RiboUnmix: Learning Shared Translational Dynamics from Biased and Noisy Ribo-seq Measurements

Gabriele Martino; Denis Skibinski; Ivo L. Hofacker; Sebastian Tschiatschek

cs.LGSep 30, 2026

RiboUnmix: Learning Shared Translational Dynamics from Biased and Noisy Ribo-seq Measurements

Authors: Gabriele Martino, Denis Skibinski, Ivo L. Hofacker, Sebastian Tschiatschek

Organizations: Doctoral School Computer Science University of Vienna

Abstract

Ribosome profiling (Ribo-seq) measures ribosome distributions along mRNAs, but observed occupancy profiles also contain experiment-specific distortions and stochastic variability. Consequently, models that accurately predict measured profiles may reproduce technical effects rather than recover the underlying biology. We ask whether jointly modeling datasets collected under different experimental conditions can reveal shared, sequence-dependent patterns of ribosome occupancy. We introduce RiboUnmix, a probabilistic multi-dataset framework in which each expected measured profile is represented as a shared sequence-dependent signal modulated by a dataset-specific multiplicative factor. A negative-binomial observation model captures variability across replicates. We evaluate RiboUnmix on a controlled synthetic benchmark combining programmed translation kinetics, ribosome traffic, stochastic count sampling, and sequence-dependent experimental distortions. Because the underlying kinetics and distortions are known, recovery of the shared profile and dataset-specific effects can be assessed separately. Both inferred components correlate strongly with their targets, demonstrating that RiboUnmix can disentangle shared kinetic patterns from experimental effects. Across four organism-specific real-data benchmarks, RiboUnmix outperforms sequence-to-profile baselines in predicting measured profiles. Models trained independently on subsets of 114 HEK-derived datasets recover concordant shared profiles for held-out transcripts, and experiments varying the number and composition of training datasets show that the learned representation remains stable. RiboUnmix thus converts variation across experiments into evidence for reproducible sequence-dependent patterns of ribosome occupancy, supporting biological hypothesis generation from diverse Ribo-seq datasets.

Figures & tables

Figure 1: Overview of RiboUnmix. A dataset-independent sequence encoder produces the positive, mean-one shared profile

\mathbf{L}_{t}

. A separate dataset-conditioned encoder predicts positional corrections

\bm{\gamma}_{t,d}

and NB2 dispersion parameters. Together with the observed replicate-specific scale

S_{t,d}^{(r)}

, these components define the expected count profile

\bm{\mu}_{t,d}^{(r)}=S_{t,d}^{(r)}(\mathbf{L}_{t}\odot\bm{\gamma}_{t,d})

. Raw replicates supervise the count likelihood, while their arithmetic consensus is used by auxiliary profile-agreement objectives.

Figure 2: Observation bias and sequencing depth. Both panels show transcript ENST00000319974.6 . (A) Mean of two simulated ribosome profiles,

\overline{\mathbf{q}}_{t}

(blue). At positions affected by the

3^{\prime}

- CC bias, stems and triangles show the expected profile

\bar{q}_{t,i}b_{t,f,i}

per unit depth. (B) Mean of two unbiased NB2 count replicates at nominal depths

C\in\{0.25,2,20\}

reads per codon.

Figure 3: Recovery of the simulated ribosome profile

\overline{\mathbf{q}}_{t}

and centered observation effects

\mathbf{g}^{\star}_{t,d}

. (A) Median transcript PCC (left axis) and RMSE (right axis) between

\widehat{\mathbf{L}}_{t}

and

\overline{\mathbf{q}}_{t}

. Here RMSE is calculated after log transforming and position-centering each positive profile. (B) Median transcript

\operatorname{PCC}(\widehat{\mathbf{L}}_{t},\mathbf{K}_{t})

. (C) Mean dataset PCC (left axis) and joint dataset–position RMSE (right axis) between the fitted and exact centered log-corrections. Colors denote separate sequencing depths.

Figure 4: Shared-profile reproducibility across heterogeneous real-data collections. (A) Transcript-level shared-profile PCC for all six balanced, source-disjoint panel pairs. Dots mark medians; bars show IQR and 5th–95th percentiles. (B) Mean PCC with each reference policy’s own

N=2

shared profile along the nested best-first path. Bars are 95% transcript-bootstrap intervals.

Model	C. elegans	E. coli	Human (HEK293T)	S. cerevisiae
iXnos	0.271 $\pm$ 0.016	0.350 $\pm$ 0.028	0.472 $\pm$ 0.014	0.444 $\pm$ 0.019
RiboExp	0.253 $\pm$ 0.018	0.374 $\pm$ 0.024	0.418 $\pm$ 0.013	0.456 $\pm$ 0.020
Riboformer	0.257 $\pm$ 0.021	0.466 $\pm$ 0.032	0.465 $\pm$ 0.017	0.435 $\pm$ 0.016
Seq2Ribo	0.201 $\pm$ 0.017	0.236 $\pm$ 0.028	0.222 $\pm$ 0.012	0.347 $\pm$ 0.014
RiboMIMO	0.325 $\pm$ 0.024	0.516 $\pm$ 0.028	0.461 $\pm$ 0.016	0.502 $\pm$ 0.019
RiboUnmix	0.335 $\pm$ 0.026	0.556 $\pm$ 0.027	0.539 $\pm$ 0.019	0.517 $\pm$ 0.024

Table 1: Pearson correlation benchmark across four organism. Entries report the median transcript-level Pearson correlation

\pm

half the width of their 95% percentile bootstrap intervals. Metrics use positive-count positions. All baseline rows use the full RiboUnmix training cohort and the same transcript reliability weights. Table 10 reports native and matched training settings, together with Spearman correlation and normalized RMSE. The largest estimate per organism is shown in bold.

