Rohit Singh @rohitsingh8080 Twitter profile

Pinned Tweet

Rohit Singh

4 days

Let me tell you a story. It'll end up at the current tech-bio and protein design scene. But the story starts about 25 years earlier. Did you know that, commercially, the human genome project precipitated the end and not the start of a genomics boom? 1/

9

72

437

Last Seen Profiles

@CryptoArix8

@RoundtableSpace

@Matheus_ofc21

@MommaCyune

@imranfyi

@HysholarG

@w93460

@ziko14

@TheBitJournal_

@ekinnkirac

@obs1on

@Meowny2

@THEHUMANPHASE

@Rdslfootball

@EclpsX

@CaseyTafel30123

@ice_theburg

@djordjevanjek

@Atrarad

@Suika_10382

@ItsOLY

@Ahmed56039340

@saicruera

@hvingsxx

@NedraD55547

@bokeplokalmalam

@8l8TGj59N4VR8B

@G_Airou

@AwulachewTesse2

@drew_2828

@RealTimeVideo

@KmalAlahdl

@GoblinGame

@yasar_gunacgun

@hotbokep9942

@robotjason

Rohit Singh

@rohitsingh8080

1 month

De novo protein design is great, but nature has millions of proteins- why not repurpose them? Introducing Raygun, a new approach to protein design. It allows you to miniaturize, magnify or modify any protein. We synthesized miniaturized variants of eGFP and mCherry! 1/

27

227

1K

Rohit Singh

@rohitsingh8080

4 years

This. Save the business and the employees, but wipe the equity.

Bill Gurley

@bgurley

4 years

Let's be candid. If you believe in business & capitalism, then there are zero circumstances where the government should bail out equity holders. $GM and $GS were mistakes. If the gov't is the lender of last resort, they should own all of the equity. I invest in equity as my job.

185

736

5K

9

39

392

Rohit Singh

@rohitsingh8080

1 year

Thrilled to share that I will be starting this Fall as an Asst Professor at @DukeU , jointly in Cell Biology and Bioinformatics & Biostats. We'll apply ML to understand interactions across scales, from protein-protein and protein-drug to cell-cell 1/

32

8

151

Rohit Singh

@rohitsingh8080

1 year

🧪🔬Protein language models are powerful, but they struggle with antibodies, especially their hypervariable regions. Meet AbMAP, our approach for fine-tuning PLMs & boosting prediction accuracy on antibodies!🔓💉 📃:

1

28

129

Rohit Singh

@rohitsingh8080

3 months

Have been thinking about ESM-3. Lots to like, some not so much. With a mug of station-wali chai on a drizzly Saturday morning, some big-picture thoughts. This 🧵won't be about all the method details, read @pranamanam or @anthonygitter 's threads for that. 1/n

4

22

106

Rohit Singh

@rohitsingh8080

4 months

Discover the drivers of gene regulation in your scRNA-seq data with Velorama. Exploiting RNA velocity, Velorama infers causal links between transcription factors and target genes, even in complex branching trajectories. 🌿 Now out in @CellSystemsCP 📜 #scRNAseq #GeneRegulation

5

26

100

Rohit Singh

@rohitsingh8080

3 years

TADs seem to play a key role in regulating transcription. By an integrative multimodal analysis spanning many cell types, @lab_berger and I wondered if we could learn how the TAD structure in a species generally relates to gene expression. 🧵

1

14

85

Rohit Singh

@rohitsingh8080

8 months

Say you're predicting something about a protein from its sequence. Starting from a variable-length input and going to a fixed-length output, you'll need to aggregate somewhere. You were planning to average-pool, perhaps? May we suggest an alternative?🧵

1

7

58

Rohit Singh

@rohitsingh8080

11 months

What does flight-path planning have to do with understanding cell fate? Curvature! The plane’s motion and the cell’s state both need to be modeled over a manifold. We introduce Sceodesic, which unlocks the differential geometry of gene coexpression 🧵:

1

7

52

Rohit Singh

@rohitsingh8080

1 year

This popped up in my Google News feed. What a happy surprise! Thanks for the coverage, @Marktechpost !

