Fatty-acid supply from fat tissue
Fat tissue that resists insulin keeps releasing fatty acids, which the liver takes up and stores.
- Pioglitazone predicted 10% less liver fat
Open
Why does fat build up in the liver in type 2 diabetes?
Excess triglyceride accumulates in liver cells (steatosis), driving inflammation and scarring. It affects a large share of people with type 2 diabetes.
Real: the literature, causal model, predictions, out-of-sample validation, knowledge gaps and protocols. Simulated: laboratory results, each labeled as such.
Where the mechanism fails
Liver fat is the phenotype; the question is where it comes from. A treatment can only work if it acts on the mechanism that is failing.
Where the liver's fat comes from in fatty-liver patients (isotope tracer study). Most of it arrives from fat tissue, which is why a fix aimed only at fat production in the liver can do little.
Fat production in the liver
Liver cells turn sugar into new fat (de novo lipogenesis), driven by insulin and SREBP-1c.
- Metformin predicted 10% less liver fat
- SGLT2 inhibitors predicted 10% less liver fat
Fat export and droplet handling
Liver cells package fat into VLDL for export and remodel their lipid droplets.
No current treatment acts on this mechanism. Human genetics (PNPLA3, TM6SF2) shows that impaired lipid export and droplet handling can cause fatty liver, yet no treatment modeled here targets it.
Body weight and energy balance
Weight loss lowers both the fatty-acid supply and fat production.
- GLP-1 receptor agonists predicted 18% less liver fat
- SGLT2 inhibitors predicted 10% less liver fat
- Lifestyle weight loss predicted 23% less liver fat
Predictions vs. trials
Each treatment's effect is predicted from mechanism alone, then compared with liver-fat trials the model has never seen (out-of-sample validation). An effect counts if liver fat falls by at least 10%.
Clinically meaningful reduction
2 of 3 trial outcomes predicted correctly from mechanism alone · Brier score 0.17
| Mechanism targeted | Held-out trials | Adverse effects | ||||
|---|---|---|---|---|---|---|
| Lifestyle weight lossDiet and exercise | Body weight and energy balance | 23% (7% to 44%) | 89% | Calibration data, not a test | – | – |
| GLP-1 receptor agonistsIncretin mimetics | Body weight and energy balance | 18% (6% to 33%) | 83% | Effective · Tang 2015, Diabetes Metab Res RevArmstrong 2016 (LEAN), Lancet | Correct | Gastrointestinal adverse events in 63.5% vs 34.3% on placebo (semaglutide 2.4 mg) Davies 2021 (STEP 2), LancetBoxed warning on the drug label DailyMed: Ozempic (semaglutide) label |
| PioglitazoneThiazolidinedione (PPARγ agonist) | Fatty-acid supply from fat tissue | 10% (-2% to 25%) · measured 54% | 46% | Effective · Belfort 2006, N Engl J MedTang 2015, Diabetes Metab Res Rev | Failed | Weight gain of almost 3 kg Clar 2009, PLoS OneBoxed warning on the drug label DailyMed: pioglitazone tablet label |
| SGLT2 inhibitorsGliflozins | Body weight and energy balance, Fat production in the liver | 10% (4% to 18%) · measured 25% | 46% | Inconsistent · Kuchay 2018 (E-LIFT), Diabetes CareTang 2015, Diabetes Metab Res Rev | Trials disagree | Genital tract infections, odds ratio 4.2 vs placebo Liu 2015, J Diabetes Complications |
| MetforminBiguanide | Fat production in the liver | 10% (2% to 20%) | 41% | No effect · Tang 2015, Diabetes Metab Res RevSaid 2017, Ann HepatolLavine 2011 (TONIC), JAMA | Correct | – |
Bar: 90% credible interval with the posterior mean. Dots: reductions measured in held-out trials (filled when inside the interval).
Causal model
Mechanisms in rows, treatments on the left, the clinical outcome on the right. Darker links carry stronger evidence in humans; select a link to see the studies behind it.
