Foldiff
Ranking validation on new protein families, better uncertainty assessment and a loop from proposed substitutions to measured stability/activity and the next experimental series.
Project brief · for researchers
From data to
the next
experiment
Two web tools for protein and cell research. Select testable hypotheses, analyse experiments and retain the evidence behind your next decision.
The goal is to make experimental planning and analysis clearer and more reproducible: assemble data, apply an appropriate method and show where the evidence ends.
Foldiff / Oligoplan is the shared site; Foldiff is the protein application and CellFate the cell application. There is currently no automatic model from a protein substitution to a cellular response. The tools serve laboratories, not treatment selection.
Proteins · engineering
Hypotheses for laboratory testing
Which substitutions are worth testing to try to improve protein stability?
For protein engineers, enzymologists and industrial R&D labs. The service turns known sequences and structures into a reasoned candidate list and materials for experimental planning.
| Task | Input | Output |
|---|---|---|
| Enzyme stabilization | Protein and selected homologue via UniProt; CDS for primers. | Candidate substitutions, ΔΔG estimates, combinations and draft primers for validation. |
| Comparison & context | Two proteins, a collection or a specific substitution. | 3D superposition, similarity, local environment and functional annotations. |
| Donors & PAE blocks | UniProt ID; a full EC number for donor search. | A bounded homologue search; separately, structure-model blocks based on PAE. |
| WT / pH measurements | Wild-type and variant measurements, conditions and replicates. | Means, spread, WT comparison; maximum activity among measured pH points. |
What powers the calculation
Sequences and annotations come from UniProt, precomputed predicted structures from AlphaFold DB, and published stability measurements from FireProtDB. Foldiff does not run AlphaFold. A linear model uses structural and chemical features to estimate ΔΔG; ThermoMPNN provides an additional neural-network channel when available. Disagreements remain visible.
ΔΔG does not imply a melting-temperature or activity gain. Combinations and primers require validation; PAE blocks are not established domains. Donor search does not establish which homologue is most thermostable.
Cells · experiment analysis
Distinguish treatment effects
Did treatment stop division, increase death, reduce secretion — or only change a metabolic signal?
For cell laboratories, senescence researchers, oncology screens and CROs. Seven modes help analyse existing results, check comparability and identify missing measurements.
| Section | User data | Practical output |
|---|---|---|
| Fibroblasts · 2 formats | Raw measurements or processed values / summaries. | Senescent and comparable non-senescent cultures: cell counts, secretion, ATP, OCR/ECAR; normalization when supported by the data. |
| Oncology · response | Doses, time, controls, counts and separate markers. | Treatment response, growth-rate correction (GR), comparison with a suitable nonmalignant model. A PRISM catalog of published candidates. |
| Oncology · dynamics | Cell counts at multiple time points. | Growth and DIP rates over a selected interval; observed treatment dynamics. |
| Oncology · events, 2 modes | Full division/death accounting or first outcomes with incomplete follow-up. | Separate event rates; competing risks with censoring for first outcomes. |
| Oncology · combinations | An A + B dose matrix, individual treatments and controls. | Deviation from Bliss and HSA references, without claiming clinical synergy. |
Why separate the readouts
ATP decreased. What happened to the cells?
Illustrative values, % of the corresponding control.
Input: CSV, TSV, Excel or a saved JSON project, plus model context, assay, units, controls, doses, times and independent replicates. Microscopy images, FCS and raw sequencing data are not currently processed.
Calculations use explicit statistical methods: summaries, paired comparisons, normalization, GR/DIP, competing risks and Bliss/HSA. Numerical results do not require an LLM. Outputs include plots, limitations and a report; exports include HTML, CSV and an archive with data, calculation code and checksums.
Cell models, candidate sources and project storage
The fibroblast mode covers human fibroblasts. Oncology covers established human cell lines in 2D monoculture with model metadata. Primary cells, organoids and arbitrary tissues are not validated scenarios.
PRISM 19Q4 v4 is a fixed snapshot of published responses in nine lines: A549, HCC827, MCF7, MDA-MB-231, HCT116, HT-29, PANC-1, A375, SK-OV-3. It supports follow-up research, not response prediction for a new line or patient. User tables from other established lines are supported when mode requirements are met.
Tables are sent over HTTPS for transient computation on a server in Russia and are not written to a CellFate database. My Projects uses browser-local storage without cross-device synchronization. Export JSON for transfer and backup.
Evidence · reproducibility
What has actually been checked
Arithmetic testing, reproduction of published data and biological usefulness are different levels of evidence. The first two exist for bounded scenarios. An independent laboratory pilot of the whole product has not yet been completed.
| Scope | Evidence | Limits |
|---|---|---|
| Foldiff · prediction | Saved FireProtDB split metrics: Spearman 0.452 on validation and 0.364 on test. | Ranking varies across splits. This is not a mutation success probability or independent product trial. |
| Fibroblasts | Wakita 2026: 79 groups, 158 mean/SD checks. Victorelli 2023: 133 observations, 28 means and 28 SDs. | Selected panels, not full paper replication. Wakita SSI matched in 4 of 6 conditions at a 0.001 tolerance. |
| Dynamics & combinations | Harris 2016: 360 measurements, 20 series; slope checks. Meyer 2019: 720 rows, 60 dose pairs; Bliss/HSA checks. | Independent recomputation of selected data. Not generalization across lines or replication of the authors’ MuSyC model. |
| First outcomes | Cornwell 2016: 689 founder cells, 15 groups; 1,360 numerical competing-risk comparisons. | The incomplete-follow-up mode was checked. Full division/death accounting has only synthetic-data checks so far. |
There is no proven overall accuracy percentage. Formula-recomputation error does not measure biological prediction accuracy. Missing measurements must limit a conclusion, rather than be replaced by an assumption.
Sources and methods
UniProtAlphaFold DBFireProtDBThermoMPNNHafner 2016 · GRWakita 2026Victorelli 2023Harris 2016 · DIPMeyer 2019Cornwell 2016PRISM 19Q4 v4
The validation counts above come from internal project audits using selected data from these studies, not endorsements of this service by the authors.
Development · collaboration
The next step involves a laboratory
The next shared priority is independent researcher evaluation on a familiar experiment: do results agree with established analysis, are limitations clear, and does the report help plan the next experiment?
CellFate
Validate full event accounting on a suitable published experiment, then add cell scenarios with their own requirements and reference datasets. Shared project storage and collaboration are a separate stage.
Planned extensions · not yet available
- MetaboFate
- A CellFate extension for constrained metabolic-flux models when exchange measurements are sufficient. Current ATP/OCR/ECAR normalization is only the first level.
- ImmuneClearance
- Target-cell clearance in co-culture: separate populations, contacts and outcomes. It requires a dedicated data format and validation.
- Reprogramming QC
- Quality control for partial reprogramming: age-associated markers alongside cell identity, viability and function. It would not automatically generate a “safe rejuvenation protocol”.
Start a conversation
Evaluate a familiar problem
Prepared examples are available without registration. A shared account is intended for your own calculations. CellFate is currently free; Foldiff uses a one-off payment for a completed result under its published pricing.