Concept · artificial-intelligence
Generative models
4 studies1 discoveryEvidence last moved Sep 27, 2026
A generative model learns how data are produced, so it can create new samples or explain observations through hidden variables. The evidence here is from computational studies using a GAN to enhance Hi-C maps, a VAE for cancer drug response, a diffusion model for disordered protein ensembles, and a mechanistic brain model whose parameters were used as classification features.
Generative models are often discussed only as content creators. These studies show they are also used to fill in low-quality data, speed up simulation and build interpretable features, and that they inherit the limits of what they were trained on.
Studies
4
Findings
4
4 supporting · 0 challenging · 1 qualifying citations
Open tensions
1
Latest change
Concept page published
Generative models
Currently
What we know
- Adversarial training helped recover realistic high-resolution maps from sparse data.
- Generative models can act as fast stand-ins for expensive simulations.
- A generative latent space can improve prediction even when the modelled mechanism adds less.
- A good generative model can give compact, interpretable features.
Largest unresolved question
Generative models trained on simulations or experiments reproduce their training source, including its biases: idpSAM underestimated chain size compared with experiments exactly as its simulations did, and DeepHiC can be no better than the deepest available experimental data.
Common misconceptions
A generative model's output is new, ground-truth data.
It is a learned approximation: idpSAM inherited its simulations' compaction bias, and DeepHiC's low-coverage inputs were simulated by downsampling, which may not capture real shallow-experiment biases.
Any generative model works equally well as a feature extractor.
In the fMRI study, accuracy fell from 98% to 77% with a feed-forward model and to 59% using only the right hemisphere, so the choice of model matters.
Related
Claim ledger
What the evidence shows
Drawn from 4 studies in this library. Mix labels say which citation roles are present; they are not a strength score. Supports means evidence for a finding; Challenges means evidence against a stated position; Qualifies marks scope.
Adversarial training helped recover realistic high-resolution maps from sparse data.
DeepHiC, a GAN, raised genome-wide structural similarity on downsampled Hi-C data to 0.89 versus 0.71 for HiCPlus and 0.15 for raw input, and a generator trained without the adversarial loss gave unstable results.
Generative models can act as fast stand-ins for expensive simulations.
A diffusion model (idpSAM) generated a 10,000-conformation ensemble of a disordered peptide in about 4 minutes versus hundreds of CPU hours of simulation, and beat a GAN baseline trained on the same data.
A generative latent space can improve prediction even when the modelled mechanism adds less.
Dr.VAE, which models drug-induced expression changes in a latent space, beat standard classifiers on at least 14 of 26 drugs and lost on three, though much of the gain seemed to come from a better latent representation.
A good generative model can give compact, interpretable features.
Using the fitted parameters of a biologically plausible brain connectivity model as features classified 36 of 37 people (aphasia patients vs controls) correctly, beating conventional fMRI features (62%–83%); less plausible models did worse.
- Can a brain-connectivity model make better fMRI classifiers?— Only 11 patients from one dataset, and groups differed in age.
Debates
Tensions and limits
Some items are genuine disagreements on the same question. Others mark different assays, populations, or outcomes.
Generative models trained on simulations or experiments reproduce their training source, including its biases: idpSAM underestimated chain size compared with experiments exactly as its simulations did, and DeepHiC can be no better than the deepest available experimental data.
Generative models trained on simulations or experiments reproduce their training source, including its biases: idpSAM underestimated chain size compared with experiments exactly as its simulations did, and DeepHiC can be no better than the deepest available experimental data.
- Can a diffusion model generate shapes of floppy proteins?
- Can a GAN sharpen cheap, low-coverage chromosome contact maps?
Study Role Design N Population Outcome Can a diffusion model generate shapes of floppy proteins? Supports Computational / modellingAutoencoder plus latent denoising diffusion model trained on simulation ensembles of disordered peptides, tested on held-out sequences against simulation, a GAN baseline and experiment No single N; training uses over a thousand simulated peptide sequences, and experimental Rg comparison uses 10 test peptides Intrinsically disordered protein regions (simulated Cα ensembles) Agreement of generated ensembles with reference simulations (distance, torsion and Rg distributions), sampling speed and agreement with experimental Rg Can a GAN sharpen cheap, low-coverage chromosome contact maps? Supports Computational / modellingcGAN trained on chromosomes 1-14 of GM12878 Hi-C data downsampled from 10-kb matrices, tested on held-out chromosomes and other cell lines against HiCPlus, HiCNN and Boost-HiC. No single N; training and testing use Hi-C matrices from several cell lines (GM12878, its replicate, K562, IMR90) split by chromosome, evaluated on 1 Mb sub-regions. Public human (and mouse) Hi-C chromatin contact datasets Structural similarity (SSIM) and Pearson correlation to real high-resolution matrices; accuracy of downstream loop and TAD-boundary detection
PaperFren reads this as a limit on how far one study travels — different assays, populations, or outcomes — not a forced fight between papers.
Timeline
How understanding moved
Study years are when the paper was published. Evidence edits are dated changes to this page's claims. Explanations are when PaperFren added a Discovery — not a claim that the science happened that day.
2026
- Protein structure predictors get the answer without learning how proteins fold
Concept page published
Generative models
Change log
What changed
Dated edits to this page's evidence: studies added or removed from a claim, claims added or withdrawn, and new explanations tagged here. Rewordings are not listed.
- Concept page published
- Protein structure predictors get the answer without learning how proteins foldEvidence: Preliminary
Papers
4 studies in this library bear on Generative models, ordered by citations.
- Does modelling how drugs change gene expression help predict response?
A variational autoencoder that learns both how drugs shift gene expression and whether cells respond predicted drug sensitivity better than standard classifiers for over half of the drugs tested.
- Can a brain-connectivity model make better fMRI classifiers?
Classifying people by the connection strengths of a fitted brain model, rather than raw voxel activity, identified aphasia patients almost perfectly and pointed to which connections mattered.
- Can a GAN sharpen cheap, low-coverage chromosome contact maps?
A generative adversarial network rebuilt sharp high-resolution chromosome contact maps from as little as 1% of the sequencing reads, beating earlier CNNs that produced blurry results.
- Can a diffusion model generate shapes of floppy proteins?
A diffusion model trained on simulations generated realistic shape ensembles for disordered peptides it had never seen, far faster than simulation, but it inherited the simulations' bias toward overly compact shapes.
Compare studies
Select 2–10 studies. Design and N are labels, not a ranking.
Nothing selected yet.
Questions
What is still open
Generative models trained on simulations or experiments reproduce their training source, including its biases: idpSAM underestimated chain size compared with experiments exactly as its simulations did, and DeepHiC can be no better than the deepest available experimental data.
Ask PaperFren about Generative models
Study this conceptflashcards and short-answer questions
Give two uses of generative models in biology besides making new samples for their own sake.
They can enhance sparse data, as DeepHiC turned downsampled Hi-C maps into ones close to deep experiments, and they can provide features, as generative embedding used fitted brain-model parameters to classify aphasia at 98% balanced accuracy. They can also replace slow simulations, as idpSAM produced disordered-protein ensembles in minutes rather than CPU-days.
Why can a fast generative surrogate never be more accurate than its training data source?
It learns to reproduce the distribution it saw. idpSAM matched its reference simulations closely but, like them, underestimated chain size against experiment; a biasing fix reduced the error but broke chains in about a third of larger peptides.