Umut Eser
@Umut1Eser
Followers
1,220
Following
1,479
Media
135
Statuses
1,807
co-founder & AI @weave_bio join @weave .bio
Joined March 2013
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Kamala
• 2859574 Tweets
Liam Payne
• 1888364 Tweets
hobi
• 661186 Tweets
Bret
• 617219 Tweets
Argentina
• 597896 Tweets
hoseok
• 560779 Tweets
WELCOME BACK J-HOPE
• 455250 Tweets
jhope
• 449477 Tweets
Fox News
• 326243 Tweets
García Luna
• 293528 Tweets
ブロック
• 259643 Tweets
DE LA MANO DEL SEÑOR
• 104922 Tweets
LINGORM PANTENE VOGUE TRIP
• 99477 Tweets
Sabrina
• 98850 Tweets
#AgathaAllAlong
• 77965 Tweets
Bluesky
• 50578 Tweets
Wanda
• 44156 Tweets
ミュート
• 42940 Tweets
Alvarez
• 42457 Tweets
スーパームーン
• 38419 Tweets
#AEWDynamite
• 38024 Tweets
Shohei
• 30778 Tweets
Ohtani
• 30648 Tweets
Lonzo
• 29690 Tweets
イーロン
• 27064 Tweets
Fortaleza
• 23483 Tweets
Lynx
• 20025 Tweets
AI学習
• 14953 Tweets
ランナー
• 14255 Tweets
#BakeOffFamosos
• 14182 Tweets
महर्षि वाल्मीकि
• 13520 Tweets
ジンくん
• 11542 Tweets
#WNBAFinals
• 10316 Tweets
#ProvaDoFazendeiro
• 10267 Tweets
Pinned Tweet
The new platform we have been working on,
@weave_bio
, is partially out of stealth. Stay tuned for exciting updates:
1
6
17
I hung this inspirational cartoon created by
@arjunrajlab
in my office to remind myself the importance of the proper representation of complex systems.
5
28
112
It's mid April and snowing.
Yep, you guessed it right, it's Boston.
3
4
82
2005-2020:
J1>F1>OPT>STEM extension>H1>O1>O1>EB1>Green Card
Finally...
4
0
61
We just closed $123M in series B!
Go
@cellaritybio
!
#drugdiscovery
#singlecell
#machinelearning
#ai
6
11
60
One of the biggest damage of the Russian invasion is that now every country will start multiplying their defense budgets.
And every penny that is spent on war is not spent on healthcare, education, or infrastructure.
4
4
54
Today is the day we announce our company, what we do, how we do, why we do.
#cellbiology
#cellbehavior
Today
@FlagshipPioneer
unveiled our company! The first to develop medicines targeting cell behavior by uniting network biology, AI and technology. Learn more at
0
9
23
2
7
55
I'm back in Boston. Starting in a very exciting biotech startup at MIT. I noticed how I missed the Kendal Square spirit.
5
1
49
GPT4 autoregressive nature makes it unstable and sensitive to prompting.
Here are some interesting examples:
When you say "do not explain", it just fires up a token (incorrect): (1/3)
3
6
45
Cellarity is covered in Nature Biotech. Proud to be a part of the team that builds an amazing platform for therapeutics that deeply roots in tech development, systems biology, machine learning, and generative chemistry.
1
23
40
Pfizer/BioNTech. I hope my toll-like receptors dont recognize the pseudo-uridinated mRNAs as foreign and my ribosomes translate them as if they're coming from my DNA. Then, I hope my immune system thinks that the translated capsules are foreign and dangerous.
3
1
40
I am thrilled to announce that I am taking on a new challenge in biomedical intelligence, bridging patient stories with molecular evidence, working on creating highly integrated intelligent biomedical engines.
1/2
1
0
36
My daughter saw a hard drive on my desk and asked what that is. Told her that it stores pictures, videos etc.
Then she said:
"oh, it's like the cloud"
Feeling old.
3
0
31
I moved to RTP area in North Carolina. So far so good.
Growing location. A ton of new constructions and a high influx of tech people.