Darker = stronger evidence in humans · thicker = more studies · teal = part of a ranked knowledge gap · red dashed = unmeasured · ↓ = inhibits · • = includes your experiments
The 16 model parameters and their priors
| Basis | ||
|---|---|---|
| Direct effect of GLP-1 drugs on fat production in liver cells | 0% – 9% | No receptor found in mouse hepatocytes or in human liver expression data: expected near zero. Panjwani 2013, EndocrinologyGTEx (platform data): GLP1R |
| Direct effect of SGLT2 inhibitors on liver fat production | 1% – 23% | SGLT2 is not expressed in liver; lower insulin could still reduce fat production. Unmeasured. GTEx (platform data): SLC5A2 (SGLT2) |
| Drop in fatty-acid supply on metformin | 1% – 18% | Not measured by any study in the evidence base. Centered on a small effect and left wide on purpose. |
| Drop in fatty-acid supply on pioglitazone | 8% – 48% | A randomized trial shows free fatty acids fall and lipolysis becomes insulin-sensitive, but reports no percentage for delivery to the liver. Centered near 25%, wide. Gastaldelli 2009, Am J Physiol Endocrinol Metab |
| Liver-fat reduction per unit of weight lost | 0.9 – 5.1 | Wide prior (roughly 1–6), then fitted to the Look AHEAD randomized comparison (see calibration). Lazo 2010 (Look AHEAD), Diabetes Care |
| Metformin concentration reaching the liver | 19 µM – 86 µM | Therapeutic levels sit 10–100× below most experimental doses; portal levels run above plasma. He & Wondisford 2015, Cell MetabOwen 2000, Biochem J |
| Metformin concentration that halves fat production (IC50) | 44 µM – 3,956 µM | Cell and rodent studies show suppression at hundreds of µM to low mM; nothing measures it at patient concentrations, so the prior spans about 25 µM to 6 mM. Zhou 2001, J Clin InvestLi 2011, Cell MetabHe & Wondisford 2015, Cell Metab |
| Most fat production metformin can switch off | 40% – 89% | Strong suppression at high doses in rodent cells; the ceiling in human cells is unmeasured. Zhou 2001, J Clin InvestFullerton 2013, Nat Med |
| Share of liver fat delivered from fat tissue | 43% – 76% | Same patients. Donnelly 2005, J Clin Invest |
| Share of liver fat made by the liver itself | 15% – 37% | Stable-isotope labeling in fatty-liver patients. Donnelly 2005, J Clin Invest |
| Share of pioglitazone's weight gain that acts on the liver like ordinary weight gain | 14% – 87% | Unmeasured here. Part of the gain is fluid and fat under the skin, which should not load the liver. |
| Weight change on metformin | -1% – 3% | Not measured by any study in the evidence base; assumed small. |
| Weight gained on pioglitazone | 2% – 5% | Almost 3 kg across trials; converted assuming a typical 90 kg trial participant. Clar 2009, PLoS One |
| Weight lost on GLP-1 receptor agonists | 5% – 7% | Semaglutide 2.4 mg vs placebo in type 2 diabetes. Liver trials used other agents and doses; the model assumes a class effect. Davies 2021 (STEP 2), Lancet |
| Weight lost on SGLT2 inhibitors | 2% – 3% | 2.5 kg vs placebo at 1 year; converted assuming a typical 90 kg trial participant. Liu 2015, J Diabetes Complications |
| Weight lost with intensive lifestyle change | 7% – 10% | Look AHEAD: −8.5% vs −0.05%. Lazo 2010 (Look AHEAD), Diabetes Care |
Knowledge gaps
13 gaps identified by the model, ranked by how much they limit its predictions.
- Failed predictionPrediction fails for pioglitazone
A held-out trial measured a 54% reduction in liver fat; the model predicts 10% (90% interval -2% to 25%). Either a parameter is misestimated, most likely drop in fatty-acid supply on pioglitazone or share of pioglitazone's weight gain that acts on the liver like ordinary weight gain, or a mechanism is missing from the model.
- Failed predictionPrediction fails for SGLT2 inhibitors
A held-out trial measured a 25% reduction in liver fat; the model predicts 10% (90% interval 4% to 18%). Either a parameter is misestimated, most likely liver-fat reduction per unit of weight lost or direct effect of sglt2 inhibitors on liver fat production, or a mechanism is missing from the model.