MIT Researchers Propose AbMAP: A Protein Language Model (PLM) Customized For Antibodies

Some of the most promising medication candidates in current therapies have been antibodies. The incredible structural diversity of antibodies, which enables them to recognize an incredibly broad...

www.marktechpost.com

0

5

43

Rohit Singh

@rohitsingh8080

1 month

The Raygun preprint is here:

Miniaturizing, Modifying, and Augmenting Nature’s Proteins with Raygun

Proteins are nature’s versatile nanomachines, but engineering them for enhanced function or novel applications remains challenging. Current methods for protein modification struggle to design...

www.biorxiv.org

2

1

42

Rohit Singh

@rohitsingh8080

3 years

Excited to share Schema, our tool for synthesizing multi-modal single-cell data. Schema enables fast exploratory analysis, is robust to noise, and helps generate informative visualizations. With @BrianHie , @ashwinn226 & Bonnie Berger. Tweetorial 👇 1/

Schema: metric learning enables interpretable synthesis of heterogeneous single-cell modalities -...

A complete understanding of biological processes requires synthesizing information across heterogeneous modalities, such as age, disease status, or gene expression. Technological advances in single...

genomebiology.biomedcentral.com

1

13

40

Rohit Singh

@rohitsingh8080

3 years

Causal understanding of gene regulation is a foundational goal in biology. In an ICLR 2022 paper with Alex Wu and @lab_berger , we show how multimodal single-cell technologies like 10x Multiome, SHARE-Seq or SNARE-Seq can help get us there.

Granger causal inference on DAGs identifies genomic loci regulating...

When a dynamical system can be modeled as a sequence of observations, Granger causality is a powerful approach for detecting predictive interactions between its variables. However, traditional...

openreview.net

1

10

35

Rohit Singh

@rohitsingh8080

2 years

Tragically tangled trajectories in your single-cell study stressing you out? Maybe give Velorama a whirl. Proudly presenting Velorama for inferring regulatory mechanisms from RNA velocity.

Causal gene regulatory analysis with RNA velocity reveals an interplay between slow and fast...

Single-cell expression dynamics from differentiation trajectories or RNA velocity have the potential to reveal causal links between transcription factors (TFs) and their target genes in gene regula...

www.biorxiv.org

2

14

32

Rohit Singh

@rohitsingh8080

1 month

This week marked my 1-year anniversary at Duke– I feel so lucky to be here! In the next few days, I hope to talk about some things we learned from this project– about protein design, language models, and their capabilities. Stay tuned... 12/

1

32

Rohit Singh

@rohitsingh8080

4 years

@gabemott @bgurley It's the only way to have management do better next time. If shareholders feel the pain, they'll select better managers in the next iteration. All the other penalties (buybacks, bonuses etc.) likely won't work. And I say this as someone who'd take a hit if the equity was wiped.

0

2

29

Rohit Singh

@rohitsingh8080

2 years

📷 Modern phones can take photos with fantastic depth perception. They exploit parallax, the offset between simultaneous-but-separate snapshots, to infer depth (2D ➡️ 3D). Turns out, we can discover cellular mechanisms from similar snapshots of cells! 🧵

1

7

29

Rohit Singh

@rohitsingh8080

1 month

How does it work? A new, non-diffusion approach. Using protein language models, we represent any protein as a 64,000-dimensional multivariate normal distribution. We directly sample in a single shot, and a 700-million parameter encoder-decoder architecture does the rest. 8/

2

29

Rohit Singh

@rohitsingh8080

15 days

🚨Hiring alert! @SoderlingLab and I are recruiting a postdoc we'll jointly mentor. Scott and I have broad, shared interests at the intersection of neurobiology, proteomics and PLMs, and we'll happily follow your lead on specific problems. Pls retweet! More details below...

1

13

28

Rohit Singh

@rohitsingh8080

1 year

First lab meeting of our research group @DukeBiostats and @DukeCellBiology . Excited to dive into fun science adventures with collaborators, new and old.

