First, a massive thank you to Myllia Biotechnology and the Virtual Cell Research Community for organizing such an elegant and uniquely challenging competition. The novel evaluation metric and the strict one-submission-per-day limit made this an incredible test of disciplined, hypothesis-driven data science.
\nI’m excited to share my team's 7th Place solution (Public LB: 4.090 / Private LB: 3.3075).
\nOur core philosophy was that data augmentation would outperform complex architecture. We utilized a very simple \"Blind\" LightMLP (~2.5M params) that takes only a gene index as input, and powered it with two major augmentation pipelines:
\nThe grand vision of virtual cell models is to learn deep mechanistic rules to predict specific perturbation outcomes. However, a fascinating takeaway from our approach is how we navigated the competition's extreme low-data regime (80 training perturbations predicting 120 completely disjoint, unseen targets).
\nWe found that in this highly constrained environment, models attempting to learn perturbation-specific rules severely overfit the training data. Ultimately, our winning strategy leaned into a different biological prior: predicting a highly robust, shared transcriptomic response based on the covariance structure of the training data.
\nAside from manually correcting the delta of the knocked-out gene itself, our model predicts the exact same generalized cellular response for every perturbation. It highlights a profound lesson in computational biology: when perturbation data is sparse, a strong, generalized biological prior is currently the safest path to generalization.
\nOur final solution was the result of 24 carefully documented iterations. If you want to see the full evolutionary journey—including our experiments with external data, and a deep dive into an interesting inverse correlation we discovered between our CV and LB scores—please visit the repository!
\nFull Write-up & Source Code:https://github.com/DrGM87/Myllia-Challenge
\nThank you again to the organizers for providing such a phenomenal dataset and pushing the boundaries of what virtual cell models can achieve!
", "messages": [ { "id": 3426134, "postDate": "2026-03-21T23:24:40.033Z", "content": "First, a massive thank you to Myllia Biotechnology and the Virtual Cell Research Community for organizing such an elegant and uniquely challenging competition. The novel evaluation metric and the strict one-submission-per-day limit made this an incredible test of disciplined, hypothesis-driven data science.
\nI’m excited to share my team's 7th Place solution (Public LB: 4.090 / Private LB: 3.3075).
\nOur core philosophy was that data augmentation would outperform complex architecture. We utilized a very simple \"Blind\" LightMLP (~2.5M params) that takes only a gene index as input, and powered it with two major augmentation pipelines:
\nThe grand vision of virtual cell models is to learn deep mechanistic rules to predict specific perturbation outcomes. However, a fascinating takeaway from our approach is how we navigated the competition's extreme low-data regime (80 training perturbations predicting 120 completely disjoint, unseen targets).
\nWe found that in this highly constrained environment, models attempting to learn perturbation-specific rules severely overfit the training data. Ultimately, our winning strategy leaned into a different biological prior: predicting a highly robust, shared transcriptomic response based on the covariance structure of the training data.
\nAside from manually correcting the delta of the knocked-out gene itself, our model predicts the exact same generalized cellular response for every perturbation. It highlights a profound lesson in computational biology: when perturbation data is sparse, a strong, generalized biological prior is currently the safest path to generalization.
\nOur final solution was the result of 24 carefully documented iterations. If you want to see the full evolutionary journey—including our experiments with external data, and a deep dive into an interesting inverse correlation we discovered between our CV and LB scores—please visit the repository!
\nFull Write-up & Source Code:https://github.com/DrGM87/Myllia-Challenge
\nThank you again to the organizers for providing such a phenomenal dataset and pushing the boundaries of what virtual cell models can achieve!