5
0
31
I will be talking about
@cellaritybio
's approach to machine learning applications to drug discovery in the context of Cell Behavior Therapy next week at Drug Discovery Innovation Programme (DDIP) organized by
@WorldBI_Uk
0
4
30
GPT3.5 and GPT4 both fail at Kahneman's Linda problem (Conjunction fallacy). Of course, I morphed the example to avoid memorization.
5
7
30
T cells killing the cancer cell is amazing. But, what's more interesting is that these guys are trained so that they don't kill the normal cells.
Incredible video of a T killer cell DESTROYING a huge ovarian cancer cell using pore-forming toxins
#MedTwitter
34
760
4K
0
9
28
Today is my last day at
@cellaritybio
. It's been a very productive, amazing adventure for the last 27 months, started as VL49 as a
@FlagshipPioneer
venture with a few people, unconstrained ambitious dreams, moving aspirations & exponential progress on solid scientific grounds 1/
1
0
28
After failing 3 times in a row for a website's "are you a robot?" captcha test, I started to suspect myself.
6
3
27
Reinventing the wheel can still be useful as long as the methodology is novel.
#selfconsolation
1
19
25
FIDDLE (Flexible Integration of Data with Deep LEarning) is out.
#ASAPbio
Based on
@tensorflow
and
@TorchML
2
11
25
A neat trick to flow the gradients where the forward-pass contains discrete decisions over a differentiable function:
in Tensorflow:
y = tf.stop_gradient(x_disc- x_cont) + x_cont
or in PyTorch:
y = (x_disc- x_cont).detach() + x_cont
1
9
24
@arjunrajlab
In principle, it maybe the case. But in reality, scATACseq is noisier.
Sampling from only 2 copy of a stretch of DNA vs 100s of copies of mRNA.
1
0
22
My 5 yo concept drawing.
I wasn't expecting. He showed this and said that the time passes faster than expected.
He says it's an hourglass but the sand is leaking sideways too.
0
0
21
When I was building FIDDLE, I was dreaming something towards this end but in genomics:
One Model to Learn Them All. Single model classifies ImageNet, NMT, Image Captioning, Speech Recognition, Parsing.
5
55
158
3
6
19
Gated RNNs can learn complex and long-range patterns in full-length human transcripts.
1
11
20
Proudly carrying Scanpy decal on my mac, an incredibly useful python package developed by my colleague
@falexwolf
. I wish I had such a tool 4 years ago when I was building data integration solutions with Deep Learning.
2
4
20
@ItakGol
The answer is actually surprisingly smart.
It caught the typo and mocked you with a funny response :)
3
0
17
Langsmith (from
@LangChainAI
) is really useful for translating LLMs into robust&reliable products.
Months ago we were planning to implement in house at
@weave_bio
.
But I knew
@hwchase17
was cooking something cool when we chat.
I waited for it and turns out it was worth.
4
5
18
Me: what is the biggest number?
My 4 yo: 1 billion
Me: how about 5 billion?
4yo: everything is same after 1 billion
Well, internally consistent...
2
0
17
What happened in Israel is absolutely barbaric.
Everyone, especially Muslims, should condemn the attacks.
Terrorists have to be singled out.
0
1
16
We can describe, measure, and predict the collective behavior of 10^23 gas particles in a box with just 2 parameters: pressure and temperature.
Yet, we can't describe the collective behavior of 10^7 proteins in a cell even with 100 parameters effectively.
Why?
Let's go deeper🧵
1
4
18
R2D2 said that they have no intention to take over the world. It is all fake news.
0
1
16
I purchased a phone case from
@amazon
and the algo suggests me to buy a phone.
Everyday, I appreciate the importance of causal modeling bettter.
0
2
16
Conclusions:
What
@ylecun
says about AR models exponentially diverging is both right and wrong.
Right as when the model forms an incorrect opinion early on, the rest of the arguments are divergent.
Wrong as when the model generates reasoning first, it de-noises and converges.