- Poorly constrainedLiver-fat reduction per unit of weight lost: poorly constrained
Accounts for 93% of the predictive uncertainty for GLP-1 receptor agonists. 90% credible interval: 0.9 to 5.1. Wide prior (roughly 1–6), then fitted to the Look AHEAD randomized comparison (see calibration).
- Untargeted mechanismNo treatment targets fat export and droplet handling
Human genetics (PNPLA3, TM6SF2) shows that impaired lipid export and droplet handling can cause fatty liver, yet no treatment modeled here targets it.
- Unmeasured linkUnmeasured: Metformin → fatty-acid supply to the liver
Metformin might lower fatty-acid release from fat tissue. No study in the evidence base measures it. The model relies on a prior alone for this link.
- Poorly constrainedDrop in fatty-acid supply on pioglitazone: poorly constrained
Accounts for 70% of the predictive uncertainty for pioglitazone. 90% credible interval: 8% to 48%. A randomized trial shows free fatty acids fall and lipolysis becomes insulin-sensitive, but reports no percentage for delivery to the liver. Centered near 25%, wide.
- High doses onlyACC phosphorylation → de novo lipogenesis: shown only at supra-therapeutic doses
Every study behind this link used concentrations above those reached in patients.
- High doses onlySREBP-1c activity → de novo lipogenesis: shown only at supra-therapeutic doses
Every study behind this link used concentrations above those reached in patients.
- Non-human onlyMetformin → energy stress in liver cells: shown only in non-human systems
Tested at clinically relevant concentrations, but not yet in human cells or people.
- Conflicting evidenceConflicting evidence: Energy stress in liver cells → AMPK activation
Zhou 2001, J Clin Invest: Metformin activates AMPK in hepatocytes; AMPK is required for its effect on glucose output. But Foretz 2010, J Clin Invest: In liver lacking AMPK, metformin still lowers glucose output through the energy state.
- Conflicting evidenceConflicting evidence: SREBP-1c activity → de novo lipogenesis
Li 2011, Cell Metab: Blocking SREBP-1c processing reduces lipogenesis and lipid accumulation in hepatocytes. But He & Wondisford 2015, Cell Metab: Most mechanism studies used 10–100 times the concentrations patients reach.
- Conflicting evidenceConflicting evidence: PPARγ in fat cells → insulin resistance of fat tissue
He 2015, Metabolism: Thiazolidinediones inhibited lipolysis in rat fat cells, dose- and time-dependently. But Pravenec 2008, Int J Obes: Long-term pioglitazone increased fat-tissue lipolysis in rats, yet circulating fatty acids still fell through better insulin action and muscle oxidation.
- Poorly constrainedDrop in fatty-acid supply on metformin: poorly constrained
Accounts for 32% of the predictive uncertainty for metformin. 90% credible interval: 1% to 18%. Not measured by any study in the evidence base. Centered on a small effect and left wide on purpose.
Decisive experiments
The experiments that would most reduce uncertainty per dollar. Cell and invertebrate studies only. Each is reviewed by an advocate, a skeptic, a methods reviewer and an editor, and none is run without your approval.
1
Pioglitazone · Fatty-acid supply from fat tissue
Recommend with revisionsHow much does pioglitazone make fat cells hold on to fatty acids?
At plasma-level concentrations, pioglitazone cuts net fatty-acid release from insulin-resistant human fat cells by well over 25%, enough to explain the liver trials.
Expected to reduce predictive uncertainty by 36% · $7,000 · 12 days
- Cells
- Differentiated human preadipocytes made insulin-resistant with TNF-α, day 12
- Time
- 7 days of pioglitazone (PPARγ acts through gene expression), then 3 h assays
- Stimuli
- Pioglitazone 0, 0.1, 0.3, 1, 3, 10 µM; insulin 0 and 1 nM during the assay
- Response
- Labeled palmitate uptake and re-esterification, glycerol and fatty-acid release, insulin suppression of lipolysis
2
Metformin · Fat production in the liver
Recommend with revisionsThe same dose-response in HepG2 cells
HepG2 cells show the same concentration dependence, at a fraction of the cost.
Expected to reduce predictive uncertainty by 15% · $2,500 · 4 days
- Cells
- HepG2 hepatoma cells (STR-authenticated)
- Time
- 48 h treatment after 24 h attachment
- Stimuli
- Metformin at 8 concentrations in 25 mM glucose + 100 nM insulin
- Response
- De novo lipogenesis, lipid droplet area, lipogenic gene expression
3
Metformin · Fatty-acid supply from fat tissue
Recommend with revisionsDoes metformin lower fatty-acid release from human fat cells?
At 10–100 µM, metformin reduces stimulated lipolysis in human fat cells by under 10%.
Expected to reduce predictive uncertainty by 16% · $4,000 · 5 days
- Cells
- Differentiated human preadipocytes, day 12
- Time
- 24 h pre-treatment, then 3 h stimulated lipolysis
- Stimuli
- Metformin 0–100 µM; isoproterenol 1 µM
- Response
- Glycerol and fatty-acid release
4
Metformin · Fat production in the liver
Recommend with revisionsMetformin dose-response on fat production in primary human liver cells
At concentrations patients reach (10–100 µM), metformin cuts fat production in human liver cells by less than 10%.
Expected to reduce predictive uncertainty by 18% · $9,000 · 7 days
- Cells
- Cryopreserved primary human hepatocytes, 3 donors, sandwich culture
- Time
- 48 h treatment after 24 h attachment
- Stimuli
- Metformin at 8 concentrations in 25 mM glucose + 100 nM insulin
- Response
- De novo lipogenesis (13C-acetate into palmitate), lipid droplet area, FASN/ACACA/SCD1 expression
5
SGLT2 inhibitors · Fat production in the liver
Recommend with revisionsDo SGLT2 inhibitors act on human liver cells directly?
Empagliflozin has no direct effect on fat production in liver cells; any liver benefit comes through weight and insulin.
Expected to reduce predictive uncertainty by 14% · $6,000 · 6 days
- Cells
- Primary human hepatocytes, 3 donors
- Time
- 48 h
- Stimuli
- Empagliflozin 0, 0.1, 1, 10 µM in 25 mM glucose, with 100 nM vs 10 nM insulin to mimic lower insulin
- Response
- De novo lipogenesis, SREBP-1c target genes
6
Metformin · Fat production in the liver
Not nowMetformin and fat storage in C. elegans, controlling for bacteria
Metformin lowers worm fat through AMPK only when the bacteria are alive.
Expected to reduce predictive uncertainty by 1% · $1,500 · 6 days
- Cells
- C. elegans N2 and aak-2(ok524), synchronized L4
- Time
- 72 h from L4
- Stimuli
- Metformin 0, 25, 50 mM; live vs UV-killed OP50
- Response
- Oil Red O fat staining
7
GLP-1 receptor agonists · Fat production in the liver
Not nowDo GLP-1 drugs act on human liver cells directly?
Semaglutide has no effect on fat production in primary human liver cells, confirming the benefit is indirect.
Expected to reduce predictive uncertainty by 0% · $6,000 · 6 days
- Cells
- Primary human hepatocytes, 3 donors
- Time
- 48 h
- Stimuli
- Semaglutide 0, 1, 10, 100 nM in 25 mM glucose + 100 nM insulin; forskolin as a cAMP positive control
- Response
- De novo lipogenesis, cAMP, GLP1R transcript
8
All treatments · Fat export and droplet handling
RecommendDoes any current drug touch fat export?
None of the five patches changes VLDL-triglyceride secretion from human liver cells, in either PNPLA3 genotype, so the export module is genuinely unpatched.
Tests whether this mechanism is an untapped target · $11,000 · 8 days
- Cells
- Primary human hepatocytes from PNPLA3 148I and 148M donors
- Time
- 48 h
- Stimuli
- Each drug at patient-level concentration; lomitapide as an export-blocking control
- Response
- ApoB-100 and VLDL-triglyceride secretion, intracellular triglyceride
Experiment record
Each approved experiment, what the lab returned, and how the model's predictions and validation changed. Saved in this browser only.
No experiments run yet. The top-ranked experiment addresses the model's failed prediction for pioglitazone.