3

1

26

Rohit Singh

@rohitsingh8080

2 years

Job alert 🚨! The Bonnie Berger lab @lab_berger at MIT CSAIL is hiring for a software engineer role with funding and benefits. Perfect for fresh college grads or folks in industry looking to transition to academia/research. Email me and Bonnie if interested! Pls retweet! 1/n

3

10

24

Rohit Singh

@rohitsingh8080

1 month

What can you do with Raygun? A whole ton of stuff-- add loops, remove them, miniaturize sensors, fit proteins into AAVs and delivery platforms. The parameters allow for fine-grained control: minor tweaks or major alterations, your choice. All delivered speedily. 7/

1

24

Rohit Singh

@rohitsingh8080

2 months

Bryan Bryson @TheBrysonLab giving a fantastic talk on why we don't yet have a great TB vaccine and how ML can help. @AccMLBio workshop

0

3

23

Rohit Singh

@rohitsingh8080

1 month

Looking forward to visiting the School of Information Technology at @iitdelhi tomorrow. I'll be giving a talk (3pm, Room: SIT 001) about protein language models, the enormous promise they offer, and some of our work in the area. Excited to talk science with folks!

0

2

21

Rohit Singh

@rohitsingh8080

1 month

Fun fact: 23% of human proteins are shorter than eGFP and mCherry. When you tag a small protein with a large reporter, might the tail wag the dog? With Raygun, we miniaturized eGFP and mCherry. Two of the generated mCherry variants shrank by 37 and 20 residues! 3/

1

21

Rohit Singh

@rohitsingh8080

1 month

We're working to put the code out, as well as pretrained weights. It will all be at

GitHub - rohitsinghlab/raygun: Template-based protein design with Raygun

Template-based protein design with Raygun . Contribute to rohitsinghlab/raygun development by creating an account on GitHub.

github.com

1

19

Rohit Singh

@rohitsingh8080

1 month

Caveat: our generated proteins are dim. We’re working on that. ESM-3 folks also found that their initial sample was dim and needed to be optimized. We discuss in the preprint that design methods might work best in conjunction with techniques like directed evolution 6/

2

0

20

Rohit Singh

@rohitsingh8080

2 years

At ISMB 2022, we’ll present Topsy-Turvy. It predicts PPIs by integrating a graph-theoretic view into a protein language model approach. Jointly with Kapil Devkota, @samsledzieski , @lab_berger and Lenore Cowen. #ISMB2022 🧵

Topsy-Turvy: integrating a global view into sequence-based PPI prediction

AbstractSummary. Computational methods to predict protein–protein interaction (PPI) typically segregate into sequence-based ‘bottom-up’ methods that infer

academic.oup.com

1

3

18

Rohit Singh

@rohitsingh8080

1 month

How to use Raygun: input a protein sequence and specify a target length and noise parameter. Target length controls indels, noise controls substitutions. Raygun will generate for you (< 1 sec/sample) a protein with those specifications, structurally similar to the original. 2/

1

3

19

Rohit Singh

@rohitsingh8080

2 months

Excited that CZI will be supporting our work with @Tatalab_Duke on decoding the drivers of spatio-temporal differentiation. Interested postdocs and students, join us! We started on this path with our DAG-based Granger causality work. There's so much more to be done! 1/

CZI Science

@cziscience

2 months

Deep learning advancements change how we address challenges & bottlenecks in #SingleCell biology. See how researchers are transforming what's possible using existing data to understand health + disease with cutting edge computational approaches

2

10

27

6

2

19

Rohit Singh

@rohitsingh8080

1 year

Apply to our joint search of @DukeCellBiology and @DukeBiostats ! I was recruited under this search, and I'd love for you to be my colleague! A brief thread on why I'm so excited about AI in Biology at @DukeU 🧵. Also happy to answer questions over DM/email!

Duke Department of Cell Biology

@DukeCellBiology

1 year

📢 Faculty Search at Duke University! We are seeking tenure-track faculty candidates who are driving innovation in deep neural network algorithms for biology! #ISMB #NeurIPS2023 #ComputationalBiology

1

35

26

1

13

18

Rohit Singh

@rohitsingh8080

1 year

Nice thread overviewing some recent work in scRNA-seq representation learning! While these works are extremely creative and will be very impactful, I wonder if overly relying on the terminology of language models might be counterproductive here. 1/

Will Connell

@wilstc

1 year

🧬🔮 Single cell foundation models have been a recent hot topic in bio-ML! A few of the recent methods and some thoughts 🧬🔮 1) Geneformer 2) scGPT 3) scFoundation 4) Exceiver

4

58

252

1

0

17

Rohit Singh

@rohitsingh8080

1 year

Super-fast, accurate computational screening of large compound libraries using protein language m our ConPLex paper is now out in @PNASNews . Joint work w @samsledzieski @lab_berger @TheBrysonLab and Lenore Cowen.