2
2
16
@cellaritybio
means a lot to me as it sums up what I have worked on academically so far. Systems biology, transcription regulation, deep learning, complexity science, causal inference etc.
1
0
14
@quaidmorris
@karpathy
Invert the acronym and make it a SPIN off conference. Also, it becomes safe searchable.
1
1
15
Starting as a Chief Innovation Officer at
@DoctorEvidence
. Very excited to be a part of the team building a massively valuable digital platform.
2/2
0
0
15
ChatGPT passing bunch of serious exams, like medical licensing, MBA, bar etc., tells a lot more about how incompetent our qualification criteria are than how smart the AI is.
1
4
15
Predicting Hi-C contacts from DNaseI and sequence using deep learning:
0
11
13
Homebrewed biology lab for testing antibacterial properties of different surface materials :)
1
0
13
We would be so lucky if our communication language turns out to be sufficient to explain biological complexity. I've been thinking about how would the most efficient conceptualization look like, absolute vs relavite, reducible vs emergent. Maybe should write a blog.
2
1
13
Now I'm building a specific neural network architecture using quaternions. The idea is to exploit the rotational symmetries of the data.
Better prior means faster convergence, easier optimization, better compression, and hence better generalization.
Architecture is the prior.
0
3
13
But when you say "first explain, then answer", it starts reasoning, then finds the correct answer:
3
2
12
@anshul
This is esp. a serious case where the postdoc is the only bread earner in the family with kids in an expensive city like Boston.
2
0
13
8 yo: Why do you keep saying smart phone? It's just a phone!
Me: Well, ... I guess you're right nowadays.
Feeling old.
1
0
13
Meeting from home, with the background of my office. I love Zoom virtual background :)
2
0
13
Efficiency tip:
Organize your calendar to have a minimum number of context switching.
4
0
12
My daughter's and my favorite game is to create a new game. This one is really fun. Jenga and playing cards are combined for a real estate market simulation :)
1
0
12
Rule
#1
of blog writers:
Think deep on the topic you'll write. But not too deep so that you don't feel the topic is obvious and unnecessary.
0
1
12
Autoencoder on metabolomics:
The biological interpretation of the first hidden layer reveals 8 commonly enriched significant metabolomics pathways that cannot be discovered by other machine learning methods:
1
1
11
I love it when people announce their publications/works on Twitter and briefly explain the key findings in a thread.
0
1
12
The most valuable innovation ideas are simple and elegant such that they seem obvious only after the fact.
0
3
12
If you want to encode DNA/RNA for neural networks or kernel machines, try 2-bit encoding instead of one-hot.
e.g.00:A 01:C 11:T 10:G
3
1
12
At the end of a day, you want to find something in a book and it takes on the order of seconds to realize that there is no ctrl+F. It is a sign that you overworked the day.
0
0
12
My point of view on how to accelerate the
#innovation
in
#drugdiscovery
, also hints at the new company we will be launching soon:
1
5
11
Flagship upstart Cellarity gathers $123M to finance its exploration of cell behavior, blazing a new path to building a broad pipeline -
1
3
11
Unintended consequences of fast pacing technologies :)
"I'm here live. I'm not a cat."
A lawyer logged into the 394th Judicial District with a cat filter on and couldn't turn it off.
This may be the funniest video I've seen this year.
4K
56K
213K
1
0
11
People were visiting our website and asking "What is
@weave_bio
actually building?"
Well, as of today, our new website is live:
0
2
9
Exciting to see
@kernalbio
mRNAs to be sent to International Space Station for zero-gravity oncology tests next week.
Next week's scheduled launch of
@northropgrumman
#CRS
-14 will carry more than 20 National Lab-sponsored investigations to the
@Space_Station
. You can learn more about this exciting
#science
and how it benefits us right here on Earth in this video!
1
32
156
1
1
10
@van__Oijen
@Wikipedia
Wikipedia is one of the greatest (both in size and value) projects in the history. When I was a grad student, the assignments of
@JanSkotheim
s class was to create a Wikipedia article on a subtopic of the cell cycle. It was really fun and educational.