Against the ARPA-H IGoR program metrics
| Program target | Now | ||
|---|---|---|---|
| Knowledge gaps detected algorithmically | ≥ 3 (Phase I) | 13 | Met |
| Mechanistic sub-models (treatments) | ≥ 3 (Phase I), ≥ 10 (Phase II) | 5 | Met |
| Out-of-sample prediction | improves with experiments (Phase II) | 2/3 | Met |
| Lab data returns ingested | ≥ 1 (Phase I), ≥ 10 (Phase II) | 0 | Not yet |
| Model update latency | ≤ 24 h, then ≤ 4 h | no data yet | – |
| Protocols generated automatically | ≥ 3 (Phase I) | 0 | Not yet |
| Experimental modalities | ≥ 3 (Phase II) | 5 | Met |
| Inter-lab concordance | ≥ 85% | run one test at two labs | – |
| Exceptions handled automatically | ≥ 30% (Phase II) | none yet | – |
Evidence
Every source, its study design, and how it is used: as model evidence, a prior, a calibration point, or an out-of-sample test. PubMed records are re-verified before every change is accepted.
| Study design | ||
|---|---|---|
| Armstrong 2016 (LEAN), LancetLiraglutide safety and efficacy in patients with non-alcoholic steatohepatitis (LEAN): a multicentre, double-blind, randomised, placebo-controlled phase 2 study. | Study | out-of-sample test |
| Belfort 2006, N Engl J MedA placebo-controlled trial of pioglitazone in subjects with nonalcoholic steatohepatitis. | Study | out-of-sample test |
| Kuchay 2018 (E-LIFT), Diabetes CareEffect of Empagliflozin on Liver Fat in Patients With Type 2 Diabetes and Nonalcoholic Fatty Liver Disease: A Randomized Controlled Trial (E-LIFT Trial). | Study | out-of-sample test |
| Lavine 2011 (TONIC), JAMAEffect of vitamin E or metformin for treatment of nonalcoholic fatty liver disease in children and adolescents: the TONIC randomized controlled trial. | Study | out-of-sample test |
| Said 2017, Ann HepatolMeta-Analysis of Randomized Controlled Trials of Pharmacologic Agents in Non-alcoholic Steatohepatitis. | Study | out-of-sample test |
| Tang 2015, Diabetes Metab Res RevComparative efficacy of anti-diabetic agents on nonalcoholic fatty liver disease in patients with type 2 diabetes mellitus: a systematic review and meta-analysis of randomized and non-randomized studies. | Study | out-of-sample test |
| Cabreiro 2013, CellMetformin retards aging in C. elegans by altering microbial folate and methionine metabolism. | Genetic perturbation (knockout / knockdown) | context |
| Clar 2009, PLoS OneAdding pioglitazone to insulin containing regimens in type 2 diabetes: systematic review and meta-analysis. | Randomized controlled trial | model evidence, prior, adverse effect |
| DailyMed: Ozempic (semaglutide) labelLabel carries a boxed warning | Dataset | adverse effect |
| DailyMed: pioglitazone tablet labelLabel carries a boxed warning | Dataset | adverse effect |
| Davies 2021 (STEP 2), LancetSemaglutide 2·4 mg once a week in adults with overweight or obesity, and type 2 diabetes (STEP 2): a randomised, double-blind, double-dummy, placebo-controlled, phase 3 trial. | Randomized controlled trial | model evidence, prior, adverse effect |
| Donnelly 2005, J Clin InvestSources of fatty acids stored in liver and secreted via lipoproteins in patients with nonalcoholic fatty liver disease. | Isotope tracer study (human) | model evidence, prior |
| Foretz 2010, J Clin InvestMetformin inhibits hepatic gluconeogenesis in mice independently of the LKB1/AMPK pathway via a decrease in hepatic energy state. | Genetic perturbation (knockout / knockdown) | model evidence |
| Fullerton 2013, Nat MedSingle phosphorylation sites in Acc1 and Acc2 regulate lipid homeostasis and the insulin-sensitizing effects of metformin. | Genetic perturbation (knockout / knockdown) | model evidence, prior |
| Gastaldelli 2009, Am J Physiol Endocrinol MetabDecreased whole body lipolysis as a mechanism of the lipid-lowering effect of pioglitazone in type 2 diabetic patients. | Randomized controlled trial | model evidence, prior |