Sam Sledzieski

@samsledzieski

1 year

Our paper on language models for drug-target interaction is out now in PNAS!

0

8

24

0

4

16

Rohit Singh

@rohitsingh8080

1 year

The 60th anniversary of @MIT_CSAIL was a geekfest, with many cool talks. Here's my bit on how to "speak" protein. It's a fast-track intro to the power of protein language models.

Using Protein Language Models For Drug Discovery

Rohit Singh is a Research Scientist at CSAIL at MIT. His interests broadly lie in the area of computational biology. He is especially excited about leveragin...

www.youtube.com

0

3

15

Rohit Singh

@rohitsingh8080

9 months

The NYT-OpenAI lawsuit reminded me of a looming issue in training biological foundation models: data-sharing while respecting privacy and commercial concerns. 🧵

1

15

Rohit Singh

@rohitsingh8080

9 months

I forgot to take a photo during his talk, so this will have to do. @benjraphael gave a fantastic talk at @DukeCellBiology on the creative and elegant spatial transcriptomics methods from the Raphael Lab. Check out Gaston, Belayer, PASTE2 and the rest!

1

0

14

Rohit Singh

@rohitsingh8080

1 month

And all we did was specify a target length + noise! No scaffold or chromophore was specified. In fact, the shortest functional variant does not even have the native chromophore! Raygun lets you start from a natural protein and build something dramatically different. 4/

1

0

14

Rohit Singh

@rohitsingh8080

5 months

Excited to be at the @iclr_conf and participating in the @gembioworkshop . If you're going to be there and want to talk about ML for protein modeling (esp PLMs) or single-cell regulatory inference, I'd love to grab coffee. Also, I'm recruiting postdocs and grad students :)

0

2

14

Rohit Singh

@rohitsingh8080

2 years

Super excited about this work with @samsledzieski , Lenore Cowen, and @lab_berger ! I want to highlight a particular thread that we explore in the paper. DTI performance seems to vary quite widely depending on where it's being reported. Why?

Sam Sledzieski

@samsledzieski

2 years

Can you accurately scan ~10 million drug-target pairs per minute? We think you can! In our new preprint (with @rohitsingh8080 , Lenore Cowen, and @lab_berger ), we develop ConPLex, a high-throughput method for predicting drug-target interaction (DTI). 🧵

1

7

34

1

12

Rohit Singh

@rohitsingh8080

3 months

Now on to the aspects of ESM-3 I'm less enthusiastic about. But first, full disclosure on my biases: 1. I think LLMs are a top-3 advance of the 21st century, but I'm not an AGI believer 2. I first-hand understand VC-funded startup constraints, but I speak as an academic 16/n

1

3

12

Rohit Singh

@rohitsingh8080

1 month

The two mCherry variants above are smaller than 96% of fluorescent proteins in FPbase. AFAICT, any FPbase entry shorter than these seems to require some cofactor like flavin or bilirubin for fluorescence. 5/

1

0

12

Rohit Singh

@rohitsingh8080

3 years

@samsledzieski Out in @CellSystemsCP now: genome-scale interpretable protein-protein interaction prediction. Language models power our mapping from 1D (sequence) to 3D (structure-based interaction). Fantastic collab with @samsledzieski @lab_berger and Lenore Cowen:

0

3

11

Rohit Singh

@rohitsingh8080

2 years

When friends with kids visit Boston, nothing -- not even the duck boats -- seems to light up those kids' eyes quite as much as walking into an empty, cavernous lecture hall and pretending to teach. Let's keep MIT open!

Burhan Azeem, Cambridge City Councillor

@realBurhanAzeem

2 years

It's the reason a lot of people fell in love with MIT.

1

2

53

0

2

10

Rohit Singh

@rohitsingh8080

3 months

The methods section is beautifully written. It's clean and clear, using math to clarify rather than to sound smart. Example: this sub-section defining a "frame" could be directly put in a lecture slide: 2/n

2

1

9

Rohit Singh

@rohitsingh8080

3 months

That is starting to happen with PLMs. Here're the ESM-3 authors on the Best Way to build a transformer for protein sequences. (example: rotary embeddings > regular positional encodings) 5/n

1

0

9

Rohit Singh

@rohitsingh8080

8 months

Free food alert. Wed 1/24 at 10:15am.

