Central Dogma Transformer III: Interpretable AI Across DNA, RNA, and Protein

Nobuyuki Ota

cs.LGMar 24, 2026

Central Dogma Transformer III: Interpretable AI Across DNA, RNA, and Protein

Authors: Nobuyuki Ota

Organizations: Independent Researcher, Burlingame, CA, USA

Abstract

Biological AI models increasingly predict complex cellular responses, yet their learned representations remain disconnected from the molecular processes they aim to capture. We present CDT-III, which extends mechanism-oriented AI across the full central dogma: DNA, RNA, and protein. Its two-stage Virtual Cell Embedder architecture mirrors the spatial compartmentalization of the cell: VCE-N models transcription in the nucleus and VCE-C models translation in the cytosol. On five held-out genes, CDT-III achieves per-gene RNA r=0.843 and protein r=0.969. Adding protein prediction improves RNA performance (r=0.804 to 0.843), demonstrating that downstream tasks regularize upstream representations. Protein supervision sharpens DNA-level interpretability, increasing CTCF enrichment by 30%. Analysis of experimentally measured mRNA and protein responses reveals that the majority of genes with observable mRNA changes show opposite protein-level changes (66.7% at |log2FC|>0.01, rising to 87.5% at |log2FC|>0.02), exposing a fundamental limitation of RNA-only perturbation models. Despite this pervasive direction discordance, CDT-III correctly predicts both mRNA and protein responses. Applied to in silico CD52 knockdown approximating Alemtuzumab, the model predicts 29/29 protein changes correctly and rediscovers 5 of 7 known clinical side effects without clinical data. Gradient-based side effect profiling requires only unperturbed baseline data (r=0.939), enabling screening of all 2,361 genes without new experiments.

Figures & tables

Figure 1: CDT-III two-stage Virtual Cell Embedder architecture. VCE-N (nuclear stage, identical to CDT-II) processes DNA Enformer embeddings and per-cell RNA expression through self-attention and cross-attention, modeling transcription to produce a cell-level RNA embedding. VCE-C (cytosolic stage) takes this RNA embedding and protein expression, modeling translation to produce a protein embedding. Each stage has an independent task head.

Gene	Cells	RNA $r$ (CDT-II)	RNA $r$ (CDT-III)	$\Delta$	Protein $r$ (expr.)
GFI1B	477	0.851	0.885	+0.034	0.969
TNFSF9	349	0.831	0.868	+0.037	0.991
TFRC	847	0.824	0.854	+0.030	0.989
CD44	233	0.795	0.858	+0.063	0.935
CD52	131	0.719	0.748	+0.029	0.962
Mean		0.804	0.843	+0.039	0.969

Table 1: Per-gene prediction performance. Per-gene mean

r

(pseudo-bulk) on five held-out validation genes. CDT-II values are from the same checkpoint used for CDT-III weight transfer. RNA per-gene mean

r

improves from CDT-II (

0.804

) to CDT-III (

0.843

) despite adding a second task. Protein per-gene mean

r

on 65 expressed proteins reaches

0.969

Figure 2: Prediction performance. (A) Single-stage vs. two-stage VCE comparison (RNA $r$ by architecture). (B) Per-gene RNA correlation: CDT-II vs. CDT-III. (C) Per-gene protein predicted vs. actual (65 expressed proteins, 5 genes).

Metric	CDT-II	CDT-III	Change
Mean CTCF enrichment	$6.6\times$	$8.59\times$	$+30\%$
Genes $>2\times$ enriched	—	27/27 (100%) †	—
Permutation $P$	$<0.001$	$<0.001$	—
Hi-C mean contact ratio	—	$1.30\times$	—
Genes with ratio $>1$	—	16/25 (64%)	—
Wilcoxon $P$	—	0.020	—

Table 2: CTCF enrichment and Hi-C validation. Protein supervision improves DNA-level interpretability.

Panel of Figure 3 — Figure 3: Protein supervision improves DNA interpretability. (A) Hi-C contact ratio per gene (top 20 attention bins vs. random 20 bins; mean $=1.30\times$ , $P=0.020$ ). (B) Attention-high vs. random bins scatter plot showing that high-attention genomic regions have stronger physical contact with promoters.