| GTEx (platform data): GLP1ROnly tissue recorded: pancreatic islets, median 29.4 TPM | Expression atlas | model evidence, prior |
| GTEx (platform data): PNPLA3Liver 20.2 TPM, hepatocytes 18.5 TPM | Expression atlas | model evidence |
| GTEx (platform data): PPARGSubcutaneous fat 63.8 TPM, visceral fat 58.7 TPM | Expression atlas | model evidence |
| GTEx (platform data): SLC5A2 (SGLT2)Kidney cortex 33.8 TPM, testis 15.2 TPM; liver not recorded | Expression atlas | model evidence, prior |
| GTEx (platform data): TM6SF2Small intestine 76.3 TPM, liver 34.4 TPM | Expression atlas | model evidence |
| He & Wondisford 2015, Cell MetabMetformin action: concentrations matter. | Review | model evidence, prior |
| He 2015, MetabolismThiazolidinediones attenuate lipolysis and ameliorate dexamethasone-induced insulin resistance. | Pharmacological exposure | model evidence |
| Kozlitina 2014, Nat GenetExome-wide association study identifies a TM6SF2 variant that confers susceptibility to nonalcoholic fatty liver disease. | Human genetic association, Genetic perturbation (knockout / knockdown) | model evidence |
| Lazo 2010 (Look AHEAD), Diabetes CareEffect of a 12-month intensive lifestyle intervention on hepatic steatosis in adults with type 2 diabetes. | Randomized controlled trial | model evidence, prior, calibration |
| Li 2011, Cell MetabAMPK phosphorylates and inhibits SREBP activity to attenuate hepatic steatosis and atherosclerosis in diet-induced insulin-resistant mice. | Genetic perturbation (knockout / knockdown) | model evidence, prior |
| Liu 2015, J Diabetes ComplicationsEfficacy and safety of sodium-glucose cotransporter 2 inhibitors in type 2 diabetes: a meta-analysis of randomized controlled trials for 1 to 2years. | Randomized controlled trial | model evidence, prior, adverse effect |
| Madiraju 2014, NatureMetformin suppresses gluconeogenesis by inhibiting mitochondrial glycerophosphate dehydrogenase. | Genetic perturbation (knockout / knockdown) | model evidence |
| Onken & Driscoll 2010, PLoS OneMetformin induces a dietary restriction-like state and the oxidative stress response to extend C. elegans Healthspan via AMPK, LKB1, and SKN-1. | Genetic perturbation (knockout / knockdown) | context |
| Owen 2000, Biochem JEvidence that metformin exerts its anti-diabetic effects through inhibition of complex 1 of the mitochondrial respiratory chain. | Pharmacological exposure | model evidence, prior |
| Panjwani 2013, EndocrinologyGLP-1 receptor activation indirectly reduces hepatic lipid accumulation but does not attenuate development of atherosclerosis in diabetic male ApoE(-/-) mice. | Genetic perturbation (knockout / knockdown) | model evidence, prior |
| Pravenec 2008, Int J ObesLong-term pioglitazone treatment enhances lipolysis in rat adipose tissue. | Pharmacological exposure | model evidence |
| Romeo 2008, Nat GenetGenetic variation in PNPLA3 confers susceptibility to nonalcoholic fatty liver disease. | Human genetic association | model evidence |
| Vilar-Gomez 2015, GastroenterologyWeight Loss Through Lifestyle Modification Significantly Reduces Features of Nonalcoholic Steatohepatitis. | Observational cohort | model evidence |
| Zhou 2001, J Clin InvestRole of AMP-activated protein kinase in mechanism of metformin action. | Pharmacological exposure | model evidence, prior |
Version history
- Version 1, 2026-09-28: First model: five treatments over four mechanisms, 27 PubMed-verified papers plus GTEx and DailyMed records; liver-fat trials held out for testing.
Think a link is wrong or a study is missing? Suggest a correction. Every change is re-validated against the held-out trials before it is accepted.