Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals

Ihor Kendiukhov

cs.LGMar 10, 2026

Discovery of a Hematopoietic Manifold in scGPT Yields a Method for Extracting Performant Algorithms from Biological Foundation Model Internals

Authors: Ihor Kendiukhov

Organizations: Department of Computer Science, University of Tübingen, Tübingen, Germany

Abstract

We report the discovery and extraction of a compact hematopoietic algorithm from the single-cell foundation model scGPT, to our knowledge the first biologically useful, competitive algorithm extracted from a foundation model via mechanistic interpretability. We show that scGPT internally encodes a compact hematopoietic manifold with significant developmental branch structure, validated on a strict non-overlap Tabula Sapiens external panel and confirmed via frozen-head zero-shot transfer to an independent multi-donor immune panel. To isolate this geometry, we introduce a general three-stage extraction method consisting of direct operator export from frozen attention weights, a lightweight learned adaptor, and a task-specific readout, producing a standalone algorithm without target-dataset retraining. In 88-split donor-holdout benchmarks against scVI, Palantir, DPT, CellTypist, PCA, and raw-expression baselines, the extracted algorithm achieves the strongest pseudotime-depth ordering and leads on key subtype endpoints (CD4/CD8 AUROC 0.867, mono/macro AUROC 0.951). Compared to standard probing of frozen scGPT embeddings with a 3-layer MLP, the extracted head is BH-significantly better on 6/8 classification endpoints while completing a full 12-split evaluation campaign 34.5x faster with approximately 1000x fewer trainable parameters. The exported operator compresses from three pooled attention heads to a single head without statistically significant loss, and further to a rank-64 surrogate. Mechanistic interpretability of the compact operator reveals a concentrated four-factor core explaining 66.2% of ablation impact, with factors resolving into explicit T/lymphoid, B/plasma, granulocytic, and monocyte/macrophage gene programs. A supplementary second-manifold validation (intercellular communication geometry) confirms that the extraction method generalizes beyond hematopoiesis.

Figures & tables

Figure 1: Hematopoietic manifold discovered in scGPT attention geometry (internal panel). Distinct branch structure emerges around stem/progenitor, erythroid, granulocytic, monocyte/macrophage, and lymphoid/T-cell regions.

Panel	Trust.	Random	Donor	Branch
Internal (H65, $d{=}10$ )	0.979	0.495	0.496	0.382
External lung ( $d{=}6$ )	0.617	0.236	0.687	0.058
External Tabula non-overlap ( $d{=}10$ )	0.985	0.594	0.596	0.453

Table 1: LET holdout robustness under shared gates. Trust. = trustworthiness; Random/Donor/Branch = Spearman correlation between predicted and target distances on random, donor-held-out, and branch-held-out test splits, respectively.

Panel	Anchors	Donors	Stages	Corr	Corr resid	Trustworthiness	Blocked $p$
External multi-donor zero-shot	165	7	7	0.500	0.476	0.993	0.0005

Table 2: Frozen-head zero-shot transfer on a separate external multi-donor panel.

Figure 2: Holdout transfer comparison. The strict non-overlap external panel passes all gates, while the lung panel fails robust transfer despite strong global association.

Method	Branch bal. acc.	Branch macro-F1	Stage bal. acc.	Stage macro-F1	CD4/CD8 AUROC	Mono/Macro AUROC	Pseudotime Spearman
H65 anchor head (10D)	0.750	0.557	0.465	0.184	0.803	0.934	0.093
H65 cell head (10D)	0.828	0.584	0.552	0.216	0.867	0.951	0.249
H65 hybrid conservative (10D)	0.815	0.599	0.542	0.217	0.858	0.949	0.132
Raw baseline (1200 log1p)	0.826	0.595	0.567	0.219	0.805	0.833	-0.044
PCA-10 baseline	0.733	0.543	0.503	0.194	0.842	0.944	-0.007
SVD-10 baseline	0.731	0.541	0.499	0.192	0.842	0.936	0.002

Table 3: Large-cell Robust-V2 pooled donor-holdout benchmark on strict non-overlap mapped cells (100k training cells per split). Values are pooled donor-level means across 88 grouped splits. Coverage varies by method/metric due to valid-run filtering (pseudotime: 88 splits for most methods, 70 for Palantir, 36 for scVI).