Known side effect	Incidence	CDT-III predicted proteins	Match
Autoimmune thyroid disease	36–42%	FCGR2A $\uparrow$ , BAFF-R $\uparrow$ , CD40 $\uparrow$	✓
ITP	2–3%	FCGR2A $\uparrow$	✓
Infection risk	66–77%	CD58 $\uparrow^{\star}$ , TFRC $\uparrow^{\star}$	✓
Infusion reaction	$>$ 90%	ICAM1 $\uparrow$ , CD29 $\uparrow$	✓
Autoimmune nephropathy	0.2–0.3%	MCAM $\uparrow$	✓
Goodpasture syndrome	$<$ 0.1%	—	✗

Table 3: CDT-III predictions vs. known Alemtuzumab side effects. Five of seven known side effects are recapitulated by CDT-III’s protein predictions, without access to any clinical data.

Panel of Figure 4 — Figure 4: In silico Alemtuzumab side effect prediction. (A) CD52 knockdown RNA prediction vs. actual (per-gene $r=0.748$ ). (B) Protein prediction vs. actual for 65 expressed proteins ( $r=0.962$ ). (C) Side effect map: predicted vs. measured protein changes for 29 proteins with detectable effects (29/29 direction agreement, color-coded by clinical relevance).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: CDT-II architecture. The model mirrors the central dogma: DNA self-attention captures genomic relationships within a $\pm$ 57 kb window, RNA self-attention captures gene co-regulation, and cross-attention models transcriptional control. A Virtual Cell Embedder integrates both modalities to predict perturbation effects. CDT-III’s VCE-N preserves this architecture exactly, enabling 100% weight transfer.

	Architecture	$\lambda$	RNA $r$	Protein $r$	Fusion transfer
CDT-II	RNA only	—	0.64	—	—
Exp A	1-Stage	1.0	0.19	0.76 ∗	✗
Exp B	1-Stage	0.11	0.24	0.77 ∗	✗
Exp C	1-Stage	0.05	0.35	0.81 ∗	✗
Exp D	1-Stage	0.01	0.37	0.80 ∗	✗
Exp E	1-Stage ( $\lambda$ warmup)	0.001 $\to$ 0.01	0.30	0.21	✗

Appendix

Table 4: Single-stage vs. two-stage VCE. Per-cell correlation (

r

) on the validation set. All models transfer CDT-II encoder and attention weights; the critical difference is the VCE fusion layer. Single-stage architectures modify fusion dimensions (

d\times 2\to d\times 3

), preventing fusion weight transfer and causing RNA prediction to collapse despite encoder transfer. The two-stage architecture preserves full weight compatibility including fusion. ∗ Protein

r

values for single-stage models are inflated by 124/189 non-expressed proteins predicted as

\sim

Attention Map	Shape	Biological Meaning	Source
DNA Self-Attention (L0)	$[8,896,896]$	Genomic interactions	CDT-II
DNA Self-Attention (L1)	$[8,896,896]$	Higher-order genomic	CDT-II
RNA Self-Attention	$[8,2361,2361]$	Gene co-regulation	CDT-II
DNA $\to$ RNA Cross-Attention	$[8,2361,896]$	Transcriptional control	CDT-II
Protein Self-Attention	$[8,189,189]$	Protein co-regulation	CDT-III
RNA $\to$ Protein Cross-Attention	$[8,189,2361]$	Translational control	CDT-III

Appendix

Table 5: Attention map inventory. CDT-III produces six interpretable attention maps, each corresponding to a specific biological process. Four are inherited from CDT-II; two are new.

