# mlns project pitch
time: 2026-02-09 03:49:22

the goal of this project is to aim for the two outputs, namely
reaction center localization that is we want to predict the respective reaction centres in a reaction
and the second one is possibly mechanistic step prediction
and additionally we would like to get the confidence or the uncertainty score for them respectively

so what was the problem that we are trying to address?
the pain point here is that chemistry cannot be generalized to one distribution cause reaction families differ by alot and also rare or new chemistries have very few labelled examples

the first point which addresses is either using standard GNNs or sequence models, but the thing is they need a full retraining or lots of labels per class for generalization otherwise if a reaction family is under-represented, which is the case most of the times, we get a lot of distribution shift hence giving rise to poor performance

so this idea's core need is that we need a few shot model which can adapt to the set of reactions by just specifying a k examples which can be practical

so in this we can represent each reaction as a graph where atoms are nodes and bonds are edges and we can embed further data into these. so the label is either node level for reaction center detection and edge level for step wise prediction

we use graphPFN which is a prior data fitted network for graphs which is especially designed for in context learning on node prediction tasks
the key feature here is you can give labelled examples as context and it predicts the labels for a new query graph without gradient updates only with optional light fine tuning

so the workflow at the inference time would something be like, here area the k example reactions from this family with their reaction centres and mech steps labelled, and given this set of data now predict the center/step for a new reaction from that family

technically graphpfn used a transformer backbone with graph aware message passing adapters, and is pretrained with many synthetic graph datasets so it can do this in context learning