Arian Amani

Arian Amani

ML Researcher

Apheris

Biography

I work on machine learning for biology, and I care most about models that explain why cells respond to a drug or perturbation the way they do, not only what they will do.

I recently joined Apheris as a machine learning researcher for molecular privacy. The work is about letting organizations study sensitive molecular data together without sharing any data, and I also contribute to Apheris’s virtual cell network. Before Apheris I spent about a year and a half at AI VIVO building generative models for drug discovery.

In parallel, from late 2022, I was a data scientist at the Wellcome Sanger Institute (Lotfollahi Lab), where I co-first-authored CellDISECT (bioRxiv 2025; under review at Nature Biomedical Engineering), co-authored SP-FM (arXiv 2026), fine-tuned single-cell foundation models for perturbation prediction, and was lead maintainer of CPA.

Interests
  • Molecular Privacy
  • Federated Learning
  • Deep Generative Models
  • Causal Representation Learning
  • Flow Matching
  • Single-Cell Perturbation
  • Drug Discovery
Education
  • Applied Computer Science & Artificial Intelligence, 2026

    Sapienza University of Rome

  • BSc in Computer Science, 2023

    Amirkabir University of Technology

Current Work

  • Apheris: Molecular privacy research, which means letting several parties train on sensitive molecular data without sharing it using federated learning, and also contributing to Apheris’s Virtual Cell Network, which I presented at the scverse conference in October 2026.
  • CellDISECT (bioRxiv, 2025; under review at Nature Biomedical Engineering, code): A causal VAE that separates covariate effects in single-cell RNA-seq data, predicts counterfactual perturbations, and helps find new cell types.
  • SP-FM (arXiv, 2026): A flow-matching method that generalizes better to conditions the model has not seen, such as unseen perturbations.
  • Why Perturbation Prediction Needs Much Better Metrics (Medium, March 2026): An essay on why mean-predictor baselines expose weak evaluation metrics in perturbation prediction.

Work Experience

 
 
 
 
 
Apheris
Machine Learning Researcher, Molecular Privacy
June 2026 – Present 5 mos Rome, Italy (remote)
  • Research on privacy-preserving machine learning for sensitive molecular data, so that organizations can collaborate without sharing the underlying data
  • Contribute to Apheris’s virtual cell network
 
 
 
 
 
AI VIVO
Machine Learning Scientist
AI VIVO
December 2024 – June 2026 1 yr 7 mos Rome, Italy (remote)
  • Built deep learning and generative models for drug discovery
  • Deployed and scaled ML pipelines on GCP with PyTorch Lightning and Docker
  • Designed multi-modal pipelines that combine molecular structure with biological assay data
  • Worked with PyTorch, Lightning, RDKit, and HuggingFace, and with computational chemistry tools such as BioSolveIt, AutoDock Vina, and Boltz-2
 
 
 
 
 
Wellcome Sanger Institute
Data Scientist
November 2022 – June 2026 3 yrs 8 mos Hinxton, United Kingdom (remote)
  • Lotfollahi Lab: co-first author of CellDISECT, a causal VAE for disentangled single-cell representations and in silico perturbation prediction (2M+ cells, 100+ counterfactual conditions). The manuscript is under review at Nature Biomedical Engineering.
  • Fine-tuned single-cell transformer foundation models ~300M parameters (scFoundation, Geneformer, scGPT) with LoRA and prompt tuning on 1500+ perturbation conditions
  • Was lead maintainer of CPA (Compositional Perturbation Autoencoder): code review and releases for the single-cell community
  • Worked with PhD students and postdocs on preprints, open-source tools, and PyTorch-based single-cell methods
 
 
 
 
 
Virasad
Computer Vision Engineer
January 2022 – May 2022 5 mos Tehran, Iran
  • Reached over 95% accuracy on tasks with very little data (15 images per class)
  • Worked on 5 client projects and led several of them on my own, improving the development pipelines along the way

Projects

CellDISECT
CellDISECT (Cell DISentangled Experts for Covariate counTerfactuals) is a causal generative model for disentangled single-cell representations and counterfactual perturbation prediction. Preprint on bioRxiv; under review at Nature Biomedical Engineering.
CellDISECT
CPA (Compositional Perturbation Autoencoder)
CPA is a deep generative framework to learn effects of perturbations at the single-cell level. It performs OOD predictions of unseen combinations of drugs, learns interpretable embeddings, estimates dose-response curves, and provides uncertainty estimates.
CPA (Compositional Perturbation Autoencoder)

Recent Publications

Quickly discover relevant content by filtering publications.
(2026). Shortest-Path Flow Matching with Mixture-Conditioned Bases for OOD Generalization to Unseen Conditions. arXiv.

PDF Cite DOI arXiv

(2025). Integrating multi-covariate disentanglement with counterfactual analysis on synthetic data enables cell type discovery and counterfactual predictions. bioRxiv.

PDF Cite Code Project DOI bioRxiv

Teaching Experience

 
 
 
 
 
Teaching Assistant
Sharif University of Technology
September 2022 – August 2023 1 yr Tehran, Iran
  • Machine Learning for Bioinformatics (graduate course), Spring 2023
    • Prepared teaching material on CNNs and autoencoders, designed assignments, and ran class contests.
  • Introduction to Machine Learning, Fall 2022
    • Designed and graded assignments for 150 students and ran a workshop on variational autoencoders.
 
 
 
 
 
Teaching Assistant
Amirkabir University of Technology
September 2021 – March 2022 7 mos Tehran, Iran
  • Introduction to Image Processing and Neural Networks, Fall 2021
    • Gave workshops and lectures on OpenCV and deep learning to 80 students.
  • Advanced Programming with C++, Spring 2022
    • Designed assignments and projects for 90 students and graded their submissions.

Accomplish­ments

Coursera
Deep Learning Specialization
  • Neural Networks and Deep Learning
  • Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization
  • Structuring Machine Learning Projects
  • Convolutional Neural Networks
  • Sequence Models
See certificate