Grounded Light-Field Cell Painting of Genetic Variants in Human Organoids

When a patient's genome is sequenced, a rare misspelling in a disease gene is more often than not uninterpretable: over 90% of missense variants across roughly 4,000 disease-associated human genes are classified as variants of uncertain significance (VUS), meaning clinicians cannot tell whether the mutation causes disease A scalable approach to resolving variants of uncertain significance. The functional assays used to resolve them run in baker's yeast or flat immortalized cells, which lack the polarized, multicellular architecture of human tissue Constructing and interpreting a large-scale variant effect map for an ultrarare disease gene: Comprehensive prediction of the functional impact of PSAT1 genotypes; Cell Painting predicts impact of lung cancer variants. The obvious fix — human organoids, miniature self-organized tissues grown from patient-derived cells — stalls at the microscope: the standard way to see inside a thick 3D organoid is a confocal Z-stack, which is slow, phototoxic, and, as one audit showed, leaves deep networks trained on it attending mostly to out-of-focus blur rather than to cells Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images.

This proposal builds a single pipeline that closes all three gaps at once. A light-field microscope — an instrument that records not just how much light arrives at each sensor pixel but from which direction, so an entire 3D volume is captured in one snapshot, with no mechanical scanning — images Cell-Painting-stained human organoids well by well. The captured light field is reconstructed into a 3D volume by explicit Gaussian splatting extended with a learnable model of defocus blur, so that out-of-focus haze is separated out physically rather than leaking into the features. A 3D feature extractor is then trained with a penalty that forces its attention onto reconstructed cell interiors, audited with the Grad-CAMO grounding metric, and its per-gene calibrated morphological readout is used to reclassify VUS whose effects depend on human tissue context — the class of variants that yeast and flat-cell assays systematically miss.

Core Claim

The claim has three parts. First, that Fourier light-field microscopy (FLFM) — which splits the microscope pupil instead of the image plane, avoiding the sampling non-uniformities of microlens-based light-field designs — combined with 3D Gaussian Adaptive Tomography (3DGAT) can reconstruct subcellular morphology in dense human organoids at screening throughput, where Richardson-Lucy deconvolution, the current standard, fails on resolution and axially stretches structures 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy. Second — and this is the genuinely new ingredient — that grafting a learnable Circle-of-Confusion blur model, as demonstrated for Gaussian splatting from defocused photographs CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images, into the organoid reconstruction decomposes each snapshot into sharp in-focus geometry plus an explicit depth-dependent defocus residual, and that training the downstream variant-effect classifier only on the grounded sharp component eliminates the "background cheating" documented for 3D Cell Painting networks Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images. Third, that the resulting assay resolves a measurable set of VUS that 2D morphological profiling in flat carcinoma lines leaves unclassified or misclassifies, because those variants' mechanisms require human tissue architecture Cell Painting predicts impact of lung cancer variants; A scalable approach to resolving variants of uncertain significance.

The closest existing work each covers one edge of this triangle. 3DGAT reconstructed transparent zebrafish larvae but never human organoids, genetic perturbations, or multi-channel phenomics 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy. Snapshot light-field imaging was demonstrated at 100 frames per second in acute mouse brain slices, but with wild-type reporters and rodent tissue, and it documented a painful trade: ray-optics refocusing blurred axially (median spread 47.5 µm) while iterative deconvolution halved that spread (22.5 µm) only by dropping the temporal signal-to-noise ratio from 1.0 to 0.6 Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators. The cmVIP pipeline showed Cell Painting can read out variant impact (gain-, loss-, or change-of-function) but only in flat A549 lung carcinoma monolayers Cell Painting predicts impact of lung cancer variants, and Grad-CAMO showed 3D networks on confocal stacks reach 85% accuracy while only 30% of their attention maps overlap actual cell boundaries — 70% latch onto out-of-focus background or neighboring cells Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images. No prior work puts snapshot volumetric optics, defocus physics, and variant-effect readouts in the same human 3D substrate.

Research Program

Three components are built in sequence. The first is the optical and reconstruction core. A FLFM instrument images entire organoids (a few hundred micrometers across) in a single exposure per fluorescence channel, which matters because sequential mechanical Z-scanning is exactly what makes confocal screening of multi-well organoid plates slow and phototoxic. The reconstruction extends 3DGAT — explicit 3D Gaussian ellipsoids optimized with a Fourier-domain loss and tile-based intensity voxelization 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy — with a learnable Circle-of-Confusion model in the spirit of CoCoGaussian, which parameterized defocus blur inside a splatting objective with a learnable aperture parameter $K$ and scale parameter $\beta$ and reached 30.14 dB PSNR on the DoF-NeRF benchmark, ahead of the prior best 29.87 dB CoCoGaussian: Leveraging Circle of Confusion for Gaussian Splatting from Defocused Images. The core move is that the Circle of Confusion radius for a structure at depth $z$ relative to a rendered focal plane $z_f$ is modeled as

