Brian L Trippe
@brianltrippe
Followers
1,684
Following
452
Media
8
Statuses
92
Bayesian statistics at @Columbia , machine learning for protein design @UWproteindesign
New York City, NY
Joined December 2016
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
LOVTITUDE FANMEETING
• 354362 Tweets
FB MADE YOU BLUSH
• 345666 Tweets
Xavi
• 199570 Tweets
#それスノ
• 148699 Tweets
#原神予告番組
• 134317 Tweets
クロリンデ
• 86991 Tweets
シグウィン
• 80985 Tweets
#한성수_방시혁이랑_손잡고_한강가
• 79951 Tweets
Hansi Flick
• 76872 Tweets
Laporta
• 68257 Tweets
ノーヒットノーラン
• 66623 Tweets
ジェシー
• 60022 Tweets
LEAVE SEVENTEEN ALONE
• 47879 Tweets
アルハイゼン
• 37605 Tweets
戸郷投手
• 31553 Tweets
ショートアニメ
• 25945 Tweets
تشافي
• 22762 Tweets
沢村栄治
• 20594 Tweets
魔神任務
• 10117 Tweets
I am pleased to share that I have accepted a position
@Stanford
to start as an assistant professor in the department of statistics, with an affiliation in
@StanfordData
, this fall!
85
29
937
Can ML build diverse protein scaffolds around functional motifs? We find equivariant diffusion models and Sequential Monte Carlo may help! Joint work with Jason Yim, and coauthors Doug Tischer,
@ta_broderick
, David Baker,
@BarzilayRegina
and Tommi Jaakkola
8
91
312
How can we collect good enough data for machine learning driven protein design? We show that random numbers are part of the picture. Work with the David Baker lab (including
@erika_alden_d
) and MSRNE (with
@KevinKaichuang
and
@lorin_crawford
). (1/4)
2
30
139
Working on this project has been a blast and we're thrilled to finally share it. An exciting future for protein design with RoseTTAFold Diffusion lies ahead!
DALL-E’s amazing images are popping up all over the web. That software uses something called a diffusion model, which is trained to remove noise from static until a clear picture is formed.
Turns out diffusion models can design proteins too!
34
530
2K
1
7
44
I'm excited to share a new paper with
@jhhhuggins
, Raj Agrawal and
@ta_broderick
coming out
@icmlconf
! We speed up Bayesian inference in high dimensional generalized linear models using low rank approximations of data, with a method we call "LR-GLM".
1/5
1
9
29
Come check out our poster today at 4pm! We find that with a sequential Monte Carlo technique known as twisting we can use conditional sampling heuristics, like reconstruction guidance, to get exact inferences with additional compute.
Twisted Diffusion Sampling - Practical and Asymptotically Exact Conditional Sampling for Diffusion Models
Theoretically grounded conditional sampling for diffusion models with applications to inpainting, Bayesian inverse problems, protein design, ...
2
8
45
0
3
28
We (
@skdeshpande91
,
@ta_broderick
and myself) have a new pre-print out now! It’s called “Confidently comparing estimators with the c-value”.
1
2
19
The motif-scaffolding problem is central in protein design: given a target motif (a structural fragment conferring function) construct scaffolds that support it. We propose a generative modeling approach to this problem with two steps.
1
3
17
The first step is modeling a distribution over protein structures. We develop ProtDiff, which tailors equivariant graph neural networks for diffusion probabilistic models (DPMs) on protein backbones. ProtDiff samples diverse topologies, including structures not in the PDB.
1
1
13
The second step, SMCDiff, samples scaffolds from the conditionals of ProtDiff, given a motif. Naive inpainting fails to generate long scaffolds. So SMCDiff directly targets conditionals of DPMs. It’s the first method to guarantee DPM conditional samples in a large-compute limit.
1
1
12
We believe generative modeling is a promising direction in the motif-scaffolding problem. We plan to release code once we have made more progress. Check out our preprint to learn more!
2
1
12
We model protein backbones only and leverage recent work on fixed-backbone sequence design. We use ProteinMPNN (link) to get sequences and validate them against AlphaFold predictions (without MSA) by TMScore. We call this metric self-consistency TM-Score (scTM).
1
1
10
Our samples often agree with AlphaFold predictions (scTM > 0.5) and those that agree exhibit a diverse array of topologies across different lengths.