Method	$n_{\mathrm{splits}}$	Mean diff	Wilcoxon BH- $q$
H65 hybrid conservative (10D)	88	-0.056	$2.7\times 10^{-7}$
H65 anchor head (10D)	88	-0.092	$2.7\times 10^{-7}$
SVD-10 baseline	88	-0.185	$9.7\times 10^{-15}$
PCA-10 baseline	88	-0.194	$2.6\times 10^{-14}$
Raw baseline (1200 log1p)	88	-0.226	$2.5\times 10^{-14}$
CellTypist stage-probability baseline	88	-0.260	$7.0\times 10^{-15}$

Table 4: Robust-V2 paired pseudotime statistics versus H65 cell head (split-level mean differences; method minus cell head).

Figure 3: Robust-V2 split-level pseudotime deltas versus H65 cell head (method minus cell head). All alternatives are below zero.

Key classification metrics
	Extracted 10D head + MLP2	Frozen scGPT + 3-layer MLP	Direct raw MLP ( $128{\rightarrow}64$ )
Branch bal. acc.	0.827	0.810	0.818
Stage bal. acc.	0.468	0.403	0.504
CD4/CD8 AUROC	0.856	0.817	0.827
Mono/Macro AUROC	0.956	0.946	0.920
Training and deployment cost

Table 5: Summary comparison of the three evaluated approaches on the 12-split repeated benchmark. Performance values are pooled donor-level means. Training times from Supplementary S23; parameter counts from model architectures. BH-sig. wins/losses count endpoints (out of 8) on which the method is BH-significantly better/worse than the extracted head.

Operator variant	Size	BH-sig. losses (of 8)	Key note
Full drift (3 pooled heads)	17.5 MB	—	Reference
Compact top1 (L2H5)	5.9 MB	0/8	Best single head
Top1 + rank 64	0.73 MB	5/8	Viable low-rank boundary
Top1 + rank 32	0.37 MB	6/8	Degraded
Hard sparse (16 factors, 60 genes)	0.12 MB	7/8	Interpretable, baseline-level

Table 6: Compaction chain: from pooled drift operator to single-head and low-rank surrogates. Losses are relative to the full extracted 10D head under paired BH-corrected tests.

Figure 4: Fixed-probe leave-one-factor-out ablation for the rank-64 compact operator. (a) Total pooled ablation impact for the 12 most impactful factors. (b) Cumulative concentration curve—the top four factors explain $66.2\%$ of total ablation impact. (c) Per-task drop magnitudes; asterisks denote BH-significant loss versus the intact model.

Factor	Role	Drop	Top read/write genes	Enriched programs
f01	Branch routing	0.430	R: CSF3R, SLC6A1, IL7R; W: CSF3R, SLC25A37, FCER1G	Mono/macro diff.; naïve CD8 + T; granulopoiesis
f02	Lymphoid contrast	0.254	R: IL7R, GSTPL, IGF2R; W: HLA-DRA, MARCH6, IGKV5	Lymphoid B vs T/NK; CD4 + T maturation
f00	Stage ordering	0.280	R: EPB41, GMR1, VPEL3; W: ISHC, LGALS14, EPB41	Classical monocyte; naïve CD8 + T program
f03	Mono/macro struct.	0.241	R: LGALS3, GT58, LYZ; W: VPEL5, ALPL, SLC28A3	Macrophage; granulocyte; B-cell markers

Table 7: Gene-program summary for the four core factors of the rank-64 compact operator. Ablation drop is the total pooled classification loss when the factor is removed. Top read/write genes and enriched marker sets are from cached local marker enrichment.

Figure 5: Core-subset recovery as a fraction of intact rank-64 performance across the exhaustive 15-subset interaction sweep. Green borders mark the best subset for each endpoint. No subset closes the gap to the intact model (all remain BH-significant); task-specialized circuitry is visible—mono/macro is well-recovered by the pair {f01, f03} alone (0.95 $\times$ ), while branch and stage require all four factors. Best subsets per endpoint are listed in Supplementary Table 22 .