$$r(z) \;=\; K\,\beta\,|z - z_f|,$$

with $K$ and $\beta$ learned per optical channel, so that light is explicitly split into an in-focus Gaussian component and a defocus residual whose size is predicted by optics rather than treated as an unexplained nuisance. The composite training objective is

$$\mathcal{L} \;=\; \mathcal{L}_{\mathrm{FDL}} \;+\; \lambda\,\mathcal{L}_{\mathrm{CoC}} \;+\; \mu\,\mathcal{L}_{\mathrm{ground}},$$

where $\mathcal{L}_{\mathrm{FDL}}$ is 3DGAT's Fourier-domain reconstruction loss, $\mathcal{L}_{\mathrm{CoC}}$ penalizes mismatch between the measured blur of each reconstructed structure and the blur predicted by $r(z)$, and $\mathcal{L}_{\mathrm{ground}}$ penalizes variant-classifier attention mass that falls outside reconstructed cell volumes (the same criterion the Grad-CAMO audit measures at evaluation time). $\lambda$ and $\mu$ are scalar weights set on a held-out calibration plate.

The second component is the grounded feature learner. A 3D convolutional network (3D EfficientNet-B0, the backbone used in both cmVIP and Grad-CAMO, so comparisons are apples-to-apples) consumes the sharp in-focus volume and produces a per-cell morphological profile, with the grounding penalty active during training rather than only audited afterward. The audit itself reuses the Grad-CAMO criterion — fraction of attention maps with significant overlap with segmented cell boundaries — so we inherit a directly comparable failure statistic: 30% grounded attention at 85% dose-prediction accuracy for the unpenalized baseline Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images.

The third component is the variant campaign, and here the cross-species literature is the design constraint, not just motivation. The meta-analysis of acute cerebral ischemia animal models showed the translational failure is one of predictive validity, not statistical power Treatment of acute cerebral ischemia using animal models: a meta-analysis; rodent–human divergence in circadian phase is severe enough that mouse liver organoids fail to predict human chronotoxicity Chronotoxici‐Plate Containing Droplet‐Engineered Rhythmic Liver Organoids for Drug Toxicity Evaluation, and human and mouse pancreatic islets even run their cellular clocks with different phase relationships Pancreatic α- and β-cellular clocks have distinct molecular properties and impact on islet hormone secretion and gene expression. The primary biological substrate is therefore human iPSC-derived liver organoids (the droplet-engineered rhythmic format is a proven, scalable platform), with endometrial organoids — established precisely because flat culture cannot model human implantation biology Establishment of a Novel Human Endometrial Organoid — as the second system. The variant panel is drawn from three sources with known ground-truth properties: the 20 well-characterized control alleles from cmVIP Cell Painting predicts impact of lung cancer variants; VUS left unresolved or discordant across assays in the ExCALIBR study, which definitively reclassified 75% (12,135 of 16,115) of clinical VUS in 40 genes but left the remaining 25% open, exactly because flat 2D assays miss tissue-context effects A scalable approach to resolving variants of uncertain significance; and community assay data from MaveDB, which holds over seven million variant-effect measurements MaveDB 2024: a curated community database with over seven million variant effects from multiplexed functional assays. Calibration is per gene, following the demonstration that genome-wide calibration of predictors causes substantial clinical misassignment while gene-specific empirical calibration against experimental data achieves under 1% discordance with ClinVar A scalable approach to resolving variants of uncertain significance; Guidelines for releasing a variant effect predictor.

Proposed Experiments

Experiment 1 tests the reconstruction core on physical phantoms before any biology. Fluorescent line-pair targets embedded in agarose of increasing turbidity (to mimic the known failure of light-field methods in scattering media beyond roughly 100–200 µm) establish resolution: the head-to-head numbers to beat are Richardson-Lucy failing to resolve lines spaced 0.96 µm apart while 3DGAT resolved 0.72 µm, and 28.3 dB versus 39.3 dB PSNR on a 3D target 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy. Success means the CoC-extended reconstruction matches or beats 3DGAT's resolution and PSNR in clear phantoms, and — the new result — degrades gracefully in scattering phantoms where plain 3DGAT's ray alignment assumptions break. No reported result found on Lacuna for light-field reconstruction fidelity inside human organoid-density scattering media, so this measurement itself is a contribution.

Experiment 2 tests grounding. Organoids stained with the Cell Painting palette are imaged once by FLFM snapshot and once by a conventional confocal Z-stack (subsampled wells only, since the confocal route is the slow comparator). A 3D EfficientNet-B0 trained without the grounding penalty on confocal stacks should reproduce the published failure mode — roughly 30% grounded attention at comparable accuracy Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images. The decisive comparison is the same architecture on our light-field volumes with and without the CoC residual and grounding penalty: the claim is supported only if grounded attention at least doubles (from 30% toward the majority of profiles) at matched or better classification accuracy.

Experiment 3 tests variant-effect sensitivity. The cmVIP control-allele protocol is replicated in the organoid setting: all 20 well-characterized alleles must be correctly assigned to gain-of-function, loss-of-function, change-of-function, or neutral — matching the 20-for-20 result in A549 cells Cell Painting predicts impact of lung cancer variants. Concordance is then measured against an orthogonal transcriptional assay on the same variant library, with cmVIP's Spearman correlation of 0.69 with transcriptional profiling and 76.8% concordance as the reference numbers, and its 83.2% rate of variants yielding a reproducible morphological signal as the sensitivity bar Cell Painting predicts impact of lung cancer variants.

Experiment 4 is the payoff measurement: incremental VUS resolution. The same variant library is profiled twice — in the organoid light-field pipeline and in the flat 2D Cell Painting format — under matched staining, sequencing depth, and classifier tuning budget. The primary endpoint is the set of variants that the organoid assay resolves to a confident call where the 2D assay yields no reproducible signal or a discordant one, with clinical validity checked against ClinVar on the subset with existing definitive labels and a target discordance rate below the under-1% benchmark set by ExCALIBR A scalable approach to resolving variants of uncertain significance. Variants that flip call direction between yeast-style complementation and organoid morphology are the most informative, since they directly instantiate the species-context failure: yeast complementation missed nearly 2,000 variants' worth of mammalian tissue context for PSAT1 Constructing and interpreting a large-scale variant effect map for an ultrarare disease gene: Comprehensive prediction of the functional impact of PSAT1 genotypes, and structural variants that trigger proteasomal degradation or rewire protein interactions are exactly the class static conservation scores miss Nielsen et al.; Zou et al..

Experiment 5 measures throughput and phototoxicity, the operational reason to prefer snapshot capture. The reference point is the demonstrated 100 frames-per-second snapshot volumetric imaging in light-field microscopy Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators; no reported result found on Lacuna for organoid wells-per-day throughput under confocal versus light-field capture, so we measure wells per day, photons delivered per well, and post-screen viability directly.

Baselines

The naive reference point is doing nothing: leaving the variants as VUS. That is the de facto standard of care, with over 90% of missense variants in roughly 4,000 disease genes unclassified A scalable approach to resolving variants of uncertain significance. Any pipeline in this proposal counts as useful only if it produces additional definitive calls that survive the under-1% ClinVar-discordance check; otherwise it has added cost without clinical content.

The strongest competitor overall is the cmVIP 2D Cell Painting pipeline — six fluorescent stains across eight cellular compartments in flat A549 cells, features from a pre-trained EfficientNet-B0, which correctly classified all 20 well-characterized control alleles, found a reproducible morphological signal for 83.2% of tested variants, and reached a Spearman correlation of 0.69 with an independent transcriptional assay at 76.8% concordance Cell Painting predicts impact of lung cancer variants. We will run this pipeline ourselves on the identical variant library in the flat-cell format, with the same staining palette, the same classifier capacity and tuning budget, and the same control alleles, so that any difference is attributable to substrate and optics rather than effort. Beating it means matching its 20-for-20 control-allele performance while resolving a strictly larger set of library variants — specifically, variants for which the 2D run yields no reproducible signal. A complementary flat-cell baseline is the U-2 OS ORF profiling format, which detected reproducible morphological signatures for 50% of tested constructs via 1,300 CellProfiler features Systematic morphological profiling of human gene and allele function via Cell Painting.

Within the light-field track, the strongest existing reconstruction baseline is Richardson-Lucy deconvolution, the field standard: on Fourier light-field data it failed to resolve line pairs spaced 0.96 µm apart (where 3DGAT resolved 0.72 µm) and scored 28.3 dB PSNR against 3DGAT's 39.3 dB, while in voltage imaging it halved axial spread only at the cost of dropping temporal SNR from 1.0 to 0.6 over 21 iterations 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy; Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators. We run Richardson-Lucy, plain 3DGAT, and our CoC-extended 3DGAT on the same raw snapshots with matched compute, judged on the phantom metrics of Experiment 1 and on downstream variant-classifier accuracy.

Finally, two ablations isolate the proposal's key ingredient. The first removes only the CoC residual and grounding penalty, keeping the identical FLFM capture and 3DGAT reconstruction; if the grounding statistic then collapses toward the published 30% Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images, the physics-grounded training signal — not the fancy optics alone — is what carries the result. The second replaces the light-field snapshot with a confocal Z-stack of the same organoids, keeping the classifier identical, to separate the value of snapshot capture from the value of the defocus model.

