GenBio was founded with the audacious goal of simulating all of biology. The natural starting point for biological simulation is the cell: the smallest unit of life. Today we're sharing a preview of our first version of that simulator, AIDO Cell.

Context-adaptive inference is doing to statistical modeling what LLMs did to software. What this looks like in the data domain outside of natural/formal languages is an open question, but the evolution of estimation might look similar to @karpathy's evolution of software

In practice, we see estimators getting steadily subsumed by context-adaptive inference. This is reflected in the field's trend toward large-scale pretraining and multi-task post-training, and it doesn't stop with language. e.g. @noahholl's TabPFN arxiv.org/abs/2207.01848

We prove the connection, showing that under squared loss with linear heads, varying-coefficient models and in-context learning reduce to the same kernel ridge estimator. The methods only differ in how much of the estimation pipeline they absorb within the model itself.

Classical estimation is based on task-specific data. Transfer and meta-learning reduce data requirements but still run gradient descent on task-specific samples. In-context learning and amortized estimators go further: describe a context, get context-specific inferences.

Statistical estimation is being absorbed into context. In a new 90-page review, we provide a unified view of statistical and foundation models through the lens of context-adaptive inference! ...

With this, new methods could be evaluated in <10 lines of code on DDR-Bench and DTR-Bench, and I can be sure agents aren't gaming the environment or evals.

We found that CellVS-Net substantially improves drug effect representation on both DDR-Bench (below) and DTR-Bench.

We found shortcomings in both molecular and cell-level methods on these benchmarks, motivating us to go beyond static expression snapshots. This led to CellVS-Net, an amortized estimator that predicts how gene networks restructure in response to unseen perts.

Drug-Target Retrieval Bench (DTR-Bench) evaluates a method’s ability to reconstruct drug-target interactions from separate perturbation modalities that act on shared mechanisms, bridging the gap between cell-level methods and classic molecular screens.