Duke CompSci

@dukecompsci

8 months

EVENT: Duke CS Colloquium on Wed., Jan. 24 at 10:30-11:30 AM ET in LSRC D106, snacks @ 10:15 AM. Rohit Singh, Duke Assistant Professor in B&B, Cell Bio, and ECE will present "Machine Learning #ML for Precise Diagnostics and Therapeutics." Join us!

0

1

5

0

1

9

Rohit Singh

@rohitsingh8080

10 months

Looking forward to sharing, today at 11:30ET, some of our work (both at my own lab and previously at @lab_berger ) on the geometry of single-cell biology. For those interested in Sceodesic, our differential geometry-based approach to gene programs, I'll be talking about it too!

Bonnie Berger Lab

@lab_berger

11 months

Next week on Weds 11/15 at 11:30am we'll hear from @rohitsingh8080 (a Berger lab alum!) on "The Geometry of Single-Cell Biology: Geodesics, Metrics, and Parallaxes." Rohit will be joining over Zoom at the link below, but we will livestream in Stata G-575 and have snacks!

0

2

6

0

1

8

Rohit Singh

@rohitsingh8080

3 months

Another sign of the field maturing: ESM-3 and AlphaFold-3 are both keeping it conceptually simple. AF3 ditched a lot of the AF2 complexity around equivariance (screenshot from their Nature advert) Similarly, ESM-3 chose not to use state-space models. 6/n

1

0

8

Rohit Singh

@rohitsingh8080

3 years

@baym Don't @ me, man. All of the previous 45 were crap-- I wouldn't compare to any of them if I could get away with it. The 46th time is the charm!

0

1

7

Rohit Singh

@rohitsingh8080

4 years

@zachweinberg @Post_Market The taxpayer should get the upside, not the current equity-holder. One way this can happen is by deciding where in the capital structure the taxpayer's money is infused. Another is by warrants, like Buffett did with GS/MS in 2008. 2/

2

0

6

Rohit Singh

@rohitsingh8080

2 months

Today (2:20pm ET) at #ISMB2024 , I'll be talking about contrastive learning with PLMs to predict drug target interactions. Joint work w @samsledzieski (and @lab_berger and @TheBrysonLab and Lenore Cowen). Sam's also presenting his new work on fine-tuning PLMs. Both at 3D-SIG.

1

0

8

Rohit Singh

@rohitsingh8080

4 months

Is your paper unlikely to be ready in time for #NeurIPS ? Don't worry. Get some good sleep first. Then cut your draft to 4 pages, telling a shorter but stronger story. Submit to !

AccML.Bio Workshop @ ICML'24

@AccMLBio

4 months

We're still accepting submissions and have extended our deadline to **May 28th, Anywhere on Earth**!

0

1

0

8

Rohit Singh

@rohitsingh8080

1 month

Super fortunate to be part of a fantastic team, with my amazing postdoc @kapil_devkota_ , and a deep, perspective-altering collaboration with @SoderlingLab and their grad students @DaichiShonai and @MaoYiwei2 . 9/

1

0

8

Rohit Singh

@rohitsingh8080

2 years

Tomorrow at #ICLR2022 , let Alex Wu show you how to get much more out of your multimodal single-cell RNA-seq + ATAC-seq dataset. And for the graph neutral network folks: come see a novel extension of Granger causal inference to DAGs.

Rohit Singh

@rohitsingh8080

3 years

Causal understanding of gene regulation is a foundational goal in biology. In an ICLR 2022 paper with Alex Wu and @lab_berger , we show how multimodal single-cell technologies like 10x Multiome, SHARE-Seq or SNARE-Seq can help get us there.

1

10

35

0

1

8

Rohit Singh

@rohitsingh8080

1 month

1

7

Rohit Singh

@rohitsingh8080

9 months

Yupp. From a societal perspective, being able to afford maids and drivers and cooks is a bug, not a feature.

Daniel Feldman

@d_feldman

9 months

an indian friend told me "americans don't hire maids and drivers and cooks because you believe in the dignity of work" no, our inequality is low enough that a normal professional can't afford to hire maids and drivers and cooks 🙃

87

200

6K

0

7

Rohit Singh

@rohitsingh8080

1 month

@samlobe 🤣. No, we are nerds who love sci-fi tropes.

