CELL-FM

A virtual microscopy model that bridges microscopy images and protein sequences.

Paper: Zheng, Hong & Huang, Virtual experiments bridge sequence and microscopy with generative models, bioRxiv (2026), doi:10.64898/2026.09.13.751243
Condensate Titration CELL-FM virtual CondenSeq experiments on 66-aa IDPs
Titration
Run the analysis to measure them. AUC is how much of the scanned range the IDP spends condensed; AAC how much of that it loses to re-dissolving at high expression. Both are normalised to [0, 1].
AUC
—
AAC
—
Titration curve
One grey dot per generated image, called condensed (1) or diffuse (0) by the ViT; the line is their moving average, the condensate probability. Blue is the area under it (AUC), orange the condensation lost to re-dissolving past the peak (AAC).
Channel
The reference nucleus every image is generated in. Run the analysis to add the IDP.
1 512
Samples
Nucleus image settings
0 100
1 150
0.5 1.5
IDP image settings
0 100
1 150
0.5 1.5
Amino acid substitution scan
Every residue type in 100 held-out IDPs replaced by each other type, each mutant titrated as above, and its AUC and AAC compared with the wild type's. A cell is the mean change over the IDPs that contain the FROM residue: left the paper's Fig. 5d (AUC), right its Supplementary Fig. 5c (AAC). Click any cell for the distribution behind it, or a residue on either axis to compare whole rows (Mutate FROM) or columns (Mutate TO) below.
F → S
Every F in the IDP replaced by S · 84 of 100 IDPs contain F
Mean Δ AUC
-0.0271
Median Δ AUC
-0.0173
Mean Δ AAC
-0.0120
Median Δ AAC
-0.0079
Per-IDP values
One row per IDP behind the selected cell; click a header to sort. Precomputed: the whole scan is condenseq/substitution_scan.npz in the model repo, checked against the matrices the paper's figures were drawn from (ana_all_mutation_log_scale.py, AAC being compute_area_if_no_reentrant_minus_auc).
CondenSeq index
F residues mutated
WT AUC
F→S AUC
Δ AUC
WT AAC
F→S AAC
Δ AAC
wild-type sequence
10036
4
0.3746
0.3296
-0.0449
0.1097
0.0931
-0.0167
ACKYGDKCQFAHGIHELRSLTRHPKYKTELCRTFHTIGFCPYGPRCHFIHNAEERRALAGARDLSA
Compare rows or columns
A row is everything one residue becomes (Mutate FROM), a column everything that becomes one residue (Mutate TO). Click residues on a matrix's axis to add or remove them, up to eight; clicking on the other axis starts a comparison there. Each point is a matrix cell, with its 95% confidence interval over the IDPs; hover for the numbers.
Compare
Mutate TO
FROM F Δ AUC
FROM F ± 95% CI (AUC)
FROM F Δ AAC
FROM F ± 95% CI (AAC)
FROM Y Δ AUC
FROM Y ± 95% CI (AUC)
FROM Y Δ AAC
FROM Y ± 95% CI (AAC)
FROM W Δ AUC
FROM W ± 95% CI (AUC)
FROM W Δ AAC
FROM W ± 95% CI (AAC)
FROM R Δ AUC
FROM R ± 95% CI (AUC)
FROM R Δ AAC
FROM R ± 95% CI (AAC)
R
-0.0126
0.0094
-0.0082
0.0041
-0.0186
0.0074
-0.0048
0.0035
-0.0296
0.0105
-0.0124
0.0044
-0.0031
0.0091
-0.0002
0.0024
How a titration is computed
CELL-FM generates one image per expression level along a log-spaced ladder, a ViT classifier calls each image condensed or diffuse, a moving average over the calls is the titration curve, and the curve is integrated against log expression into AUC and AAC.
Each image is CELL-FM's prediction from the IDP sequence and a protein expression level alone, drawn in one shared cell: the nucleus channel is the same real CondenSeq nucleus for every image, so differences come from the sequence and the level. Sampling is fixed at 512 levels from 20 to 2048 a.u., 100 ODE steps and seed 6, as in the offline runs. The IDP library and the condensate phenotypes the models learned from are CondenSeq's (Kappel et al., Nat. Methods 2025).

If you use CELL-FM, please cite:

@article{zheng2026virtual,
  title   = {Virtual experiments bridge sequence and microscopy with generative models},
  author  = {Zheng, Dihan and Hong, Kibeom and Huang, Bo},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.09.13.751243},
  url     = {https://www.biorxiv.org/content/10.64898/2026.09.13.751243v1}
}