Neural optimal transport handles the unpaired data that single-cell perturbation screens produce, but most OT maps cannot condition on the treatment context. We propose the Conditional Monge Gap, which learns OT maps conditioned on arbitrary covariates (drug, dosage, cell type, or combinations). Aggregating across conditions turns this into cross-task learning, unlocking strong generalization to unseen drugs and dosages and capturing the heterogeneity of the perturbed population far better than other conditional models.

An extension of our previous work (see below) on using (ATR-FTIR)Spectroscopy and Deep Learning for prediction of Idiopathic Recurrent Spontaneous Miscarriage (IRSM). This work focuses on the classification of IRSM using ATR-FTIR Spectroscopy, which is a non-invasive and cost-effective technique and improves on our previous work on using Raman Spectroscopy in a similar problem setting.

Using Raman Spectroscopy and Machine Learning for prediction of Idiopathic Recurrent Spontaneous Miscarriage (IRSM). Improved this work using ATR-FTIR Spectroscopy in a follow-up study, and currently working on a multi-omics approach to the same problem.

Using different statistical tests and empirical analyses to identify risk factors associated with mortality in Hypersensitivity pneumonitis, a rare lung disease. Checking for Publication bias and heterogeneity in the data, and using meta-analysis to combine results from different studies.

Can you design a protein to bind a stretch of DNA of your choosing? My Master's thesis argued that you can get surprisingly far by learning the distribution of DNA-binding protein sequences with a discrete diffusion model, and then steering it at sampling time with Seq2Contact as the guidance signal. The nice thing about that framing is how many design problems it absorbs: inpainting a DNA-binding domain between existing functional domains, CRISPR-Cas design, or running the whole thing backwards to ask which DNA a given protein would prefer. Some of the designs went on to express in the lab. In collaboration with Prof. Riddhiman Dhar at IIT Kharagpur; the image shows ESMFold structures of sampled (not cherry-picked) sequences, coloured by confidence.

When cAMP binds CAP, the protein rearranges somewhere else entirely, and transcription follows. The question here was whether that line of communication can be read off the dynamics rather than guessed at from structure: simulate the complex, then look for residues that move together. Three candidate allosteric networks fell out, via AlloReverse. The harder version — learning the dynamics directly, so that allosteric discovery generalises past one system — is still open. In collaboration with Prof. Soumya De at IIT Kharagpur.

Worked on using the Geneva HRCT dataset to segment lungs and classify diseases using Deep Learning. The final model uses a U-Net architecture for segmentation and a CNN for classification. The model was trained on a subset of the dataset and tested on CT scan images obtained externally through collaborations with Hospitals.

Worked on using Calcium ion imaging time-series data to extract functional networks in beta-islets of the pancreas. Used Voronoi Delaunay triangulation (see image) to extract the network, and used graph theory to analyze the network. Code is incomplete and will be updated soon.