Figure 6: Pseudotime-annotated manifold view used for trajectory interpretation. The projection emphasizes developmental ordering from stem/progenitor regions toward branch-specific lineage endpoints.

Metric	Anchor-trained	Cell-trained
Corr (projected distance vs target)	0.607	0.614
Residualized corr	0.605	0.612
Trustworthiness	0.972	0.969
CD4/CD8 AUC(abs), global 3D	0.744	0.886
Monocyte/Macrophage AUC(abs), global 3D	0.759	0.937
Normalized within-stage spread	0.217	0.314

Table 8: Anchor-trained vs cell-trained head on the same strict non-overlap external panel (global 3D view unless noted).

Metric	Anchor-trained	Cell-trained	Hybrid strong-topology	Hybrid conservative
Corr	0.607	0.614	0.473	0.634
Residualized corr	0.605	0.612	0.472	0.632
Trustworthiness	0.972	0.969	0.969	0.959
CD4/CD8 AUC(abs), global 3D	0.744	0.886	0.934	0.911
Monocyte/Macrophage AUC(abs), global 3D	0.759	0.937	0.914	0.851
Normalized within-stage spread	0.217	0.314	0.348	0.318

Table 9: Transfer-head tradeoff frontier on strict non-overlap external anchors (global 3D metrics).

Method	$n_{\mathrm{rows}}$	Mean signed $\rho$	Mean $\|\rho\|$	Positive-share (%)
H65 anchor head (10D)	307	0.093	0.235	65.1
H65 cell head (10D)	352	0.249	0.439	71.0
H65 hybrid conservative (10D)	307	0.132	0.364	64.5
Raw baseline (1200 log1p)	307	-0.044	0.266	37.5
PCA-10 baseline	307	-0.007	0.259	38.1
SVD-10 baseline	307	0.002	0.279	51.8

Table 10: Robust-V2 pseudotime orientation summary across pooled donor rows. Signed

\rho

is the primary metric;

|\rho|

and positive-sign share are orientation-aware summaries.

Metric	Cell head mean	Best comparator (mean)	Mean diff	Wilcoxon BH- $q$
Branch balanced accuracy	0.828	Raw baseline (0.826)	-0.0018	$1.5\times 10^{-1}$
Branch macro-F1	0.584	Palantir baseline (0.621)	+0.0358	$1.4\times 10^{-10}$
Stage balanced accuracy	0.552	Raw baseline (0.567)	+0.0148	$1.8\times 10^{-7}$
Stage macro-F1	0.216	Palantir baseline (0.242)	+0.0254	$1.3\times 10^{-6}$
CD4/CD8 AUROC	0.867	H65 hybrid conservative (0.858)	-0.0088	$1.1\times 10^{-15}$
CD4/CD8 balanced accuracy	0.798	H65 hybrid conservative (0.794)	-0.0025	$1.3\times 10^{-2}$

Table 11: Robust-V2 metric-level paired significance summary versus H65 cell head (comparator minus cell head; higher is better for all metrics).

Figure 7: Robust-V2 splitwise paired deltas versus the H65 cell head for the eight non-pseudotime endpoints. Each panel shows comparator-minus-cell-head mean differences with 95% bootstrap confidence intervals across grouped donor-holdout splits. Negative values favor the cell head. Red points indicate BH-significant cell-head advantages; gold points indicate BH-significant comparator advantages.

Method	ARI	AMI	NMI	V-measure
H65 anchor head (10D)	0.269	0.546	0.547	0.547
H65 cell head (10D)	0.330	0.520	0.521	0.521
H65 hybrid conservative (10D)	0.387	0.618	0.619	0.619
Raw baseline (1200 log1p)	0.393	0.626	0.627	0.627
PCA-10 baseline	0.305	0.581	0.582	0.582
SVD-10 baseline	0.326	0.591	0.592	0.592

Table 12: Canonical 100k split unsupervised branch-inference benchmark using Scanpy [ 68 ] neighbors

+

Leiden [ 59 ]

+

PAGA. Values are donor-level means across all 4 held-out donors.

Method	$n_{d}$	ARI	AMI	NMI	V-measure
H65 anchor head (10D)	2	0.613	0.670	0.670	0.670
H65 cell head (10D)	4	0.120	0.293	0.294	0.294
H65 hybrid conservative (10D)	3	0.255	0.410	0.410	0.410
Raw baseline (1200 log1p)	3	0.338	0.437	0.437	0.437
PCA-10 baseline	2	0.145	0.288	0.288	0.288
SVD-10 baseline	3	0.358	0.461	0.462	0.462

Table 13: Canonical 100k split unsupervised branch-inference benchmark using Palantir fate assignments after a fair-scope rerun with frozen donor-local evaluation subsets and explicit audit logging.

Figure 8: Supervised branch classification versus unsupervised branch recovery on the canonical 100k donor-holdout split. The task distinction is visible: branch information can be decodable under supervision even when unsupervised branch recovery remains wrapper- and coverage-sensitive.

Figure 9: Branch-inference wrapper comparison. The Scanpy wrapper yields stable full-coverage rankings, whereas Palantir’s strongest mean scores are tied to reduced donor support.

Method	TSP10	TSP14	TSP17	TSP26
Raw baseline (1200 log1p)	ok	fail	ok	ok
PCA-10 baseline	ok	ok	fail	fail
SVD-10 baseline	ok	ok	fail	ok
Frozen scGPT avg-pool	fail	ok	fail	fail
Frozen scGPT CP- <cls>	fail	ok	fail	ok
H65 anchor head (10D)	ok	fail	ok	fail

Table 14: Palantir donor-level success matrix on the canonical 100k donor-holdout split.

Metric	H65 cell head (10D)	scGPT avg-pool	scGPT CP- <cls>
Branch balanced accuracy ( $n_{d}=4$ )	0.866	0.880	0.861
Branch macro-F1 ( $n_{d}=4$ )	0.529	0.526	0.503
Stage balanced accuracy ( $n_{d}=4$ )	0.610	0.593	0.570
Stage macro-F1 ( $n_{d}=4$ )	0.188	0.154	0.145
CD4/CD8 AUROC ( $n_{d}=3$ )	0.890	0.825	0.828
Mono/Macro AUROC ( $n_{d}=2$ )	0.926	0.798	0.775

Table 15: Embedding-ablation comparison on the held-out donor split (20k train, 30,137 test). Values are donor-level means.

Metric	$n_{\mathrm{obs}}$	H65 cell head (4D) + linear	Frozen scGPT avg-pool + MLP	$\Delta_{\mathrm{base-head}}$	95% CI	BH- $q$
Branch macro-F1	40	0.597	0.547	-0.045	[-0.067, -0.023]	0.010
Branch balanced accuracy	40	0.801	0.810	0.010	[-0.001, 0.024]	0.192
Stage macro-F1	40	0.236	0.157	-0.078	[-0.087, -0.068]	0.0013
Stage balanced accuracy	40	0.540	0.403	-0.127	[-0.164, -0.086]	0.0020
CD4/CD8 AUROC	29	0.845	0.817	-0.026	[-0.033, -0.018]	0.0013
CD4/CD8 balanced accuracy	29	0.780	0.744	-0.034	[-0.045, -0.025]	0.0013

Table 16: Repeated grouped-split comparison between a 4D extracted H65 cell head with linear probes and frozen scGPT avg-pool embeddings with 3-layer MLP probes (

n_{\mathrm{split}}=12

). Negative deltas favor the extracted head.

Figure 10: Repeated 12-split comparator between a 4D extracted head and frozen scGPT avg-pool embeddings with 3-layer MLP probes. Left: pooled means. Middle: paired splitwise deltas. Right: runtime decomposition.