Why not just…?

Why not just keep profiling variants in 2D Cell Painting? This is the most serious objection, and we treat it as the primary endpoint rather than arguing around it. The 2D route is proven and cheap: cmVIP's numbers are strong, and even the large-scale ExCALIBR effort resolved 75% of clinical VUS in its 40 genes using flat 2D assays A scalable approach to resolving variants of uncertain significance. But the same papers concede the ceiling: cmVIP's authors state that variants depending on differentiated state, tissue polarity, or multicellular signaling fail to manifest in A549 monolayers Cell Painting predicts impact of lung cancer variants, and ExCALIBR left 25% of VUS unresolved with cross-assay discordances attributed to non-physiological contexts A scalable approach to resolving variants of uncertain significance. Experiment 4 settles the question empirically: if the organoid pipeline does not resolve a measurable set of variants that the matched 2D run misses, the proposal's premise is wrong and the 2D pipeline should win on cost.

Why not just use confocal Z-stacks in the organoids? Confocal microscopy is the standard tool for organoid imaging, and for a handful of static endpoints it will work fine — this is a real threat, not a straw man. The evidence against it at screening scale is twofold: sequential plane-by-plane exposure is slow and phototoxic across multi-well, multi-channel campaigns (the operational gap that motivated light-field imaging at 100 frames per second Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators), and, more damaging, the Grad-CAMO audit found that 3D networks trained on confocal stacks achieved 85% accuracy while 70% of their attention latched onto out-of-focus background or neighboring cells — meaning high accuracy can coexist with biologically hollow features Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images. The confocal ablation in the baselines addresses this directly: if the confocal-plus-grounding-penalty variant matches the light-field pipeline on grounding, throughput, and variant resolution, then the defocus physics is unnecessary and a software fix to standard confocal screening suffices — a cheaper outcome we would happily report.

Why not just rely on in silico predictors like AlphaMissense or MetaRNN? Predictors are free and instantaneous, and we use them — for panel triage, following the preclassification practice that scored 62% of over 90,000 unobserved variants A scalable approach to resolving variants of uncertain significance. But the same study showed that definitive clinical classification strictly requires empirical assay data, that genome-wide calibration of predictors produces substantial misassignment, and that gene-specific experimental calibration is what drove discordance below 1% A scalable approach to resolving variants of uncertain significance. Sequence-based predictors also cannot see mechanisms that live outside the sequence: partial unfolding that triggers proteasomal degradation, disrupted post-translational modification sites, or cell-context-specific interaction rewiring Nielsen et al.; Zou et al.. The proposal's readout is designed to capture precisely those mechanisms, which is why the predictor is a triage tool here and not the endpoint.

Why not just use Richardson-Lucy deconvolution on the light-field data? This one is largely settled by existing head-to-head numbers rather than by argument: on Fourier light-field data, Richardson-Lucy failed at 0.96 µm line spacing where 3DGAT resolved 0.72 µm, scored 28.3 dB versus 39.3 dB PSNR, and axially stretched spherical objects 3D Gaussian Adaptive Reconstruction for Fourier Light-Field Microscopy — and axial stretching directly corrupts the volumetric shape features a morphological assay depends on. In dynamic imaging it also dropped SNR from 1.0 to 0.6 over 21 iterations Subcellular resolution three-dimensional light-field imaging with genetically encoded voltage indicators. We still run it as a baseline under matched compute, because it remains the community default and the fair comparison requires it.

Decisive Checks

The proposal stands or falls on three numbers, all pre-registered. Grounding: the CoC-and-penalty pipeline must roughly double the 30% grounded-attention rate of unpenalized 3D Cell Painting networks at matched accuracy Grad-CAMO: Learning Interpretable Single-Cell Morphological Profiles from 3D Cell Painting Images; if grounding does not move, the physics ingredient is decorative. Control alleles: the organoid pipeline must match the 20-for-20 control-allele classification of cmVIP Cell Painting predicts impact of lung cancer variants; failure here means the organoid assay is noisier than the flat one it replaces. Incremental resolution: it must resolve a non-trivial set of library variants that the matched 2D run does not, at under-1% discordance with ClinVar on the labelable subset A scalable approach to resolving variants of uncertain significance; if the resolved set is empty, the honest conclusion is that flat-cell assays already saturate morphological readouts for these genes, and the cross-species context argument — strong as the organoid and chronotoxicity literature makes it Chronotoxici‐Plate Containing Droplet‐Engineered Rhythmic Liver Organoids for Drug Toxicity Evaluation; Treatment of acute cerebral ischemia using animal models: a meta-analysis — does not bite for this assay class, at least not through morphology alone.