Gene	CTCF	Top 10%	Enrich.	Gene	CTCF	Top 10%	Enrich.
CD82	7	7	10.0 $\times$	SLC3A2	6	5	8.33 $\times$
CD69	7	7	10.0 $\times$	TNFSF9	5	4	8.00 $\times$
CD40	9	9	10.0 $\times$	CD44	4	3	7.50 $\times$
ITGB1	2	2	10.0 $\times$	CD58	4	3	7.50 $\times$
ENTPD1	2	2	10.0 $\times$	CD55	4	3	7.50 $\times$
KLRG1	3	3	10.0 $\times$	LGALS3	4	3	7.50 $\times$

Appendix

Table 6: CTCF enrichment for all 28 genes with ENCODE K562 ChIP-seq data. 14/27 genes with CTCF sites show perfect 10.0

\times

enrichment (all CTCF sites in top 10% attention bins).

Exp	$\lambda$	Strategy	RNA $r$	Prot $r$	Epochs	Notes
A	1.0	Fixed	0.19	0.76	148	Protein-dominated loss
B	0.11	Fixed	0.24	0.77	209	Equal-weight by variance
C	0.05	Fixed	0.35	0.81	386	Best single-stage protein
D	0.01	Fixed	0.37	0.80	472	Best single-stage RNA
E	0.001 $\to$ 0.01	Warmup	0.30	0.21	$\sim$ 300	Curriculum failed

Appendix

Table 7: Full results for all five single-stage VCE experiments.

Rank	Protein (Gene)	Gradient	Category
1	CD32 (FCGR2A)	$+0.059\uparrow$	Fc receptor / ADCC
2	CD146 (MCAM)	$+0.044\uparrow$	Endothelial adhesion
3	CD54 (ICAM1)	$+0.043\uparrow$	Inflammatory adhesion
4	CD81 (CD81)	$+0.031\uparrow$	B cell co-receptor
5	CD58 (CD58)	$+0.030\uparrow$	T cell activation (LFA-3)
6	CD224 (GGT1)	$+0.030\uparrow$	Oxidative stress

Appendix

Table 8: Top 20 proteins predicted to change upon CD52 knockdown (gradient-based analysis).

Gene	Top 20 Hi-C	Random 20 Hi-C	Ratio	Note
CD55	156.4	56.8	$2.75\times$	Complement regulator
CD82	66.2	29.0	$2.28\times$	Tetraspanin
IL21R	68.6	31.2	$2.20\times$	IL-21 receptor
FCGR2A	93.8	45.1	$2.08\times$	Fc receptor
CD83	71.5	36.0	$1.99\times$	Dendritic cell
ITGB1	51.3	29.9	$1.72\times$	Integrin $\beta$ 1

Appendix

Table 9: Per-gene Hi-C contact ratios for all 25 genes with sufficient coverage (4DN K562 data, 5 kb resolution). Three genes excluded due to zero Hi-C records: B2M , TFRC , GFI1B . Mean ratio

=1.30\times

; 16/25 genes show ratio

>1

(

P=0.020

, Wilcoxon signed-rank).

Metric	All proteins (189)	Expressed proteins (65)
Pearson $r$	0.913	0.939
Spearman $\rho$	0.752	0.865
Direction agreement	145/189 (76.7%)	63/65 (96.9%)
Top 20 direction agreement	20/20 (100%)	—

Appendix

Table 10: Gradient analysis validation: NTC mean vs. perturbation cell input. Concordance between protein-level gradients (

\partial\hat{y}_{\text{Prot},j}/\partial x_{\text{RNA, CD52}}

) computed using NTC mean input (no perturbation data) vs. actual CD52-perturbed cell input.

$\|\text{log2FC}\|$ threshold	$N$ genes	Opposite	Same	% Opposite
$>0$ (all nonzero)	26	16	10	61.5%
$>0.005$	16	10	6	62.5%
$>0.01$	12	8	4	66.7%
$>0.02$	8	7	1	87.5%
$>0.05$	4	4	0	100.0%

Appendix

Table 11: Sensitivity analysis of mRNA–protein direction divergence. For each mRNA change threshold, we report the number of gene–protein pairs with opposite vs. same direction of change following CD52 knockdown. mRNA: full transcriptome log2FC; protein: DSB-normalized effects (65 expressed proteins, 27 matched pairs).