1
1
9
Pseudorandom numbers are crucial in computer science (e.g. cryptography) and statistics (e.g. randomization in trials), but rarely feature in biological assays. So it’s neat to find that they’re useful here too!
(4/4)
1
1
6
But we can get rid of this bias if we use random numbers to flip a (biased) coin to choose how to bin each cell. We show how this works with a mix of statistical theory, simulations, and wet-lab experiments.
(3/4)
1
1
5
By adding randomness to sort-seq assays (fluorescence activated cell sorting + high-throughput sequencing) we can get precise multiplexed measurements. This solves an open problem in sort-seq assays: deterministic binning introduces systematic bias that limits precision. (2/4)
1
1
5
Also check out our other ICML paper (led by Raj Agrawal) on "The Kernel Interaction Trick", which allows us to use Bayesian inference to efficiently identify pairwise interactions between covariates in regression models!
5/5
0
0
4
You might wonder: wait, what about RISK, i.e. the loss averaged over all possible realizable datasets? Well it turns out we can construct T1 and T2 where (A) T2 has smaller risk than T1 but (B) T2 incurs larger loss than T1 on *the majority of datasets*
1
0
4
0
0
1
Basically, loss averaged over all possible but unrealized datasets may not be close to loss incurred on the single observed dataset. c-values try to answer a different question: which estimator works better on *my observed dataset*.
1
0
2
0
0
2
See code for all applications at
1
0
2
@_aliaabbas
@ta_broderick
@BarzilayRegina
Thanks for your interest! You can find a version of the code in the supplement of our ICLR paper on OpenReview
1
0
2
Enter the c-value! It quantifies how confident we are that T2(Y) achieves smaller loss than T1(Y). Informally, a high c-value reassures us that using the comparatively more complicated T2(Y) results in smaller loss than using T1(Y). (Thm 2.2).
1
0
1
We show c-values are useful for evaluating Bayesian estimates in a range of applications including hierarchical models of educational testing data at different schools, shrinkage estimates that utilize auxiliary datasets, and selecting between different GP kernels.
1
0
1
Say you want to estimate an unknown parameter T* based on data Y. You have two potential estimates T1(Y) and T2(Y). T2 might be complicated (e.g. from a hierarchical model) and T1 is a more common baseline (e.g an MLE). When is it safe to abandon T1(Y) in favor of T2(Y)?
1
0
1
Ever used a hierarchical model to estimate a parameter and wondered if you’re actually doing better than a simpler baseline like maximum likelihood?
We present a method addressing this question!
1
1
1
Unlike variational Bayes, this approximation is conservative; LR-GLM doesn't underestimate uncertainty. We also show how increasing computational budget increases the information extracted from data about the model. 3/5
1
0
1
@ml4proteins
@erika_alden_d
@EliWeinstein6
Looking forward to seeing you there!
Zoom link:
password: ml4prot
0
1
1
We can even use the c-value to choose between T1 and T2: if the c-value exceeds a confidence level alpha, use T2. We show that we rarely incur higher loss as a result of the data-based selection than if we had just stuck with the default T1 (Thm 2.3).
1
0
1
To compute c-values, we construct lower bounds on the difference in loss L(T*, T1(Y)) - L(T*, T2(Y)) that hold uniformly w/ prob alpha. We demonstrate our construction w/ many examples including empirical Bayes shrinkage, Gaussian processes, & logistic regression.
1
0
1
One idea: Use a loss function L (e.g. squared error) and report the estimate with lower loss! I.e. use the more complicated T2 iff L(T*, T2(Y)) < L(T*, T1(Y)).
Problem: T* is unknown, so we can't operationalize this process.
1
0
1
Additionally, we provide theoretical guarantees on approximation quality, with non-asymptotic bounds on approximation error of posterior means and uncertainties. 4/5
1
0
1
With LR-GLM, we make inference with Laplace approximations and MCMC faster by up to full factor of the dimensionality. The rank of the approximation defines a trade-off between the computational demands and accuracy of the approximation. 2/5
1
0
1
Finally, constructing the uniform bounds on L(T*, T1(Y)) - L(T*, T2(Y)) is challenging & sometimes we need approximations. Fortunately,
@jhhhuggins
helped guide us towards some non-asymptotic bounds for Gaussian models (Thm 4.1), and an extension to logistic regression (Sec 5.2)!
1
0
1