Machine Learning Scientist at Synlico Boise, Idaho
At Synlico, I work on deep reinforcement learning methods for causal discovery in single-cell genomics, with a focus on inferring gene regulatory networks as directed causal graphs.
My broader research focuses on deep neural architectures for graph-structured data, including graph transformers, geometric deep learning, molecular representation learning, and efficient global attention.
Current research Synlico
At Synlico, I research deep reinforcement learning-based causal discovery methods for inferring gene regulatory networks as directed causal graphs from high-dimensional single-cell genomics data.
This work includes GPU-accelerated training and evaluation workflows for causal structure learning and collaboration across machine learning and bioinformatics.
Selected research
My prior research through the Rensselaer-IBM AI Research Collaboration focused on graph transformer architectures for structural graph representations, molecular geometry, and efficient attention.
EGT is an early graph transformer architecture that introduced evolving structural edge channels for joint node and pair representations across graph-level prediction, edge classification, and link prediction.
Author and maintainer of the PyTorch and TensorFlow implementations.
SSA subsamples transformer self-attention during training and formulates the Information Pathways Hypothesis, reducing attention memory and computation by 4–8× across graph, vision, and language tasks.
TGT extends EGT with third-order interactions among groups of three atoms and a 2D-to-3D framework that predicts interatomic distances from 2D molecular graphs for property prediction.
Author and maintainer of the official PyTorch implementation.
TGT has remained the top-ranked non-ensemble submission on PCQM4Mv2 for more than two years and records the highest EwT among direct methods on the Open Catalyst 2020 (OC20) IS2RE leaderboard.
Selected publications
ICML 2024 · PMLR 235:20768–20792
KDD 2023 · pp. 810–821
KDD 2022 · pp. 655–665
I am based in Boise, Idaho. For research and professional correspondence, email snirjhar@gmail.com.