Figure 6: mRNA vs. surface protein response divergence after CD52 knockdown. Left: Paired bar chart for 12 genes with $|\text{mRNA log2FC}|>0.01$ (DSB-normalized protein), showing mRNA change (blue) and protein change (red, scaled) for each gene. Dark colors indicate opposite-direction pairs (8/12, 66.7%); light colors indicate same-direction pairs. The dominant pattern—mRNA decreases while surface protein increases—is consistent with post-translational regulation. Right: Sensitivity analysis showing the percentage of gene–protein pairs with opposite directions at increasing mRNA change thresholds. The divergence rate rises from 61.5% to 100% as the threshold increases, indicating that genes with the strongest mRNA responses show the most pronounced protein-level discordance.

Figure 7: CD55 cross-attention genome track. DNA

\to

RNA cross-attention profile (top) aligned with CTCF ChIP-seq peaks (bottom) across the

\pm

57 kb Enformer window. Red lines: CTCF sites in top 10% attention (3/4). TSS indicated by dotted line. Hi-C contact ratio:

2.75\times

(Table 9 ).

Figure 8: FCGR2A cross-attention genome track. Same format as Figure 7 . CTCF site in top 10% attention (1/2) located upstream of TSS, consistent with distal regulatory contact. Hi-C contact ratio: $2.08\times$ .

Bioinformatics

See all 41 papers

Central Dogma Transformer II: An AI Microscope for Understanding Cellular Regulatory Mechanisms
Feb 9, 2026Nobuyuki OtaTransformer Interpretability Single-Cell Perturbation Modeling
Independent Researcher, Burlingame, CA, USA
When Do Biological Reasoning Models Use Their Biological Inputs?
Oct 1, 2026Ada Fang, Nikitha Thoduguli, Lukas Fesser +3Faithfulness of Language Model Explanations Causal Interventions in Language Models
Harvard University · Massachusetts Institute of Technology
AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling
Apr 22, 2026Zhenyu Wang, Geyan Ye, Wei Liu +1Multimodal Graph Learning
AI for Life Sciences Lab, Tencent, Shenzhen, China

Feb 9, 2026J/K move · Enter open · S save

Central Dogma Transformer II: An AI Microscope for Understanding Cellular Regulatory Mechanisms

Nobuyuki Ota

Independent Researcher, Burlingame, CA, USA

Transformer Interpretability Single-Cell Perturbation Modeling Transformer Attention Gene Regulatory Network Inference Genomics

Open paper AskPDF arXiv

Motivation: Interpretability is not optional in biology: understanding gene regulation requires models whose learned structure can be directly interrogated, not merely accurate predictors whose internals resist mapping onto regulatory relationships. We ask whether an architecture mirroring the central dogma yields attention and gradient maps that recover known regulatory elements and networks in inspectable form. Results: Central Dogma Transformer II (CDT-II) mirrors the central dogma in its architecture -- DNA self-attention, RNA self-attention, and DNA-to-RNA cross-attention -- requiring only genomic embeddings and raw per-cell expression. On K562 CRISPR interference (CRISPRi) data with five genes held out entirely, CDT-II predicts perturbation effects (per-gene mean r = 0.84), recovers the GFI1B regulatory network (6.6-fold enrichment, P = 3.5 x 10^-17), and concentrates cross-attention on ENCODE regulatory elements including CTCF sites (mean 7.67x across 28 target genes, P < 0.001). Gradient attribution predicts consequences of perturbing therapeutic targets (mean r = 0.82). For TFRC, target of the anti-TfR1 antibody PPMX-T003, it identifies erythrocyte-structure, iron-dependent DNA-synthesis and oxidative-stress genes, matching anemia and ferroptosis reported clinically and preclinically -- without clinical data as input. CDT-II acts as an AI microscope, surfacing clinically relevant regulatory structure from perturbation experiments alone. Availability: Source code is available at https://github.com/nobusama/CDT2. Pre-computed embeddings, training data, and model weights are available at https://huggingface.co/datasets/nobusama17/CDT2-data.

Abstract

Bioinformatics

Explore similar work

Central Dogma Transformer II: An AI Microscope for Understanding Cellular Regulatory Mechanisms

When Do Biological Reasoning Models Use Their Biological Inputs?

AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling