A deep learning emulator for single-cell informed deconvolution and state prediction of cancer cells in clinical cohorts

Abstract Motivation Characterizing the intratumoral heterogeneity of Lung Adenocarcinoma (LUAD) from bulk transcriptomics is critical for precision medicine, yet conventional deconvolution methods are limited by high computational latency and sensitivity to transcriptomic noise, hindering their application to large-scale clinical trials. To overcome these barriers, we present a Deep Neural Network (DNN) emulator that shifts the deconvolution task from iterative optimization to rapid functional inference. Results Utilizing a harmonized LUAD single-cell atlas, we engineered a synthetic training environment of 1,000 heterogeneous bulk profiles to train a pathway-guided architecture. By internalizing the complex transformation mappings of marker-based deconvolution, our model reproduces TOAST estimates of four previously defined LUAD cancer cell states on held-out synthetic samples (ensemble mean Spearman ρ = 0.87, MAE = 0.067), agrees with the reference-based method BayesPrism about as well as TOAST itself does (mean ρ = 0.57 for both), and runs at 1.4 ms per sample for a single model and 12.5 ms for the ensemble, similar to TOAST and about 100 to 900 times faster than BayesPrism. We applied this emulator to large-scale clinical cohorts to estimate the proportion of the high-risk, dedifferentiated state, characterized by hypoxia and EMT. The predicted proportion of this state was associated with shorter overall survival in all six cohorts (pooled hazard ratio 1.70, 95% CI 1.49–1.95, for top versus bottom tertile), suggesting that our emulator provides a scalable and biologically grounded framework for patient stratification and the translation of single-cell insights into clinical practice. Availability and implementation The analysis code is freely available on GitHub (https://github.com/bsoni08/emulator_luad_stratification) and implemented in Python using TensorFlow alongside R scripts. Contact Inquiries can be directed to [email protected], [email protected], or [email protected].

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-11726977/v1
Primary Topic
Single-cell and spatial transcriptomics
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A deep learning emulator for single-cell informed deconvolution and state prediction of cancer cells in clinical cohorts

Ines Grazina de Matos, Christopher Yau, Bhavesh Soni, Stephan Gasser
3 links
Research Square
Single-cell and spatial transcriptomics
preprint

A deep learning emulator for single-cell informed deconvolution and state prediction of cancer cells in clinical cohorts

Ines Grazina de Matos, Christopher Yau, Bhavesh Soni, Stephan Gasser
preprint en

Abstract

Abstract Motivation Characterizing the intratumoral heterogeneity of Lung Adenocarcinoma (LUAD) from bulk transcriptomics is critical for precision medicine, yet conventional deconvolution methods are limited by high computational latency and sensitivity to transcriptomic noise, hindering their application to large-scale clinical trials. To overcome these barriers, we present a Deep Neural Network (DNN) emulator that shifts the deconvolution task from iterative optimization to rapid functional inference. Results Utilizing a harmonized LUAD single-cell atlas, we engineered a synthetic training environment of 1,000 heterogeneous bulk profiles to train a pathway-guided architecture. By internalizing the complex transformation mappings of marker-based deconvolution, our model reproduces TOAST estimates of four previously defined LUAD cancer cell states on held-out synthetic samples (ensemble mean Spearman ρ = 0.87, MAE = 0.067), agrees with the reference-based method BayesPrism about as well as TOAST itself does (mean ρ = 0.57 for both), and runs at 1.4 ms per sample for a single model and 12.5 ms for the ensemble, similar to TOAST and about 100 to 900 times faster than BayesPrism. We applied this emulator to large-scale clinical cohorts to estimate the proportion of the high-risk, dedifferentiated state, characterized by hypoxia and EMT. The predicted proportion of this state was associated with shorter overall survival in all six cohorts (pooled hazard ratio 1.70, 95% CI 1.49–1.95, for top versus bottom tertile), suggesting that our emulator provides a scalable and biologically grounded framework for patient stratification and the translation of single-cell insights into clinical practice. Availability and implementation The analysis code is freely available on GitHub (https://github.com/bsoni08/emulator_luad_stratification) and implemented in Python using TensorFlow alongside R scripts. Contact Inquiries can be directed to [email protected], [email protected], or [email protected].

Research Square
F. Hoffmann-La Roche AG (CH), University of Oxford (GB)
Good health and well-being
Single-cell and spatial transcriptomics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon