AlphaGenome 2: source-linked RNA and regulatory findings
Research update · 14 November 2026
Eleven research contexts connect existing observations with AlphaGenome 2 (AG2) predictions. They cover human SMN1, SMN2, PLS3, MDM4 and PERP/ARFGEF3, plus a distinct mouse Hspa8 question. Each context preserves the source, exact identity, model output and strongest counterfinding. These are different questions and reused sources; the count is not a count of independent biological confirmations.
Model channels labelled observed refer to channel provenance. The predicted variant response is not a new experimental measurement. Model scores, attribution and splice-support fractions do not establish binding, causality or treatment efficacy.
AlphaGenome 2 (AG2): SMN2 allele and expression-model context
A13 · A43 · Source-linked research context
Existing natural-allele cohort associations and AlphaGenome 2 GeneMaskLFCScorer outputs are shown as distinct evidence. The model reports small DIFF_MEAN values across a shared chr5 interval; these are predictions, not measured RNA changes.
Source and interpretation
The Ruhno allele set provides allele identities and cohort associations, not isolated allele interventions. In the model output, observed describes channel provenance; it does not mean the variant effect was experimentally observed. A gene-first follow-up asks which sequence contributes to the expression objective; attribution is not the fraction of expression explained.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
GeneMaskLFCScorer DIFF_MEAN, common interval GRCh38 chr5:[69551712,70600288) · 4 rows
Allele
C1 adult tissue UBERON:0006469
In-vitro motor neuron CL:0000100
Channel status
A-549G; GRCh38 chr5:70075972 A>G
0.0017555952
Not reported
motor-neuron value not recoverably saved
A-44G; GRCh38 chr5:70076477 A>G
0.0029186606
0.0025262833
C1 observed; motor-neuron imputed_supported
C-1897T; GRCh38 chr5:70074624 C>T
0.00044804811
0.0006043911
C1 observed; motor-neuron imputed_supported
C-478T; GRCh38 chr5:70076043 C>T
-0.0001321435
-0.00032401085
C1 observed; motor-neuron imputed_supported
Interpretation limits
A44G has an AG1/AG2 sign disagreement; the AG1 serving equation and a common effect scale are unresolved.
A549G in CL:0000100 is missing and was not recovered; values for that channel in other alleles are imputed_supported.
C1 is heterogeneous adult tissue, not isolated SMA motor neurons. Phase, causal effect, training overlap, and gene-mask annotation equivalence remain unresolved.
Gao HEK293 shortened-minigene PSI measurements differ by substitution: C>A increased inclusion while C>G reduced it. Native-genome AG2 outputs have mixed signal and fraction behavior and do not reproduce a measured PSI endpoint. Separate c.840T>C and c.859G>C positive-source comparisons and A-44G are included with their counterfindings.
Comparison boundary
The source assay is a HEK293 minigene. The AG2 prediction uses native genomic context and predicted junction outputs. Cell, construct, and endpoint differ, so the values are not a direct validation comparison. A source-matched measured endpoint is the next useful evidence.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Additional source-positive comparisons
The same C1 spinal-cord configuration gives the expected inclusion-support direction for c.840T>C, but the opposite direction for source-positive c.859G>C. A-44G has no matched isolated-allele negative-control assay here. This does not establish calibrated splice-effect prediction. C1 spinal tissue and the HEK293 minigene are different contexts.
The native substitutions are not authenticated as identical to the complete shortened minigene constructs.
For both substitutions, donor and acceptor fractions increase while raw junction changes are mixed; competing-junction decline, especially skip J68, drives the fraction arithmetic.
These model outputs are not measured PSI, rescue, or causal mechanism evidence; the diagnostic decomposition is post hoc on the same responses.
AlphaGenome 2 (AG2): PLS3 rs73226811: eQTL and RNA-model context
A44 · Source-linked research context
The exact GRCh38 chrX:116225577 C>A allele is rs73226811. A GTEx v8 Testis conditional eQTL reports NES +0.0769585; AG2 DIFF_MEAN is negative in Testis and slightly positive in an imputed_supported motor-neuron channel.
Identity and context
PLS3-AS1 is a distinct gene and is not combined with PLS3. The gene-centered model window does not attribute this distal allele; an exact linked distal source or a separately documented mechanism would be needed.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Exact allele source and model values · 3 rows
Context
Value
Interpretation
GTEx v8 Testis conditional eQTL
NES +0.0769585
Association; rank 2; tssDistance 664403
AG2 Testis UBERON:0000473
DIFF_MEAN -0.00096169114
Observed channel provenance
AG2 in-vitro motor neuron CL:0000100
DIFF_MEAN +0.00008480996
Imputed_supported channel
Interpretation limits
The eQTL is an association that may reflect linkage disequilibrium or conditioning, not an isolated-allele causal effect.
Fine-mapping returned zero records in the queried endpoint; this does not establish that fine-mapping evidence is absent.
Five assay labels are correlated outputs, not independent replications; possible GTEx training overlap remains unresolved. No motor-neuron effect or SMA protection is established.
Two reference RNA-expression attribution maps identify a shared leading TSS/first-exon tile. The terminal-exon tile ranks second in C1 spinal tissue and third in the in-vitro motor-neuron context; an internal-exon tile ranks second in the latter. These are sequence-attribution maps, not measured regulatory effects.
Next source question
Separate exonic from non-exonic attribution in the two specified follow-up regions and reuse existing CAGE, PRO-cap, and accessible-chromatin evidence by coordinate before considering any additional inference.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
GRCh38 half-open 500-bp attribution regions · 4 rows
Context
Rank
Region
Annotated overlap
Both objectives
1
chr5:[70049500,70050000)
TSS and 98 bp of first exon
C1 spinal tissue
2
chr5:[70077000,70077500)
482 bp of terminal exon
In-vitro motor neurons
3
chr5:[70077000,70077500)
482 bp of terminal exon
In-vitro motor neurons
2
chr5:[70067000,70067500)
287 bp of internal exons
Interpretation limits
High attribution overlaps RNA objective exons and does not identify an enhancer.
Attribution shares are not fractions of expression explained; no null or functional perturbation is available.
AlphaGenome 2 (AG2): SMN2 regulatory-coordinate and chromatin context
A43 · Source-linked research context
Three fixed regulatory intervals were cross-checked against six distinct donor ATAC experiments, comprising three healthy and three ALS donors. No selected conservative peak overlaps were found; this is a bounded source-selection result, not evidence of biological inactivity.
Interpretation
The coordinate overlap check does not test whether the regions regulate SMN2 in relevant cells. It only reports the selected interval and peak comparison.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Fixed intervals and selected ATAC comparison · 1 rows
Source set
Donors
Selected conservative peak overlap
Scope
Six donor ATAC experiments
3 healthy; 3 ALS
0 at all three fixed intervals
No SMA donors; not an inactivity result
Interpretation limits
The experiments contain no SMA donors and do not establish SMN2 target-gene causality.
A shared exclusion mask overlaps none of the three intervals, so this specific mask does not explain the zero selected overlaps; broader activity and other sources remain untested.
AlphaGenome 2 (AG2): MDM4 reference splice topology and sample-source boundary
A13 · Source-linked research context
Reference splice-event summaries, local junction edges, strict read paths, and six named BAM paths describe different evidence units. Their coordinate topology can be compared, but they do not establish a selective MDM4 isoform switch in the SMA cell line.
Counterfinding
The same named edges can have event-specific counts. SE inclusion counts are not the sum of two edge columns, and A3SS inclusion evidence spans a short boundary. Strict read-name groups are separately filtered local assignments, not validated phased transcripts.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Relationship to the existing human RNA observations
The source cohort contains 15 deposited libraries from one GM03813 fibroblast line. Local 68-nt, 39-nt and skip paths are separate alignment evidence. A processing log naming six BAM paths does not authenticate their input contents or join AG2 reference edges to sample-level treatment effects. Original event membership and denominators remain unresolved.
Distinct evidence units retained · 4 rows
Evidence unit
What it supports
What it does not establish
Reference splice event
Event-level inclusion or choice
Local read paths or isoform fractions
Reference junction edge
Coordinate topology
Phased transcript chain
Strict joined read path
Filtered local assignment
Historical run membership or sample effect
Six BAM path names
Names recorded in processing log
Verified sample-level quantitative join
Interpretation limits
Generic primary dermal fibroblasts and static reference DNA do not match GM03813 SMA or treatment.
Individual edges are not phased chains. SE event counts and A3SS choices cannot be converted directly into three isoform fractions.
The exact run build, processing options, input hashes, and event membership for the six BAM paths remain unbound; do not join the separately named paths to sample effects.
AlphaGenome 2 (AG2): PLS3 reference annotation and gene-first model context
A44 · Source-linked research context
PLS3 reference annotations differ by Ensembl release. A gene-first splice scorer returned four alternative edges as not_returned and a positive canonical edge score; neither result establishes absence of alternative splicing or a measured expression change.
Existing RNA evidence
Four verified transcript-quantification tables from one ENCODE motor-neuron experiment include PLS3.17 and canonical-stable ENST00000355899.7. Exact coordinate-bound junction counts are the next relevant measured source; keep assay and phasing explicit.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Distinct PLS3 reference regions — GRCh38 half-open · 2 rows
Context
Region
Selection rank
Boundary
C1 spinal tissue; GTEx polyA RNA
chrX:[115603000,115603500)
1
500 bp; 90 bp exon overlap in release 116
In-vitro motor neuron; ENCODE total RNA
chrX:[115563500,115564000)
2
500 bp; rank 1 lies in excluded control window;139 bp exon overlap in release116
Reference annotation and model output · 3 rows
Measure
Release/result
Interpretation
Exonic overlap
Release 112: 0 bp in both selected regions; release 116: 90 bp in C1 and 139 bp in motor neurons
Reference annotation only
Alternative splice edges
4 not_returned
Not an absence or zero result
Canonical edge
1.958001852
Model score; not measured PSI
Interpretation limits
Release 112 and release 116 annotations differ; exonic overlap cannot be treated as a biological change.
not_returned is not a zero score or absence claim. No exact junction counts were supplied in the inspected experiment files.
Transcript quantification tables represent two biological replicate labels and two output methods, not four independent samples; newer alternative stable IDs are absent from the target catalogs.
AlphaGenome 2 (AG2): PERP and ARFGEF3: neuronal regulatory-source context
A21 · Source-linked research context
The fixed PERP-linked element EH38E2507697 has a verified neuronal assay connection, but target-gene causality remains open. A nearby shared 450 bp element has multiple target IDs, including PERP and ARFGEF3; spatial overlap and model attribution do not establish exclusive regulation.
Strongest counterfinding
At EH38E2507697, the reported neuronal source values are ATAC z=1.34, H3K4me3 z=-0.12, H3K27ac z=-0.56, with CTCF missing. The source establishes a neuronal measurement connection, not an active enhancer or PERP functional link. The adjacent ARFGEF3 evidence also remains non-specific.
Next source question
Identify exact biosamples and files for the element and obtain measured neuronal PERP transcript or 3D linkage with target identity. Preserve the PERP.12 versus current .13 annotation distinction.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
A separate 450-bp source-selected element
At GRCh38 chr6:[138115850,138116300), the EPIraction source contains eight target-gene IDs including PERP and ARFGEF3. This is a different region from EH38E2507697. A separately requested ARFGEF3 RNA attribution map has about 2.98-fold mean absolute attribution versus each immediate same-width flank; the element contributes 0.0978818% of the absolute attribution in that returned window. Different gene-centered windows prevent a cross-gene effect comparison. An exact allele or target-resolved functional observation is needed to distinguish target hypotheses.
Context comparability
The additional dermal-fibroblast RNA attribution used a minus-strand track, whereas the motor-neuron RNA track was unstranded. Their numerical difference is not an isolated cell-type effect. These DNA/RNA regulatory comparisons do not test the separate PERP–DAG1 protein-interaction hypothesis.
Bounded source and model findings · 4 rows
Finding
Result
Boundary
SCREEN neuronal element
ATAC z 1.34; H3K4me3 z -0.12; H3K27ac z -0.56; CTCF missing
Ancillary collection; no target-link result
ATAC cCRE/flank ratio
5.686308134
Model consistency, not independent validation
H3K27ac cCRE/flank ratio
0.867984887
No local elevation in this comparison
Shared element
450 bp; eight target IDs
No exclusive PERP or ARFGEF3 assignment
Separate source-selected ARFGEF3-map question · 1 rows
Fixed region, GRCh38 half-open
Local mean-absolute-attribution ratios
Own-window absolute-attribution share
Source target identities
chr6:[138115850,138116300)
2.98160 versus left flank;2.98040 versus right flank
0.0978818%
8 target-gene IDs; not an exclusive PERP enhancer
Interpretation limits
Human SMA/control PERP ranges overlap; donor mapping and processing remain unresolved. Separate NGN2 comparison is adjusted P=.929.
The motor-neuron assay source includes one ALS donor; iNPC/iPSC-derived motor-neuron biosamples differ. CTCF is missing and no displayed target-link result confirms PERP causality.
No selected thresholded peak overlaps the original cCRE/tile/fixed ±5 kb regions; this is not evidence of zero activity. The shared element is not target-specific.
For the mouse Hspa8 G470R allele, one paired broad-brain prediction returned small raw-score ratios for fixed upstream and downstream junctions. A separate diagnostic skip ratio is larger but uses a low reference score. This does not establish a cis-RNA effect.
Context separation
The source mouse allele is mm10 chr9:40804249 G>C, p.Gly470Arg. NM_031165.5:c.1408G>C is our current annotation crosswalk, not an original-paper transcript version. The prediction uses broad mouse brain UBERON:0000955, ENCODE polyA RNA; this is not an SMA motor-neuron or developmental-stage model. The known protein-level and SMN2-trans observations are separate from this local Hspa8 cis-RNA question.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
AlphaGenome 2 (AG2): Reported PLS3 c.367+182C>T blood association
A44 · Source-linked research context
A reported association links PLS3 c.367+182C>T to blood expression in a 50-person SMA and 21-control cohort. The genomic coordinate, transcript identity, carrier count, and original supplement were not established, so no AG2 variant call was made.
Next source question
Recover the original supplement and bind the exact transcript and genomic allele before considering any variant-level model comparison. Keep this candidate separate from rs73226811.
Reported association and strongest counterfinding
The source reports p=0.04 and a carrier median of 19.1 copies/ng cDNA. The pooled 10.8 value is inconsistently labelled mean or median and is not an authenticated noncarrier comparator. Carrier count is not bound. Family severity directions differ, and higher expression does not establish consistent clinical benefit.
Reported cohort context · 4 rows
Item
Reported information
Status
Cohort
50 SMA; 21 controls
Source report
Variant
PLS3 c.367+182C>T
Genomic identity pending
Carrier count
Not bound
Missing
AG2 result
None
No call made
Interpretation limits
The exact variant identity is unresolved; no genomic coordinate or transcript is inferred.
The reported blood association has no bound carrier count and no consistent clinical benefit.
Venous-blood reference RNA differs from full-length ddPCR; a literal blood RNA label was imputed. This is not a motor-neuron result.
AlphaGenome 2 (AG2): SMN1 c.863G>T splice-source and model context
A13 · Source-linked research context
SMN1 NM_000344.3 c.863G>T maps to GRCh38 chr5:70951969 G>T. The shortened HEK293 minigene showed exon 7 inclusion decreasing from 91.7% to 14.8%; native blood and fibroblast observations are separate materials. AG2 fraction changes are tiny and negative while raw J78 rises in all three imputed channels.
Bounded interpretation
This is a source/model comparison, not reproduction of the strong experimental splice effect. The finite reference-transfer question is closed; the useful next evidence would be source-bound human native-RNA allele/junction data with material and denominator specified.
Model and output provenance
These are API predictions made with the AlphaGenome 2 client. ALL_FOLDS is the requested model selector; an immutable served weight revision was not independently supplied. Channel labels such as observed or imputed describe output provenance, not an experiment on the variant. Model outputs and source measurements retain their own units and denominators.
Source and model comparison · 4 rows
Evidence
Result
Context
HEK293 shortened minigene
Exon 7 inclusion 91.7% to 14.8%
Experimental construct
AG2 inclusion-support fractions
Negative changes across 3 channels; 0.0055–0.0173 percentage points
Coupled imputed channels
AG2 raw J78
Increases in all 3 channels
Raw junction score, not PSI
Native RNA
Blood from 2 patients; fibroblast from 1 patient
Distinct source materials; no matched denominator
HEK293 imputed model channels — ALT-minus-REF fractions expressed in percentage points · 3 rows
The model fraction changes are only 0.0055–0.0173 percentage points and are not PSI; rising raw J78 is compatible with falling J78 fraction because the denominator also rises.
Native RNA observations are not a matched model denominator and no quantitative conversion is established.
SMN1 and SMN2 remain distinct paralogs; do not infer SMN2 behavior from this result.