", "rawMarkdown": "First, a massive thank you to **Myllia Biotechnology** and the Virtual Cell Research Community for organizing such an elegant and uniquely challenging competition. The novel evaluation metric and the strict one-submission-per-day limit made this an incredible test of disciplined, hypothesis-driven data science. \n\nI’m excited to share my team's 7th Place solution (Public LB: 4.090 / Private LB: 3.3075). \n\n### 🧬 Extremely Short Summary of Our Approach\nOur core philosophy was that **data augmentation would outperform complex architecture**. We utilized a very simple **\"Blind\" LightMLP** (~2.5M params) that takes only a gene index as input, and powered it with two major augmentation pipelines:\n1. **3-Tier PseudoBulk:** Bootstrapping single-cell data to multiply our 80 ground-truth samples.\n2. **Leave-One-Out (LOO) Synthetic Perturbations (The Breakthrough):** Simulating knockouts genome-wide by finding control cells where specific genes are naturally low-expressed. This generated ~15,000 synthetic perturbation samples, teaching the model to generalize across all 5,127 genes.\n3. **External Data as Regularization:** We used Replogle K562 data *only* to add beneficial noise and prevent overfitting, rather than as input features.\n\n### 🔬 An Interesting Divergence in Modeling Strategy\nThe grand vision of virtual cell models is to learn deep mechanistic rules to predict specific perturbation outcomes. However, a fascinating takeaway from our approach is how we navigated the competition's extreme low-data regime (80 training perturbations predicting 120 completely disjoint, unseen targets). \n\nWe found that in this highly constrained environment, models attempting to learn perturbation-specific rules severely overfit the training data. Ultimately, our winning strategy leaned into a different biological prior: **predicting a highly robust, *shared* transcriptomic response** based on the covariance structure of the training data. \n\nAside from manually correcting the delta of the knocked-out gene itself, our model predicts the exact same generalized cellular response for every perturbation. It highlights a profound lesson in computational biology: when perturbation data is sparse, a strong, generalized biological prior is currently the safest path to generalization. \n\n### 🔗 Deep Dive & Source Code\nOur final solution was the result of 24 carefully documented iterations. If you want to see the full evolutionary journey—including our experiments with external data, and a deep dive into an interesting inverse correlation we discovered between our CV and LB scores—please visit the repository!\n\n**Full Write-up & Source Code:**[https://github.com/DrGM87/Myllia-Challenge](https://github.com/DrGM87/Myllia-Challenge)\n\nThank you again to the organizers for providing such a phenomenal dataset and pushing the boundaries of what virtual cell models can achieve!", "votes": 5 } ], "comments": [], "raw_markdown_by_id": { "3426134": "First, a massive thank you to **Myllia Biotechnology** and the Virtual Cell Research Community for organizing such an elegant and uniquely challenging competition. The novel evaluation metric and the strict one-submission-per-day limit made this an incredible test of disciplined, hypothesis-driven data science. \n\nI’m excited to share my team's 7th Place solution (Public LB: 4.090 / Private LB: 3.3075). \n\n### 🧬 Extremely Short Summary of Our Approach\nOur core philosophy was that **data augmentation would outperform complex architecture**. We utilized a very simple **\"Blind\" LightMLP** (~2.5M params) that takes only a gene index as input, and powered it with two major augmentation pipelines:\n1. **3-Tier PseudoBulk:** Bootstrapping single-cell data to multiply our 80 ground-truth samples.\n2. **Leave-One-Out (LOO) Synthetic Perturbations (The Breakthrough):** Simulating knockouts genome-wide by finding control cells where specific genes are naturally low-expressed. This generated ~15,000 synthetic perturbation samples, teaching the model to generalize across all 5,127 genes.\n3. **External Data as Regularization:** We used Replogle K562 data *only* to add beneficial noise and prevent overfitting, rather than as input features.\n\n### 🔬 An Interesting Divergence in Modeling Strategy\nThe grand vision of virtual cell models is to learn deep mechanistic rules to predict specific perturbation outcomes. However, a fascinating takeaway from our approach is how we navigated the competition's extreme low-data regime (80 training perturbations predicting 120 completely disjoint, unseen targets). \n\nWe found that in this highly constrained environment, models attempting to learn perturbation-specific rules severely overfit the training data. Ultimately, our winning strategy leaned into a different biological prior: **predicting a highly robust, *shared* transcriptomic response** based on the covariance structure of the training data. \n\nAside from manually correcting the delta of the knocked-out gene itself, our model predicts the exact same generalized cellular response for every perturbation. It highlights a profound lesson in computational biology: when perturbation data is sparse, a strong, generalized biological prior is currently the safest path to generalization. \n\n### 🔗 Deep Dive & Source Code\nOur final solution was the result of 24 carefully documented iterations. If you want to see the full evolutionary journey—including our experiments with external data, and a deep dive into an interesting inverse correlation we discovered between our CV and LB scores—please visit the repository!\n\n**Full Write-up & Source Code:**[https://github.com/DrGM87/Myllia-Challenge](https://github.com/DrGM87/Myllia-Challenge)\n\nThank you again to the organizers for providing such a phenomenal dataset and pushing the boundaries of what virtual cell models can achieve!" } }