{ "id": 438963, "title": "🔥Papers including Code for this Competition", "url": "/competitions/open-problems-single-cell-perturbations/discussion/438963", "author_name": "", "post_date": "2023-09-13T07:51:59.455871500Z", "votes": 41, "comment_count": 2, "views": 0, "content": "
    \n
  1. Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution
    2. \n
  2. CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data
    3. \n
  3. Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling
    4. \n
  4. Predicting Cellular Responses with Variational Causal Inference and Refined Relational Information
    5. \n
  5. Conditional out-of-sample generation for unpaired data using trVAE
    6. \n
\n

Check out other discussions about AI trends, papers etc:
\nRecent AI Releases & Announcements
\nTrending AI Research
\nTop 5 AI Web Tools Of The Week

", "messages": [ { "id": "2435841", "postDate": "09/13/2023 07:51:59", "content": "
    \n
  1. Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution
    2. \n
  2. CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data
    3. \n
  3. Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling
    4. \n
  4. Predicting Cellular Responses with Variational Causal Inference and Refined Relational Information
    5. \n
  5. Conditional out-of-sample generation for unpaired data using trVAE
    6. \n
\n

Check out other discussions about AI trends, papers etc:
\nRecent AI Releases & Announcements
\nTrending AI Research
\nTop 5 AI Web Tools Of The Week

", "rawMarkdown": "1. [Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution](https://arxiv.org/pdf/2204.13545v2.pdf)\n - **Code:** https://github.com/theislab/chemCPA\n - **Summary:** Single-cell transcriptomics allows for the detailed study of individual cell responses to changes, but using this for high-throughput drug screens is technically challenging and expensive. The new method, chemCPA, uses an encoder-decoder architecture to analyze the effects of drugs that haven't been studied yet. By training on existing bulk RNA datasets, it improves its accuracy without needing extensive single-cell testing. This can lead to cost-saving and faster drug discovery by predicting outcomes in-silico.\n2. [CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data] (https://arxiv.org/pdf/2210.17283v2.pdf)\n - **Code:** https://github.com/causalbench/causalbench\n - **Summary:** Causal inference is crucial in fields like medicine, but testing its methods in real-world settings is tough due to the need for controlled and interventional observations. Evaluations on made-up datasets don't truly represent real-world performance. Thus, we present CausalBench, a tool for testing network inference methods on real single-cell perturbation data. The tool uses new performance metrics and reveals that current methods don't scale well. Surprisingly, methods with interventional data aren't better than observational-only methods, unlike synthetic tests. CausalBench offers a consistent way to assess progress in real-world causal data use.\n3. [Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling](https://arxiv.org/pdf/2211.03553v4.pdf)\n - **Code:** https://github.com/genentech/svae\n - **Summary:** Variational Auto-Encoder (VAE) models are popular for analyzing biological data like single-cell genomics. However, interpreting the hidden variables in these models in biological terms remains a challenge. While there are VAE variants that attempt to disentangle these variables, successful disentanglement using traditional methods has proven hard. Recent approaches suggest using non-stationary data and assuming sparse changes to achieve this. In this work, we use these advances for single-cell genomics with genetic or chemical changes. We treat each change as a random intervention on a small set of latent variables. After testing our methods on simulated data, we applied them to real large-scale gene alteration datasets. Our models, which rely on the sparse change assumption, outperformed other methods in transfer learning tasks.\n4. [Predicting Cellular Responses with Variational Causal Inference and Refined Relational Information](https://arxiv.org/pdf/2210.00116v2.pdf)\n - **Code:** https://github.com/yulun-rayn/graphvci\n - **Summary:** In our study, we introduce a new graph-based method to predict how a cell will react to certain changes, which is crucial for drug development and tailored treatments. We use gene regulatory networks (GRNs) to better predict individual cell responses. To make the GRNs more adaptable, we've enhanced them with an updating technique that offers deeper insights into gene relationships and improves the model. We also introduce an efficient estimator for measuring the impact of perturbations, a step not previously taken. Our tests show our method outperforms leading deep learning models in predicting individual responses.\n5. [Conditional out-of-sample generation for unpaired data using trVAE](https://arxiv.org/pdf/1910.01791v2.pdf)\n - **Code:** https://github.com/theislab/trVAE\n - **Summary:** Generative models excel at creating complex samples based on simple descriptors but struggle when generating out-of-sample. The conditional variational autoencoder (CVAE) doesn't effectively link conditions during training, which leads to issues. We addressed this by using maximum mean discrepancy (MMD) in the post-bottleneck decoder layer. This modification, termed \"transformer VAE\" (trVAE), enhances sample reconstruction and transformations, resulting in better generalization. When tested on complex data, trVAE outperformed existing methods, especially in predicting cell responses to treatments and diseases using single-cell gene data. It also improved accuracy metrics significantly.\n\n**Check out other discussions about AI trends, papers etc:**\n[Recent AI Releases & Announcements](https://www.kaggle.com/discussions/general/437852)\n[Trending AI Research](https://www.kaggle.com/discussions/general/437099)\n[Top 5 AI Web Tools Of The Week](https://www.kaggle.com/discussions/general/437038)", "votes": null }, { "id": "2451688", "postDate": "09/22/2023 17:25:53", "content": "

