Vikram Sundar

Researcher in ML and Structural Biology | ML Scientist at @generate_biomed and @HertzFoundation Fellow | @vikramsundar.bsky.social

[Vikram Sundar @vikramsundar]

The FLIGHTED preprint is out, with some exciting updates! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt. https://t.co/hsTfpd0E7e 1/

[Vikram Sundar @vikramsundar]

Excited to present the first major project of my PhD: FLIGHTED, a method of generating protein fitness landscapes from noisy high-throughput experimental data! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt.

I'll be presenting a poster at @workshopmlsb at NeurIPS. 1/

[Vikram Sundar @vikramsundar]

@boqiangtu and I will be presenting about FLIGHTED and DHARMA at the @gembioworkshop at #ICLR2024. We're excited to talk about our new ability to generate massive protein fitness datasets. See you there!

[Vikram Sundar @vikramsundar]

The FLIGHTED preprint is out, with some exciting updates! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt. https://t.co/hsTfpd0E7e 1/

[Vikram Sundar @vikramsundar]

@boqiangtu and I will be presenting about DHARMA and FLIGHTED at @gembioworkshop at ICLR in May. I'm looking forward to talking to anyone interested in protein design or fitness modeling there! Finally, thanks to my co-authors for making this work possible. 14/14

[Vikram Sundar @vikramsundar]

Our results show that FLIGHTED matters: model ranking changes post-FLIGHTED and performance generally improves, showing that accounting for noise makes your conclusions both more accurate and better. 8/

[Vikram Sundar @vikramsundar]

The 160k TEV protease variant dataset is available on Zenodo here (along with all other data relevant to the paper). We hope this is a useful resource for people developing or benchmarking protein fitness models: https://t.co/6oM1BmZPif 13/

[Vikram Sundar @vikramsundar]

New to the preprint, we've added comprehensive benchmarking on two protein fitness datasets of 160k variants: (1) the popular GB1 dataset, corrected with FLIGHTED-Selection, and (2) a dataset we generated on TEV protease using DHARMA. 7/

[Vikram Sundar @vikramsundar]

The FLIGHTED preprint is out, with some exciting updates! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt. https://t.co/hsTfpd0E7e 1/

[Vikram Sundar @vikramsundar]

Excited to present the first major project of my PhD: FLIGHTED, a method of generating protein fitness landscapes from noisy high-throughput experimental data! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt.

I'll be presenting a poster at @workshopmlsb at NeurIPS. 1/

[Vikram Sundar @vikramsundar]

DHARMA's use of base editing to measure fitness means we can generate very large protein fitness datasets through cheap sequencing, but these datasets are extremely noisy. With FLIGHTED, we can account for this noise, improving performance and generating calibrated errors. 6/

[Vikram Sundar @vikramsundar]

We find that single-step selection experiments are actually quite noisy, but FLIGHTED is able to generate calibrated errors robustly across a wide range of experimental parameters. 5/

[Vikram Sundar @vikramsundar]

We applied FLIGHTED to two experimental systems: (1) single-step selection (phage display, FACS, etc.) and (2) DHARMA (a new assay developed by @boqiangtu and @kesvelt that links protein fitness to base editing). 4/

[Vikram Sundar @vikramsundar]

FLIGHTED relies on Bayesian modeling and stochastic variational inference to generate calibrated errors on fitness measurements from experimental data. It does not need to be trained on ground-truth fitnesses; we use biological knowledge of the experimental system instead. 3/

[Vikram Sundar @vikramsundar]

FLIGHTED is a robust and easy-to-use method of generating protein fitness landscapes from noisy high-throughput experimental data. This should make it easier for ML scientists and protein engineers to incorporate underlying experimental noise into their modeling. 2/

[Alexis Ross @alexisjross]

One of my favorite things about grad school has been getting to play chamber music again--Had a blast playing Tchaikovsky Piano Trio with @vikramsundar and @erencshin 😊

[Vikram Sundar @vikramsundar]

Thanks to my co-authors for making this work possible! I'll be at NeurIPS tomorrow through Saturday and am looking forward to meeting anyone interested in protein design or fitness modeling. 10/10

[Vikram Sundar @vikramsundar]

For more details about FLIGHTED, see the paper here: https://t.co/14XAwNb9qr Code is available on Github: https://t.co/P4q7vxqJcE. A preprint with more data and results will be posted soon as well. 9/

[Vikram Sundar @vikramsundar]

FLIGHTED-DHARMA allows us to generate very large protein fitness datasets quite cheaply, which we are currently exploring - stay tuned for more exciting results! 8/

[Vikram Sundar @vikramsundar]

DHARMA measures protein fitness by linking fitness to base editing of a canvas and using sequencing of the canvas, but this process is extremely noisy. With FLIGHTED, we can measure and account for the noise, improving performance and generating calibrated errors. 7/

[Vikram Sundar @vikramsundar]

Excited to present the first major project of my PhD: FLIGHTED, a method of generating protein fitness landscapes from noisy high-throughput experimental data! Co-authored with @boqiangtu, @lindseyguan, and @kesvelt.

I'll be presenting a poster at @workshopmlsb at NeurIPS. 1/

[Vikram Sundar @vikramsundar]

When applying FLIGHTED to the popular GB1 dataset, we find that benchmarking is strongly dependent on whether noise is accounted for. In order to accurately benchmark our models, we should use methods like FLIGHTED. 6/