1

0

7

Rohit Singh

@rohitsingh8080

10 months

Bonnie told me to retweet that this is happening now

Rohit Singh

@rohitsingh8080

10 months

Looking forward to sharing, today at 11:30ET, some of our work (both at my own lab and previously at @lab_berger ) on the geometry of single-cell biology. For those interested in Sceodesic, our differential geometry-based approach to gene programs, I'll be talking about it too!

0

1

8

0

1

7

Rohit Singh

@rohitsingh8080

1 year

The code is available on github () and as the Python package "abmap". On github, we already have pre-trained models built on top of ESM-1b (ESM2 coming soon!), ProtBert, and PROSE. Give them a spin!

GitHub - rs239/ablm: Protein language model customized for antibodies

Protein language model customized for antibodies. Contribute to rs239/ablm development by creating an account on GitHub.

github.com

0

3

7

Rohit Singh

@rohitsingh8080

3 years

What do PPI and DTI have in common? The "I"! When adapting protein language models for DTI prediction, first tuning them on PPI prediction improves feature informativeness. Poster at NeurIPS MLSB Workshop on Dec 13 #NeurIPS2021

1

0

6

Rohit Singh

@rohitsingh8080

1 year

Most importantly, I want to thank my lovely and nourishing village: @lab_berger and members of the Bonnie Berger lab, Lenore Cowen, @PerrimonLab , @TheBrysonLab @QuinceyJustman @samsledzieski @alexpywu @BrianHie Ashwin Narayan, @ilauhsoj @tsudhirg @jinboxu_chicago @peng_illinois

1

0

6

Rohit Singh

@rohitsingh8080

3 months

Now onto the hairier issues of IP and sharing. Both EvolutionaryScale and Isomorphic are trying to thread a needle: being seen as open while protecting their commercial interests. Their solution: "ok for academic use, but not for drug discovery". This won't work long-term 20/n

1

0

6

Rohit Singh

@rohitsingh8080

3 months

First, on emphasis: I consider the ESM-3 team among the best in ML+biology, and I admire their design sensibilities. However, their new focus on de novo generation rather than representation learning and property prediction misses where PLMs are most frequently used. 17/n

1

0

6

Rohit Singh

@rohitsingh8080

3 months

Reading the architecture, I was reminded of @MoAlQuraishi 's review of AlphaFold2. To paraphrase him, the suprising thing is that nothing's too surprising. There's no magical layer-- just lots of data, great engineering, and lots of compute. 3/n

1

6

Rohit Singh

@rohitsingh8080

2 years

This is a great example of why data-processing decisions in single-cell genomics are so fraught. Many analyses (I'm guilty here too!) do some form of count normalization (explicitly or implicitly) as initial preprocessing. The figure below emphatically says "rethink that!"

Dharmesh Bhuva

@bhuva_dd

2 years

A simple visual inspection of library sizes (total detections) in spatial molecular data reveals tissue structure regardless of the technology or the tissue being investigated. This suggested that library size contained biological information. (3/7)

1

2

8

1

6

Rohit Singh

@rohitsingh8080

1 year

The amazing Anne Trafton of @MITnews covered ConPLex. She has done a great job of distilling the key ideas for a broad audience!

New model offers a way to speed up drug discovery

A model known as ConPLex can predict whether potential drug molecules will interact with specific protein targets, without having to perform the computationally intensive calculation of the molecul...

news.mit.edu

0

3

6

Rohit Singh

@rohitsingh8080

4 years

@zachweinberg @Post_Market It needn't look like a bankruptcy in every situation. There're two parts here: 1) how much is the equity really worth, and 2) who should get the upside from here on. While the first is a problem that'll need time to sort out, there are ways to address just the second problem. 1/

1

0

5

Rohit Singh

@rohitsingh8080

3 months

As ESM-3 shows, I believe transformers are going to remain the architecture of choice for PLMs. Long contexts (the relative strength of SSD models like Hyena) aren't really needed for proteins, and transfomer's attention mechanism is just unreasonably effective. 7/n

1

0

6

Rohit Singh

@rohitsingh8080

6 months

I don't know when I became the old unix guy. Yeah perl has the best regexes why do you ask?

0

1

Rohit Singh

@rohitsingh8080

4 years

I wonder how many quant models broke due to negative prices. If your code handled this without breaking, you're either a very careful quant or a very careless one. The median case, in my expectation, is that the code crashed.