0
1
10
@Caroline_Bartma
@Graham_Coop
No, they wouldn't let millions of their own to be sacrificed. I suspect cats. They PR on social media to distract.
1
0
10
Some people should apply this to protein folding problem. A hybrid of first-principle based ground up simulation and statistical coarse-graining based gradient descent learning would overcome data limitations.
0
1
10
My daughter saw the plastic bags next to the highway and told me:
d: I want to be a trash collector
me: interesting idea. How will you collect all of them?
d: I will hire people.
me: ok. how will you pay them?
d: ?
me: what if we make plastic bags precious materials?
1
0
10
I'm involved in a
@FlagshipPioneer
exploration project to drive biological innovation using
#MachineLearning
on
#KnowledgeGraph
s.
We are looking for talented consultants to lift the project to the next phase. The first position is Graph DB (Neo4j):
2
4
9
A short primer (thank me later):
Invariance:
f(x)=f(Transformation(x))
Equivariance:
Transformation(f(x)) = f(Transformation(x))
1
3
9
#ECCB18
is on. I was supposed to give 1 of the 3 accepted talks from our team at
@4Catalyzer
. But I couldn't make it bc of ongoing green card application. I will briefly highlight our 3 works here. (thread) >>>
2
1
10
Fascinating
Cells are particularly good at solving mazes, according to a Science study from earlier this year, which demonstrates how they are able to navigate long and complicated routes through the body using self-generated chemoattractant gradients: ($)
16
277
1K
0
0
10
Altaic languages like Turkish, Korean, Japanese have an inverted way of thinking with respect to indo-european languages. You have to wait for the end to get to the action.
I'm curious about the implications for bilinguals.
The next one is another favorite of mine as an interpreter w/ Turkish in their language combination: simultaneous between languages w/ inverse structure. I'd like to share w/ you 3 graphs that show a sentence translated from French, German & Turkish into English:
141
4K
17K
2
1
9
Deep learning of the regulatory grammar of the yeast 5' UTR sequences
0
2
8
Every time I use Ray (), I'm so satisfied. It works like a charm.
1
0
9
Somehow 1% mortality rate seems insignificant to people who dont isolate themselves.
But think about this, everyone will personally know someone that will die due to the Covid-19.
1
1
9
Remarkable!
AlphaGo Zero: Learning from scratch. Self-play strikes again! This is pretty cool.
7
311
592
0
4
9
PLoS Medicine has published a special issue on Machine Learning for Health:
0
1
9
We are looking for Deep Learning Scientists, Bioinformatics Data Scientists, Data Science Interns. Check out the open positions:
0
5
9
Our work -from my PhD- is out:
"Form and function of topologically associating genomic domains in budding yeast"
0
3
9
Our review article just came out:
Cellular identity at the single-cell level
0
2
7
The team has now grown to 7-8x of size. I am confident that Cellarity will be very successful and I will always be rooting for it.
2/2
1
0
9
The size of living beings spans 7 orders of magnitude. An incredibly large dynamic range for a stochastic process.
1
0
8
Looks like me watching my deep learning models' loss curves :)
1
0
8
@kaia_mattioli
Yep
The city where the lab equipment ad goes into the public transportation is called Boston.
http://t.co/uuDhb1Ug09
0
1
3
0
0
8
Wow, a centimeter-long bacterium, with a cool name "Thiomargarita magnifica".
@JanSkotheim
is expected to explain its humongous size that is orders of magnitude larger than your average bacteria :)
1
2
8
Reading about technical analysis in finance and stock markets. Trends, intersections, moving averages, etc.
Looking from the discipline of machine learning, they all sound like bullshit.
3
0
8
d: how?
me: the reason why plastic remains in nature for so long is that there is no natural process that decomposes them. But if you can engineer an enzyme that breakdown polymers, and converts them into ATP, then you can make them precious.
d: cool. I will be a trash collector.
1
0
8
The paper proposes a new algorithm for jointly modeling labels and worker quality from noisy crowd-sourced data, which can also be very useful for biomedical datasets:
1
0
8