Metric	$n_{\mathrm{obs}}$	H65 cell head (10D) + MLP2	Frozen scGPT avg-pool + MLP	$\Delta_{\mathrm{head-base}}$	95% CI	BH- $q$
Branch balanced accuracy	40	0.827	0.810	0.017	[0.012, 0.024]	0.0011
Branch macro-F1	40	0.556	0.547	0.010	[-0.003, 0.022]	0.194
Stage balanced accuracy	40	0.468	0.403	0.063	[0.048, 0.077]	0.0011
Stage macro-F1	40	0.180	0.157	0.024	[0.020, 0.027]	0.0011
CD4/CD8 AUROC	29	0.856	0.817	0.038	[0.032, 0.044]	0.0011
CD4/CD8 balanced accuracy	29	0.779	0.744	0.034	[0.028, 0.040]	0.0011

Table 17: Repeated grouped-split comparison between a 10D extracted H65 cell head with 2-layer MLP probes and frozen scGPT avg-pool + 3-layer MLP probes (

n_{\mathrm{split}}=12

). Positive deltas favor the extracted head.

Figure 11: Probe-depth tradeoff for the extracted 10D H65 cell head on the 40-split Robust-V2 family. The linear head remains the best overall default.

Metric	Frozen scGPT + MLP	Direct raw1200 $1\times 64$	Direct raw1200 $128\rightarrow 64$	Extracted 10D + MLP2
Branch balanced accuracy	0.810	0.814	0.818	0.827
Branch macro-F1	0.547	0.550	0.557	0.556
Stage balanced accuracy	0.403	0.499	0.504	0.468
Stage macro-F1	0.157	0.190	0.186	0.180
CD4/CD8 AUROC	0.817	0.828	0.827	0.856
CD4/CD8 balanced accuracy	0.744	0.758	0.759	0.779

Table 18: Repeated grouped-split comparison between direct tiny-from-scratch raw-expression MLPs, frozen scGPT avg-pool + MLP, and the extracted 10D head + MLP2. Values are pooled donor-level means across 12 splits.

Component	Scope	H65 cell head (4D) + linear	Frozen scGPT avg-pool + MLP
Reference 1200-gene matrix extraction	one-time shared	164.7 s
Method-specific cache build	one-time	99.5 s (drift + head training)	5163.8 s (avg-pool extraction)
Per-split evaluation	mean / split	8.8 s	159.2 s
Trainable parameters / task	downstream	5–170	172,610–174,690

Table 19: Runtime decomposition for the repeated-split 4D-head versus frozen scGPT+MLP comparator.

Figure 12: T-cell signal exists in latent space but is partially hidden in default 3D projection. The T-cell lens exposes this information for visualization.

Figure 13: Phase-2 latent intervention sweeps. Positive monotonic shifts are strongest for stem $\rightarrow$ myeloid and stem $\rightarrow$ erythroid directions.

Figure 14: Fixed-probe leave-one-factor-out audit for the best sparse surrogate. The sparse model preserves the same mechanistic core as the rank64 parent: the top four retained factors explain $68.9\%$ of total sparse ablation impact.

Factor	Role	Drop	Top read/write genes	Enriched programs
f01	Branch routing	0.430	R: EPB41, GBR1, VPEL3, GPX1, TAL001; W: DIF2R, EPB41, VPEL5, RBXP3L, RMT39	Classical and non-classical monocyte identity
f02	Lymphoid contrast	0.254	R: IL7R, GSTPL, DIF2R, ChrFTB, GPFC, FNL; W: HLA-DRA, MARCB9, AC96579.19, IGKC, IGLC1	Naïve CD8 + T-cell; thymic CD4 + ; B-cell lineage
f00	Stage ordering	0.280	R: IL7R, TENTSC, GSPT1, GVPC, GBK1, PRBX2; W: CSF3R, SLC2BA1, S100A8, ELR2, ACSL1, PRBB	Naïve CD8 + T-cell; T-regulatory; granulopoiesis
f03	Mono/macro struct.	0.241	R: LGALS3, BLC2L1, LV2, THBS1, C10A, BLVR8; W: VPEL1, ALPL, SLC28A3, FIG30R, ALT8, TENTSC	Macrophage; phagocyte; granulocyte/neutrophil

Table 20: Gene-program summary for the four core sparse-surrogate factors. Drop is the total pooled ablation impact. Read/write genes and enrichment are from cached local marker analysis. The same biological programs persist after hard-sparse pruning.

Figure 15: Head/layer attribution screen across all 96 scGPT attention units. Color shows external transfer screen score; stars mark top-ranked units.