Thanks a lot for sharing this info. I have been trying to find exactly this

", "rawMarkdown": "Thanks a lot for sharing this info. I have been trying to find exactly this", "votes": null }, { "id": "2487591", "postDate": "10/18/2023 17:01:05", "content": "

it's a nice to know info

", "rawMarkdown": "it's a nice to know info", "votes": null } ], "comments": [ { "id": 2451688, "author_name": "neelshelgaonkar", "author_url": "", "post_date": "09/22/2023 17:25:53", "content": "

Thanks a lot for sharing this info. I have been trying to find exactly this

", "votes": null, "replies": [] }, { "id": 2487591, "author_name": "luckysh", "author_url": "", "post_date": "10/18/2023 17:01:05", "content": "

it's a nice to know info

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2435841": "1. [Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution](https://arxiv.org/pdf/2204.13545v2.pdf)\n - **Code:** https://github.com/theislab/chemCPA\n - **Summary:** Single-cell transcriptomics allows for the detailed study of individual cell responses to changes, but using this for high-throughput drug screens is technically challenging and expensive. The new method, chemCPA, uses an encoder-decoder architecture to analyze the effects of drugs that haven't been studied yet. By training on existing bulk RNA datasets, it improves its accuracy without needing extensive single-cell testing. This can lead to cost-saving and faster drug discovery by predicting outcomes in-silico.\n2. [CausalBench: A Large-scale Benchmark for Network Inference from Single-cell Perturbation Data] (https://arxiv.org/pdf/2210.17283v2.pdf)\n - **Code:** https://github.com/causalbench/causalbench\n - **Summary:** Causal inference is crucial in fields like medicine, but testing its methods in real-world settings is tough due to the need for controlled and interventional observations. Evaluations on made-up datasets don't truly represent real-world performance. Thus, we present CausalBench, a tool for testing network inference methods on real single-cell perturbation data. The tool uses new performance metrics and reveals that current methods don't scale well. Surprisingly, methods with interventional data aren't better than observational-only methods, unlike synthetic tests. CausalBench offers a consistent way to assess progress in real-world causal data use.\n3. [Learning Causal Representations of Single Cells via Sparse Mechanism Shift Modeling](https://arxiv.org/pdf/2211.03553v4.pdf)\n - **Code:** https://github.com/genentech/svae\n - **Summary:** Variational Auto-Encoder (VAE) models are popular for analyzing biological data like single-cell genomics. However, interpreting the hidden variables in these models in biological terms remains a challenge. While there are VAE variants that attempt to disentangle these variables, successful disentanglement using traditional methods has proven hard. Recent approaches suggest using non-stationary data and assuming sparse changes to achieve this. In this work, we use these advances for single-cell genomics with genetic or chemical changes. We treat each change as a random intervention on a small set of latent variables. After testing our methods on simulated data, we applied them to real large-scale gene alteration datasets. Our models, which rely on the sparse change assumption, outperformed other methods in transfer learning tasks.\n4. [Predicting Cellular Responses with Variational Causal Inference and Refined Relational Information](https://arxiv.org/pdf/2210.00116v2.pdf)\n - **Code:** https://github.com/yulun-rayn/graphvci\n - **Summary:** In our study, we introduce a new graph-based method to predict how a cell will react to certain changes, which is crucial for drug development and tailored treatments. We use gene regulatory networks (GRNs) to better predict individual cell responses. To make the GRNs more adaptable, we've enhanced them with an updating technique that offers deeper insights into gene relationships and improves the model. We also introduce an efficient estimator for measuring the impact of perturbations, a step not previously taken. Our tests show our method outperforms leading deep learning models in predicting individual responses.\n5. [Conditional out-of-sample generation for unpaired data using trVAE](https://arxiv.org/pdf/1910.01791v2.pdf)\n - **Code:** https://github.com/theislab/trVAE\n - **Summary:** Generative models excel at creating complex samples based on simple descriptors but struggle when generating out-of-sample. The conditional variational autoencoder (CVAE) doesn't effectively link conditions during training, which leads to issues. We addressed this by using maximum mean discrepancy (MMD) in the post-bottleneck decoder layer. This modification, termed \"transformer VAE\" (trVAE), enhances sample reconstruction and transformations, resulting in better generalization. When tested on complex data, trVAE outperformed existing methods, especially in predicting cell responses to treatments and diseases using single-cell gene data. It also improved accuracy metrics significantly.\n\n**Check out other discussions about AI trends, papers etc:**\n[Recent AI Releases & Announcements](https://www.kaggle.com/discussions/general/437852)\n[Trending AI Research](https://www.kaggle.com/discussions/general/437099)\n[Top 5 AI Web Tools Of The Week](https://www.kaggle.com/discussions/general/437038)", "2451688": "Thanks a lot for sharing this info. I have been trying to find exactly this", "2487591": "it's a nice to know info" }, "source": "meta" }