PerturbLDM: conditional latent diffusion for modelling single-cell perturbation responses

Verdict. Here we introduce PerturbLDM, a latent-diffusion framework for conditional generation of single-cell transcriptional responses. It intersects mitochondrial stress/dysfunction themes (dynamics (fission/fusion); OXPHOS / ETC).

What the authors report

Single-cell perturbation profiling maps intervention-induced phenotypes, yet experiments measure only a fraction of the perturbation-context space. Learning context-dependent perturbation effects could enable response prediction beyond measured conditions.

Key results stated in the abstract include the following. Here we introduce PerturbLDM, a latent-diffusion framework for conditional generation of single-cell transcriptional responses. In PBMCs, it captured six of seven interferon and antiviral programmes and the interferon-associated FAO-OXPHOS programme more accurately than scGen. Together, these results support conditional response generation across data scales and biological settings.

Why it matters for mitochondrial biology

Within mitochondrial research, this work maps primarily to OXPHOS, mitochondrial dynamics, redox biology, therapeutics. It is relevant to mitochondrial dysfunction discourse because the abstract invokes dynamics (fission/fusion), OXPHOS / ETC. That does not by itself establish a validated disease mechanism; it indicates thematic proximity. Because a therapeutic or interventional angle is present, the piece is of interest for mitochondrial-targeted drug hypothesis generation—subject to full-text validation of endpoints and safety context. OXPHOS/ETC involvement, if confirmed, would place the work in the core of bioenergetic pathophysiology rather than peripheral organelle biology. Server: biorxiv. Posted 2026-08-12. Synthesis confidence is bounded by abstract completeness.

Study design (abstract-level)

Following Tahoe-100M pretraining, it outperformed leading methods across 13,942 held-out combinations of observed drugs, doses and cell lines, with higher matched-control effect correlation than an additive marginal baseline in 95.2% of conditions. The Tahoe-100M-pretrained model was further used to rank PANACEA compounds by pathway similarity, placing shared-mechanism pairs among nearest neighbours.

Principal findings

1. Here we introduce PerturbLDM, a latent-diffusion framework for conditional generation of single-cell transcriptional responses.
2. In PBMCs, it captured six of seven interferon and antiviral programmes and the interferon-associated FAO-OXPHOS programme more accurately than scGen.
3. Together, these results support conditional response generation across data scales and biological settings.

Limitations of this brief

• This Mitos brief is an abstract-level synthesis of a preprint; it is not peer review and not a substitute for reading the full paper.
• Preprint status: findings may change with revision or journal review.
• Effect sizes, n numbers, statistics, and full experimental controls are typically incomplete at abstract resolution.
• Primary source: biorxiv DOI 10.64898/2026.08.07.743610 (posted 2026-08-12).

Open scientific questions

• Which specific experimental panels in the full paper establish the strongest causal claim, and how robust are the controls?
• Are OXPHOS defects primary drivers or secondary consequences of broader cellular stress?
• What dose, timing, and off-target profile would be required to take the intervention seriously as a therapeutic hypothesis?
• How do these findings sit relative to prior literature on the same pathway—replication, contradiction, or incremental extension?

Bottom line

For mitochondrial biologists focused on OXPHOS, mitochondrial dynamics, redox biology, this preprint is worth full-text review if the topic matches your program. Abstract-level takeaway: Here we introduce PerturbLDM, a latent-diffusion framework for conditional generation of single-cell transcriptional responses. Confirm methods, effect sizes, and controls in the full PDF before citing the result as established.

Bibliographic record

Abstract-based editorial synthesis by Mitos. Not peer review.