Unit	Corr resid	Trustworthiness	CD4/CD8 AUC 3D	Mono/Macro AUC 3D
L2H5 (screen rank 1)	0.634	0.985	0.711	0.873
L0H3 (screen rank 2)	0.627	0.981	0.743	0.769
L6H6 (screen rank 3)	0.621	0.984	0.765	0.867
L6H4 (screen rank 4)	0.610	0.986	0.765	0.810
L0H7 (screen rank 5)	0.636	0.977	0.746	0.765

Table 21: Top single heads after full refit on internal anchors and zero-shot evaluation on strict non-overlap external anchors.

Endpoint	Best subset	Size	Best / intact
Branch BA	$\{\textbf{f01},\textbf{f02},\textbf{f00},\textbf{f03}\}$	4	69.7%
Branch F1	$\{\textbf{f01},\textbf{f02},\textbf{f00},\textbf{f03}\}$	4	71.7%
Stage BA	$\{\textbf{f01},\textbf{f02},\textbf{f00},\textbf{f03}\}$	4	44.0%
Stage F1	$\{\textbf{f01},\textbf{f02},\textbf{f00},\textbf{f03}\}$	4	45.5%
CD4/CD8 AUROC	$\{\textbf{f02},\textbf{f00},\textbf{f03}\}$	3	83.1%
CD4/CD8 BA	$\{\textbf{f01},\textbf{f00},\textbf{f03}\}$	3	82.9%

Table 22: Best subset from the exhaustive 15-subset interaction sweep for each audited endpoint. Every listed subset remains BH-significantly below the intact rank64 model.

Figure 16: H38 internal canonical manifold. Best internal embedding using 30 interaction-path pairs with square-root-weighted distances. This branch captures a structured intercellular-communication axis system distinct from the hematopoietic developmental geometry in the main text.

Figure 17: H38 negative-control distributions. Observed H38 alignment was not cleanly separated from all control draws under early external conditions, motivating the rescue program below.

Run (key change)	External corr	$\Delta$ vs PCA2
v4 (pair hygiene + hard negatives)	0.8850	-0.0909
v5 (cell-trained transfer head)	0.9624	-0.0199
v7 (metric head + cleaned pair set + ensemble)	0.9707	-0.0061
v9 (cell-metric head + ensemble)	0.9800	+0.0032
v10 (MLP-geom/dim6 extras)	0.9747	-0.0021
v11 (raw diverse 100k-cell internal training + ensemble)	0.9825	+0.0057

Table 23: H38 rescue progression on external zero-shot evaluation (minimum 40 cells per anchor group).

Single-Cell Transcriptomics

Single-cell transcriptomics measures gene activity cell by cell to map cell types, states, interactions, and responses across tissues and conditions.

See all 65 papers

Feb 25, 2026

What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses

When biological foundation models such as scGPT and Geneformer process single-cell gene expression, what geometric and topological structure forms in their internal representations? Is that structure biologically meaningful or a training artifact, and how confident should we be in such claims? We address these questions through autonomous large-scale hypothesis screening: an AI-driven executor-brainstormer loop that proposed, tested, and refined 141 geometric and topological hypotheses across 52 iterations, covering persistent homology, manifold distances, cross-model alignment, community structure, and directed topology, all with explicit null controls and disjoint gene-pool splits. Three principal findings emerge. First, the models learn genuine geometric structure. Gene embedding neighborhoods exhibit non-trivial topology, with persistent homology significant in 11 of 12 transformer layers at p < 0.05 in the weakest domain and 12 of 12 in the other two. A multi-level distance hierarchy shows that manifold-aware metrics outperform Euclidean distance for identifying regulatory gene pairs, and graph community partitions track known transcription factor target relationships. Second, this structure is shared across independently trained models. CCA alignment between scGPT and Geneformer yields canonical correlation of 0.80 and gene retrieval accuracy of 72 percent, yet none of 19 tested methods reliably recover gene-level correspondences. The models agree on the global shape of gene space but not on precise gene placement. Third, the structure is more localized than it first appears. Under stringent null controls applied across all null families, robust signal concentrates in immune tissue, while lung and external lung signals weaken substantially.

Ihor Kendiukhov

Department of Computer Science, University of Tübingen, Tübingen, Germany

Neural Representation Geometry Genomic Foundation Models Geometric Representation Learning Topological Representation Learning Single-Cell Transcriptomics

Ask PDF

Feb 19, 2026

Systematic Evaluation of Single-Cell Foundation Model Interpretability Reveals Attention Captures Co-Expression Rather Than Unique Regulatory Signal

We present a systematic evaluation framework - thirty-seven analyses, 153 statistical tests, four cell types, two perturbation modalities - for assessing mechanistic interpretability in single-cell foundation models. Applying this framework to scGPT and Geneformer, we find that attention patterns encode structured biological information with layer-specific organisation - protein-protein interactions in early layers, transcriptional regulation in late layers - but this structure provides no incremental value for perturbation prediction: trivial gene-level baselines outperform both attention and correlation edges (AUROC 0.81-0.88 versus 0.70), pairwise edge scores add zero predictive contribution, and causal ablation of regulatory heads produces no degradation. These findings generalise from K562 to RPE1 cells; the attention-correlation relationship is context-dependent, but gene-level dominance is universal. Cell-State Stratified Interpretability (CSSI) addresses an attention-specific scaling failure, improving GRN recovery up to 1.85x. The framework establishes reusable quality-control standards for the field.

Ihor Kendiukhov

Department of Computer Science, University of Tübingen, Tübingen, Germany

Mechanistic Interpretability Neural Network Interpretability Attention Head Analysis Gene Regulatory Network Inference Single-Cell Transcriptomics Single-Cell Perturbation Modeling

Ask PDF

Mar 3, 2026

Sparse autoencoders reveal organized biological knowledge but minimal regulatory logic in single-cell foundation models: a comparative atlas of Geneformer and scGPT

Background: Single-cell foundation models such as Geneformer and scGPT encode rich biological information, but whether this includes causal regulatory logic rather than statistical co-expression remains unclear. Sparse autoencoders (SAEs) can resolve superposition in neural networks by decomposing dense activations into interpretable features, yet they have not been systematically applied to biological foundation models. Results: We trained TopK SAEs on residual stream activations from all layers of Geneformer V2-316M (18 layers, d=1152) and scGPT whole-human (12 layers, d=512), producing atlases of 82525 and 24527 features, respectively. Both atlases confirm massive superposition, with 99.8 percent of features invisible to SVD. Systematic characterization reveals rich biological organization: 29 to 59 percent of features annotate to Gene Ontology, KEGG, Reactome, STRING, or TRRUST, with U-shaped layer profiles reflecting hierarchical abstraction. Features organize into co-activation modules (141 in Geneformer, 76 in scGPT), exhibit causal specificity (median 2.36x), and form cross-layer information highways (63 to 99.8 percent). When tested against genome-scale CRISPRi perturbation data, only 3 of 48 transcription factors (6.2 percent) show regulatory-target-specific feature responses. A multi-tissue control yields marginal improvement (10.4 percent, 5 of 48 TFs), establishing model representations as the bottleneck. Conclusions: These models have internalized organized biological knowledge, including pathway membership, protein interactions, functional modules, and hierarchical abstraction, yet they encode minimal causal regulatory logic. We release both feature atlases as interactive web platforms enabling exploration of more than 107000 features across 30 layers of two leading single-cell foundation models.

Ihor Kendiukhov

Department of Computer Science, University of Tübingen, Tübingen, Germany

Genomic Foundation Models Single-Cell Transcriptomics Neural Network Interpretability Transformer Interpretability

Ask PDF

Abstract

Single-Cell Transcriptomics

Explore similar work

What Topological and Geometric Structure Do Biological Foundation Models Learn? Evidence from 141 Hypotheses

Systematic Evaluation of Single-Cell Foundation Model Interpretability Reveals Attention Captures Co-Expression Rather Than Unique Regulatory Signal

Sparse autoencoders reveal organized biological knowledge but minimal regulatory logic in single-cell foundation models: a comparative atlas of Geneformer and scGPT