Bloomberg

@business

4 years

BREAKING: WTI crude oil futures trade at negative price for first time

636

12K

14K

0

5

Rohit Singh

@rohitsingh8080

1 month

@DrAnneCarpenter @CarlosEAlvare17 @doc_jlmeier Now that you have said it, I can't unsee it 😂

0

1

5

Rohit Singh

@rohitsingh8080

3 years

Excited about presenting Schema, our method for integrating multi-modal single-cell data, at #ISMB #MLCSB tomorrow. Find me there or at the poster session later. Joint work with @BrianHie @ashwinn226 and @lab_berger . #ismbeccb2021

0

4

5

Rohit Singh

@rohitsingh8080

3 months

Wearing my finance hat, there's a whole separate conversation to be had if these IP issues can ever be resolved given the business-plan constraints on companies like EvolutionaryScale, Isomorphic and Profluent. But that's for some other time... 24/n

0

5

Rohit Singh

@rohitsingh8080

3 months

I can't help but think that the brilliant Foldseek idea of "structure, but as sequence" inspired many folks. Speculating wildly, I suspect ESM-3 team started working on it in early-2023. That'd be in line with SaProt from @duguyuan (which came out in Oct 2023). 13/n

3

1

5

Rohit Singh

@rohitsingh8080

2 years

If you're at #NeurIPS2022 go say hi to @samsledzieski at *MLSB*. He'll present how we combine protein language models with contrastive learning to find better drugs for protein targets. That helps us be accurate on unseen proteins/drugs while recognizing decoys as "not drugs"

0

5

Rohit Singh

@rohitsingh8080

1 year

If presented as LM frameworks for scRNA-seq data, it is not clear to me that the kind of distributional semantics that underlie a language model have been demonstrated to hold as strongly for scRNA-seq data. Take protein language models (PLMs) for instance... 3/

1

0

5

Rohit Singh

@rohitsingh8080

1 year

I'm on 🧵as rohitsingh_8080 . Haven't yet actively explored there, though.

0

5

Rohit Singh

@rohitsingh8080

3 months

Great backstory by @duguyuan on SaProt development here ⬇️. Using FoldSeek tokens for AFDB can be viewed as knowledge distillation from AF2. You'd think that this lazy/simple distillation could be improved by using fancy GNNs on actual AF2-predicted structures. You'd be wrong.

fajie yuan

@duguyuan

3 months

@rohitsingh8080 Before choosing foldseek tokens, one key question is why we cannot directly use explicit structures with MLM, we spent over one year struggling here, and find it does not work with predicted structures if u have to use MLM loss- a key highlight in Saprot.see fig 2.

0

2

0

5

Rohit Singh

@rohitsingh8080

1 year

With scRNA-seq gene expression, a similar distributional semantics *probably* holds but AFAIK hasn't been conclusively demonstrated yet. Doing so requires a massive diversity of cell types, cell states and perturbational states and needs to clearly rule out technical factors. 5/

2

0

5

Rohit Singh

@rohitsingh8080

1 year

I think these papers are super-cool-- they are introducing very innovative methods! My point is that the LM framework is not needed to motivate their contributions-- the DL techniques they leverage (e.g., attention and transformers) are used well beyond language models. 2/

1

0

5

Rohit Singh

@rohitsingh8080

2 years

My New Englander daughter to this Tropic of Cancer-raised dad: "finally, it feels like Nature is committing to this winter."

0

5

Rohit Singh

@rohitsingh8080

3 months

So what did I not like? Well, my chai ran out a while ago so maybe next time. 15/n

1

0

5

Rohit Singh

@rohitsingh8080

3 months

You can learn a better *sequence* representation if you have structure information during training. Somehow, the learned sequence attentions are better. This is a direction for someone to figure out why! 12/n

2

0

5

Rohit Singh

@rohitsingh8080

1 month

@srivsesh @sokrypton This is a perceptive observation. I'll do a longer thread later, but in short, I think part of the issue are structure prediction tools and metrics like pLDDT etc.

