthikeyan
Eric Topol @erictopol.bsky.social · 12/01/2025
Good combo! A nasal vaccine with self-amplifying RNA (much less dose required for v potent immune response), complete protection effective vs both H5N1 and H7N9 viruses in experimental models www.biorxiv.org/content/10.1...
biorxiv.org
An intranasal, NLC-delivered self-amplifying RNA vaccine establishes protective immunity against pre-pandemic H5N1 and H7N9 influenza
Seasonal and pandemic influenzas are continuous threats to human health, requiring rapid development of vaccines to multiple evolving viral strains. New RNA vaccine technologies have the adaptability ...
19551114
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 30/12/2024
Ahahaha I think my brain autocorrected LLM to pLM and selectively dropped for natural language when I saw your skeet. Still a cool approach if you would like to be the first to show that it works and is useful for natural language tasks 😇
010
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 30/12/2024
www.biorxiv.org/content/10.1...
biorxiv.org
Aggregating Residue-Level Protein Language Model Embeddings with Optimal Transport
Protein language models (PLMs) have emerged as powerful approaches for mapping protein sequences into embeddings suitable for various applications. As protein representation schemes, PLMs generate per...
130
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 30/12/2024
@rohitsingh8080.bsky.social and co have a really cool paper out on optimal transport for pooling embeddings for pLMs. Worth checking out forsure
110
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 30/12/2024
On the merits of reading papers, attending conferences, preprinting manuscripts… Overseen at the Library of Congress.
000
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 21/12/2024
Unsure if people on bsky are aware of o3s existence. But if anyone wants to talk I’ve been itching all day
010
Reposted by Dhuvi Karthikeyan
Q-Immuno lab at UCL @qimmuno.bsky.social · 19/12/2024
How to learn generalizable rules across complex sequence-function maps? 💡In our new preprint we propose a framework for learning two-point statistics and use it to discover biophysical rules of TCR co-specificity that generalize to unseen ligands ! ✨ arxiv.org/abs/2412.13722
Intuition behind why two-point statistics might provide an easier path to generalization
095
Reposted by Dhuvi Karthikeyan
Diego del Alamo @delalamo.xyz · 11/12/2024
Ideal test is antibody CDRH3s which often conformationally rigidify during affinity maturation. MSAs only report inter-species evolution, not intra-species maturation, so existing methods fail. Pics from doi.org/10.3389/fimmu.2018.03065 & doi.org/10.3389/fmolb.2020.00182
Decrease in conformational entropy of an antibody between germline/naive sequence (left, green) and affinity-matured sequence (yellow, right)2D free energy landscape before and after affinity maturation. Features are sin/cos of CDRH3 backbone dihedrals
0102
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 10/12/2024
Huge fan of SEDD for DNA sequences! Congratulations again on a well thought out and well executed paper :)
010
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 04/12/2024
Given the subtext here I wouldn’t be surprised if both X and Bsky develop extremely close but distinct niches: the former to blast information and reach broader audiences, and the latter as a place of discourse with members of a particular community.
021
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 04/12/2024
🧙🪄Summoning Alex to 💙🦋 so the Bio x ML community can celebrate ESM-C with him on here as well
030
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 02/12/2024
Also worth checking out: academic.oup.com/bioinformati... TLDR: performance gains from tokenization depend largely on dataset composition and specific task (complexity?).
020
Reposted by Dhuvi Karthikeyan
Jamie Heather @jamieheather.bsky.social · 26/11/2024
My usual go to is VDJdb. Set antigen species to human, minimal confidence score to 3, and paired A+B. It's not perfect, but it filters out most of the cruft. There's also a new DB called TRAIT which has a nice pack of TCRs that are being/have been trialed, which are mostly against human epitopes.
171
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 26/11/2024
CA Coworking 🌉🌅
000
Reposted by Dhuvi Karthikeyan
Sander Wuyts, PhD 🧬 @swuyts.bsky.social · 21/11/2024
Calling all T-cell aficionado's. Come and join us next year (27-28 May 2025) for the fourth Antwerp T-Cell Receptor conference organised by @pmeysman.bsky.social, @krislaukens.bsky.social and Benson Ogunjimi. I'll also be there! www.uantwerpen.be/en/conferenc... #ImmunoSky
12512
Reposted by Dhuvi Karthikeyan
Mohammed AlQuraishi @moalquraishi.bsky.social · 20/11/2024
#GoodThingsComeToThoseWait Sorry it's taking us a while. Promise that the wait will be well worth it! #OpenFold3 Love all the great work that's already been done #HelixFold3 #Ligo #Chai1 #Protenix #Boltz1
311926
Reposted by Dhuvi Karthikeyan
Dr. Glaucomflecken @glaucomflecken.bsky.social · 19/11/2024
Surely I misheard. There’s no way Dr. Oz, the OG grifter, president of grift university, has been tapped to lead the largest and most influential US healthcare entity.
1782304312
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
Huge thank you to everyone who made TCR-TRANSLATE possible. Thank you Colin-learning LLMs from the creator of T5 helped in more ways than I can count. Thank you @PIs for pushing a deep understanding of the models. Thank you everyone @the conferences whose feedback made our work that much better <3
000
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
So where does that leave us? TCR-TRANSLATE is just the first step on an orthogonal path to solving TCR:pMHC cross-reactivity. A ton of work remains, including wet lab validation, so please reach out if you'd like us to test your favorite pMHC/otherwise collaborate! www.biorxiv.org/content/10.1...
biorxiv.org
TCR-TRANSLATE: Conditional Generation of Real Antigen Specific T-cell Receptor Sequences
The paradoxical nature of T-cell receptor (TCR) specificity, which requires both precise recognition and adequate coverage of antigenic peptide-MHCs (pMHCs), poses a fundamental challenge in immunolog...
122
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
TCRT5 even generates better than random sequences against the notoriously challenging IMMREP2023 "private" antigen set, and gets 1/12 known sequences for one of them. (PS - Thank you @Giulio (not on here, yet) for a brilliant model of V(D)J recombination and thymic selection in soNNia)
100
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
This key finding helped us choose our flagship TCRT5-FT model which we evaluate with a more qualitative lens. We find that TCRT5: >preferentially samples sequences with high biological pGen >generates antigen-specific sequences to unseen epitopes >generates sequences not seen during training
100
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
POLYSPECIFIC TCRS! It turns out the bidirectional models sample more empirically de-risked CDR3b sequences, or sequences that appear a lot in the training data and have been shown to bind highly dissimilar inputs. Think super-sequences, with high empirical utility.
100
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
🗣️ @Protein Designers, something for yall: We tried bi-directional and multi-task translation (target-conditioned receptor generation and receptor-conditioned target generation) to see if self-consistency improves the sequence quality for TCR generation, and we found something rather interesting...
100
Dhuvi Karthikeyan @dkarthikey1.bsky.social · 19/11/2024
TCR-TRANSLATE frames this is a seq2seq task (w tricks from low-resource machine translation to handle sparsity). Given a target of interest, can we generate a faithful conditional distribution over cognate sequences and sample from it to get an epitope-specific TCR repertoire?
100