Bashor Lab in Houston, Texas | Mammalian Synthetic Biology et al.
Houston, TX
- Tweets 236
- Following 814
- Followers 1,040
- Likes 829
One thing that has become clear in developing and using CLASSIC for a variety of applications (stay tuned!), is how using ML/AI fundamentally reshapes synbio DBTL by bringing focus to bear on dataset generation and model training as central experimental goals. (13/16)
1
1
131
What we found were surprising strategies for attaining digital logic, demonstrating that, while our general notion of how to build these circuits was directionally correct, the model was able to precisely identify non-intuitive compositions that actually worked. (12/16)
1
1
142
We then experimentally validated model predictions from across the space, finding that compositions from the most productive AND- and OR-like regions indeed demonstrated digital logic. We also looked into the data to understand how these designs generated their behavior. (11/16)
1
1
148
We then built sub-libraries for the parts that were initially held out to sample the rest of the space. Fine-tuning using this data proved to be an excellent strategy for expanding the model to predict the full design space while retaining predictive power. (10/16)
1
1
163
We approached this problem by sub-sampling the overall design space, holding out parts to constrain the space, collecting data for several hundred thousand circuits and training a “base” model, which we found could accurately predict the constrained space. (9/16)
1
1
181
Including all the possible compositions and locus topologies for the set of parts we were working with, the design space for these circuits exceeded a 3.4 BILLION member theoretical space - one we almost certainly couldn’t explore experimentally! (8/16)
1
1
202
In one demo, we used a combination of (1) selective design space sampling coupled with (2) multi-step model training, to explore whether a dual input, 3-gene circuit could be designed to do one synbio’s favorite things . . . Digital logic! (7/16)
1
1
228
In synbio, DBTL cycles usually involve iterative, “low-n” build & test, often guided by biophysical models. When circuits don’t work as expected due to non-linear interactions between parts and with the host cell, the result can be frustration + long project timelines. (4/16)
1
1
279
In the paper, we describe CLASSIC (combining long and short-read sequencing to investigate genetic complexity), a platform that uses pooled DNA assembly/screening to build & test large libraries (105-106) of circuits with combinatorially diverse genetic part compositions. (3/16)
1
1
375
This work was done by a team of two superlatively talented PhD students here at @Rice, @kshitij_raiguy and @ronanoconnell, who developed the experimental and computation capabilities for this study over several years of working closely together in the lab. (2/16)
1
1
724
We also hope our work will help enable new cell-based therapeutic modalities in which cells programmed with phospho-signaling circuits are able to rapidly sense and respond to changes in disease/physiological state. 10/n
1
2
482
We demonstrated that these capabilities can be used to implement therapeutically relevant sense-and-respond function, e.g., programming cells to sense TNF-α (inflammatory signal) and respond by secreting IL-10 (anti-inflammatory) to modulate T cell activation. 8/n
1
2
346
This allows us to not only construct phosphorylation cycles in human cells, but to link them together into multi-cycle pathways that activate rapidly and reversibly, or connect receptor inputs to downstream outputs (e.g., transcriptional activation). 7/n
1
2
351
The cycles are constructed from a simple protein domain part toolkit: kinase and phosphatase domains directed to interact with obligate (leucine zippers) and phosphorylation-dependent (SH2s) interaction domains. 6/n
1
2
383
To briefly recap: we developed our framework by engineering reversible phosphorylation cycles in which kinase and phosphatase activities respectively phosphorylate and dephosphorylate a protein substrate. 5/n
1
3
415
Last year we released the preprint and shared a detailed X-torial (nitter.siddz.com/BashorLab/status/17014…) for this work, which was led by the legendary @XiaoyuYang6, a student who recently graduated from the Rice Systems and Synthetic Biology (@RiceSSPB) PhD program. 4/n
1
1
827
We also think they will be great for developing cell-based theranostics capable of maintaining closed-loop control over a variety of diseases, including metabolic, autoimmune, and endocrine disorders. 26/n
6
1,250
In summary, we think our system begins to check some of the boxes for a useful design abstraction, but we mostly hope it helps stimulate conversation about frameworks for phosphorylation circuit engineering, what parts to use, and how to construct and model new circuits. 24/n
1
3
699
Finally, we configured this circuit to sense the cytokine TNF-α (causes inflammation) and respond by secreting IL-10 (anti-inflammatory), enabling feedback control over activation and proliferation of stimulated T cells. 23/n
1
5
681
We then combined sensor and amplifier modules into a complete membrane-to-nucleus signaling pathway that senses external ligand and responds by transcriptional activation. 22/n
1
2
657
We thought this was a pretty good performance for a synthetic circuit—its response timescale was actually on par with native signaling pathways such as TGF-β/SMAD and JAK/STAT. 21/n
1
2
671
To assess the kinetics of ligand-induced phosphorylation, we engineered a fast-time scale reporter based on protein co-localization. The data showed that our circuits could turn on pretty fast, with a t1/2 ~ 23 mins for activation and t1/2 ~ 20 mins for deactivation. 20/n
1
4
1,254
We can build, tune, and connect push-pulls to each other. So, what about connecting them to something useful? We engineered a synthetic sensor circuit in which synthetic receptors shift phosphorylation equilibrium in response to extracellular ligand binding. 19/n
1
5
695
This design works! We observed sequential phosphorylation and even engineered a two-step cascade that amplifies signal from the first phosphorylation event to the second. 18/n
1
1
700
We used SH2 domains, which bind to phosphorylated tyrosine motifs, to engineer connections between an upstream push-pull and a downstream one. 17/n
1
1
679
The simplicity and reliability of our part set led us to establish a non-equilibrium thermodynamic model that we could use to predictively tune phosphorylation equilibrium. 16/n
1
2
667
To complete the push-pulls we also designed recruited phosphatases, which could efficiently reverse synthetic substrate phosphorylation. 15/n
1
2
5,175
We assembled the domains into push-pulls by recruiting kinase or phosphatase domains to the substrate using leucine zipper domains and used multi-color flow cytometry to measure component expression levels as well as phosphorylation. 13/n
1
1
728
To construct synthetic push-pulls in human cells, we developed a set of protein domain parts: kinases and phosphatases, phosphorylatable substrate motifs, and domains that mediate obligate (leucine zippers) or phosphorylation-dependent (SH2s) protein-protein interactions. 12/n
1
1
778
Push-pulls are made up of kinase and phosphatase activities that mutually phosphorylate and dephosphorylate a protein substrate. 11/n
1
1
801
Such a design logic should enable circuit connections to be made easily and scalably, and should allow connections to diverse inputs and outputs. 8/n
1
2
909
This includes Pankaj Mehta and Jason Rocks from @BU_Tweets, former lab member @idmjky, our very own @krai_ostat, @walters_aj101, Jing (Vicky) Liu, and Jason Nguyen, as well as @OlsonCell from UTHealth, Jim Collins from MIT and Nichole Daringer from Rowan University. 3/n
1
1
1,657
Congrats to Rice alum Dr. Sebastián Castillo-Hair and the Seelig Lab on an exciting preprint! It’s great to see them demonstrate the power of coupling libraries with machine learning - this time for mRNA therapeutics! @castillohair @seeliglab biorxiv.org/content/10.1101/…
3
22
2,051
Summer is on the horizon! Starting last summer we began hosting monthly SynBio Supergroup talks where trainees volunteer to present their research. We are one of the few universities with a dedicated program for synthetic biology @RiceSSPB with 58 affiliated faculty!
1
7
756
Congratulations to Dr. Piepergerdes on defending his PhD thesis!! The freshly crowned Dr. @PieperTC explored circuit guided MSC differentiation to chondrocytes using scRNA sequencing. We can’t wait to see what he does next! 👀 #PhD #stemcells
@Rice_BIOE
21
1,102
Big congrats to our grad student @mattlee228 for receiving the NSF GRFP Award!! Matt is using high throughput library and #ML approaches to engineer cell-type specific promoters in mammalian cells.
Bet that book he's been reading had some good tips!
1
27
2,222
At this stage, since tools for developing the right models are mostly accessible already, we believe the onus is very much on synthetic biologists to roll up their sleeves and develop new approaches for generating the training data! (Image: Fig 4A from our preprint)
4
472
We are perhaps most excited about the possibility of using HT genetic design to unlock AI-guided, data driven SynBio—something we believe is critical for the future of the field. We hope techniques like CLASSIC will play a role in realizing this vision! Watch this space! 23/23
8
994
To summarize: CLASSIC is a platform that can be used for profiling gene circuit libraries en masse, and the data you get can be used to map genetic design spaces, learn genetic part composition-to-function rules, train ML, and generally make SynBio projects go faster! 21/23
1
3
716
With library measurements in-hand, we analyzed our data to infer circuit design rules. We found there's 2 ways to engineer high fold change: Low affinity ZF + strong AD, or medium affinity ZF with medium ADs. Promoter strength, spacing, and binding sites matter! 20/23
1
2
633
What about the circuits we didn’t measure? And how accurate are the ones we did? We again turned to ML to complete the design space and sharpen our measurements. After all ML is used ubiquitously to enhance data quality…you just gotta be honest about it 😉 @SamsungMobile? 19/23
1
626
CLASSIC was able to measure 121k circuits (~73% of the total design space). We validated the accuracy of our measurements using individually constructed circuits and were excited to see outstanding agreement between the two sets of measurements (yes, it scales!). 18/23
1
1
634
Quick mention here: Large libraries = BIG datasets. Fortunately, our students @krai_ostat @ronan_oconnell were able to figure out a scalable analysis pipeline, and how to sell it, with a @YouTube style clickbait thumbnail. 17/23
1
3
690
So, we used CLASSIC to assemble a library of circuits from a number of different part categories, and profiled a library of 10^5 circuits in ONE experiment. Did we mention, this is all integrated at single copy in human cells? 16/23
1
1
623
To find out, we turned to a simple zinc finger transcription factor circuit that we have worked on for years. This circuit executes a simple function: turn on GFP expression in the presence of a small molecule drug inducer, 4-OHT. 14/23
1
1
607
The smartest and most insightful people reading this thread will have probably already turned to their neighbor and ask the big question: “will it scale?”. We know we did @sarahcpr, thanks for the tips! 13/23
1
5
625
Nah, just kidding, it’s 2023 ofc we used #ML @SiliconHBO. We created a random forest model that could identify part trends and accurately predict expression based on part composition alone. Our model could even identify unexpected part synergies! 12/23
1
5
592
We validated the CLASSIC measurement against individually constructed circuits and saw remarkable agreement, which was encouraging! Now the only thing left to do was to fit our data to a detailed mechanistic model... 11/23
1
3
596
We constructed a 384-member library from combinations of these parts, integrated it at single copy into a human cell line, sorted based on fluorescence level, sequenced, and inferred expression for each member of the library. 10/23
1
1
1
593
As an initial validation of CLASSIC, we gathered genetic parts commonly used in mammalian SynBio (promoters, Kozaks, and terminators) with the goal of mapping how they contribute to the expression of a fluorescent protein. 9/23
1
2
616
Circuit libraries are introduced into cells, sorted into bins by expression phenotype, and then short-read sequenced @illumina to identify barcodes in each bin. Circuit designs can then be matched to behavior based on their barcode counts! 8/23
1
3
657
CLASSIC uses pooled DNA assembly to create construct libraries containing short barcodes. Long-read @nanopore sequencing is used to associate each part composition in the library with a unique barcode 7/23
1
1
707
Introducing CLASSIC (combining long- and short-range sequencing for investigating genetic complexity), a platform that allows synthetic circuit engineers to build and test >10^5 genetic circuit designs in a single experiment. How does it work? 6/23
1
1
3
744
After working for some time in mammalian cells, where circuit tuning can be particularly challenging (and where DBTL cycles are slowwww), we gave up on trying to figure out how to build exactly the right thing, and decided instead to just build EVERYthing! 5/23
1
2
710
It turns out that designing gene circuits is also really hard. With the potential for context-dependent interactions of genetic parts with each other and with the host cell, finding the right part composition that yields a target behavior can take many design iterations. 4/23
1
1
2
759
First, a useful analogy: There is an entire corner of the internet dedicated to cakes that didn’t quite turn out the way they were intended to. With so many different flavors, flour types, frostings, icings, and oven settings, it’s not unusual to simply miss the mark. 3/23
1
1
4
870
This work was driven by a team of two incredibly talented and dedicated grad students in the Bashor Lab. @ronan_oconnell and @krai_ostat, with contributions from many others in the group, as well as collaborators @traingene (@RiceUniversity) and Pankaj Mehta (@BU_Tweets). 2/23
1
4
1,295