0

5

Rohit Singh

@rohitsingh8080

3 months

Science is serendipitous. Many drugs are found by following a tangent off a basic-science discovery process. A license that precludes the chance for such serendipity means I'd need to judge each project at the start for its therapeutic potential. How do I even do that? 23/n

1

0

5

Rohit Singh

@rohitsingh8080

4 years

@tangming2005 In support of what you said about sparsity of multi-modal RNA+ATAC-seq data :), there's a good discussion of that in a recent paper by Gonzalez-Blas et al. from the Aerts lab:

1

2

5

Rohit Singh

@rohitsingh8080

1 year

... The power of PLMs comes from the distributional semantics of protein sequences, arising from evolutionary constraints. These distributional semantics have been the bedrock of bioinformatics sequence analysis (e.g., MSAs) for decades, long before transformers were invented. 4/

1

0

4

Rohit Singh

@rohitsingh8080

3 months

Almost immediately, we updated our method to simply slot in Foldseek embeddings along with PLM embeddings. Just that got us a big boost! TT3D ("Topsy-Turvy 3D") came out in 2023: 10/n

TT3D: Leveraging precomputed protein 3D sequence models to predict protein–protein interactions

AbstractMotivation. High-quality computational structural models are now precomputed and available for nearly every protein in UniProt. However, the best w

academic.oup.com

1

4

Rohit Singh

@rohitsingh8080

3 months

All hail the @thesteinegger ! At ISMB 2022, Lenore Cowen, @samsledzieski , @kapil_devkota_ and I were at the 3D-SIG session to present our work on Topsy-Turvy for PLM-based PPI prediction. Foldseek was presented there too, and we were floored! What a life that idea has taken! 9/n

1

0

4

Rohit Singh

@rohitsingh8080

4 months

The paper enhances bioRxiv preprint (and RECOMB 2023 paper) with additional details: we show how you could incorporate data across time-points, and dig deeper into the lag between fast and slow transcription factors.

Causal gene regulatory analysis with RNA velocity reveals an interplay between slow and fast...

The dynamics of cellular development can be mapped by single-cell techniques like RNA velocity. Our method, Velorama, unravels the underlying regulatory drivers of these dynamics by leveraging a...

www.cell.com

1

0

4

Rohit Singh

@rohitsingh8080

1 month

@michaellazear 🤣. Much as I'd love to say yes, the name comes from the team's love for sci-fi and period cartoons. One of the other names in the running was "Zap"

0

4

Rohit Singh

@rohitsingh8080

9 months

Foundation models in scRNA-seq and genomic LLMs are a fantastic idea and most of the problems I have mentioned above *are* solvable. They really require only two guiding principles from NIH etc: a) avoid a tragedy-of-the-commons situation, and b) reduce frictions in data sharing.

0

2

Rohit Singh

@rohitsingh8080

1 month

@amyxlu ISTG it was conceived before. We actually panicked for a hot second, debating if we should change the name given the... dance. I loved CHEAP btw, great work!

0

4

Rohit Singh

@rohitsingh8080

1 year

Intriguingly, the high-density region also seems to be where therapeutic antibodies map heavily to. We hope this direction of research will lead to an analog to the Lipinski's Rule of 5 for antibodies. 🔎💊

1

0

3

Rohit Singh

@rohitsingh8080

3 years

Once again, my procrastination has turned out to be wise, in hindsight. Patting myself on the back for not having planted saplings over the weekend.

0

4

Rohit Singh

@rohitsingh8080

1 year

A long-standing goal in biology has been understanding how a change in one cell has tissue-level effects. With rich single-cell multimodal datasets and the astonishing progress in ML (e.g., protein language models and diffusion), we may finally have the data and tools needed.

2

0

4

Rohit Singh

@rohitsingh8080

4 months

Velorama is predicated on the idea of doing Granger casual inference on a directed acyclic graph. I think that core approach is relevant in a variety of causal inference contexts where a strict ordering of observations is an unrealistic assumption.

1

0

2

Rohit Singh

@rohitsingh8080

9 months

Shout-outs also to @uthsavc , @cong992 and others from the Raphael lab for the great work!

1

0

4

Rohit Singh

@rohitsingh8080

1 month

@fakefriedberg 🤣 no, inspired by the classific sci-fi trope of a laser gun that shrinks or expands items.

1

0

4

Rohit Singh

@rohitsingh8080

1 year

If you're a grad student, postdoc, or SW engineer interested in this, pls reach out! We'll use language models and diffusion to identify and design protein-X interactions. We'll also decipher regulatory and cell-cell interactions using multimodal and spatial single-cell data.

1

2

4