Appendix figures & tables28 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Synthetic-data generation. Sequence-dependent dwell times define the programmed kinetic profile

\mathbf{K}_{t}

and parameterize an open extended TASEP with 10-codon ribosome footprints. Two independent simulations produce the simulated ribosome profiles

\mathbf{q}_{t}^{(1)}

and

\mathbf{q}_{t}^{(2)}

. Sequence features of idealized 30-nt ribosome-protected fragments determine the positions affected by each observation bias, illustrated here by a

3^{\prime}

- CC bias. These multipliers modify the expected counts before NB2 sampling.

Bias feature	Multiplier	P-sites	Transcripts
3 ′ CC	5.0–6.0 $\times$	943,724	19,266
3 ′ GG	4.0–4.4 $\times$	577,402	19,240
3 ′ AA	2.8–3.2 $\times$	785,775	19,160
3 ′ UU	5.8–6.2 $\times$	646,563	19,068
5 ′ CC	5.2–5.6 $\times$	706,527	19,242
5 ′ GG	4.2–4.6 $\times$	738,405	19,252

Appendix

Table 2: Artificial sequence-dependent bias conditions. Multiplier strengths are synthetic stress-test choices, not estimates of experimentally measured Ribo-seq bias. AU-rich fragments contain more than 70% A and U nucleotides. GC-rich fragments contain more than 70% G and C nucleotides.

Figure 6: Synthetic-generation layers for transcript ENST00000319974.6 . (A) The deterministic programmed kinetic profile

\mathbf{K}_{t}

(black dashed), the mean simulated ribosome profile

\overline{\mathbf{q}}_{t}

(blue), and a translucent band spanning the two separately simulated ribosome profiles

\mathbf{q}_{t}^{(1)}

and

\mathbf{q}_{t}^{(2)}

at every position. (B) The exact

3^{\prime}

- CC bias multiplier

\mathbf{b}_{t,f}

; red triangles mark affected P-site coordinates and the horizontal reference denotes

b_{t,f,i}=1

. (C) Panel C shows the mean simulated ribosome profile

\overline{\mathbf{q}}_{t}

and the biased expected profile per unit depth

\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,f}

before count sampling. (D) The biased expectation and the two saved NB2 count replicates divided only by the nominal depth

C=20

for display. The biased expectations and depth-adjusted counts are not renormalized by their positional means. Zeros and peaks are retained without smoothing, clipping, interpolation, or pseudocounts. The same two simulations are reused across bias conditions and depths.

\overline{\mathbf{q}}_{t}

is their mean is their mean simulated ribosome profile.

Figure 7: Ten deterministic observation effects across sequencing depths. Transcript ENST00000306954.5 is the shortest transcript in the prespecified 100–149-sense-codon range with at least one saved affected P-site for every condition. Rows correspond to the ten bias conditions, and columns to nominal sequencing depths

C\in\{0.25,2,20\}

expected reads per codon. Each cell shows the mean simulated ribosome profile

\overline{\mathbf{q}}_{t}

(light gray), the exact biased expectation per unit depth

\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,f}

(black dashed), and the arithmetic mean of the two sampled NB2 count replicates divided only by

C

(blue). Red triangles identify the exact affected coordinates. The three panels within a row share one y-axis range; different rows may differ. The biased expectation and sampled profiles are not normalized by their positional means. All zeros and peaks are retained without smoothing, clipping, interpolation, or pseudocounts. All conditions and depths reuse the same two finite simulations, whose mean simulated ribosome profile is

\overline{\mathbf{q}}_{t}

Figure 8: Replicate agreement across sequencing depths. Each box summarizes the within-transcript Pearson correlation between

\mathbf{Y}_{t,f}^{(1)}

and

\mathbf{Y}_{t,f}^{(2)}

for one bias condition and depth. Boxes show medians and interquartile ranges; whiskers show the 5th and 95th percentiles across 19,283 transcripts.

Figure 9: Agreement with expected and unbiased profiles. (A) Pearson correlation of each count replicate

\mathbf{Y}_{t,f}^{(r)}

with its expected profile per unit nominal depth

\mathbf{q}_{t}^{(r)}\odot\mathbf{b}_{t,f}

, averaged over the two replicates within each transcript. (B) Pearson correlation of the mean count profile

\overline{\mathbf{Y}}_{t,f}

with the expected profile per unit nominal depth,

\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,f}

. (C) Pearson correlation of

\overline{\mathbf{Y}}_{t,f}

with the mean simulated ribosome profile

\overline{\mathbf{q}}_{t}

, before observation bias is applied. Multiplication or division by nominal depth does not change Pearson correlation. Colors denote sequencing depth. Points show medians, thick intervals show interquartile ranges, and thin intervals show the 5th–95th percentiles across 19,283 transcripts.

Figure 10: Agreement between bias conditions before and after count sampling. (A–C) Median within-transcript Pearson correlation between $\overline{\mathbf{Y}}_{t,f}$ and $\overline{\mathbf{Y}}_{t,g}$ for each pair of conditions at $C=0.25,2,20$ reads per codon. (D) The corresponding correlations between the expected profiles per nominal depth $\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,f}$ and $\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,g}$ . All matrices use the same condition order and color scale.

Figure 11: Agreement with the programmed kinetic profile. Blue and orange show within-transcript Pearson correlations of the two count replicates with

\mathbf{K}_{t}

. Green shows the correlation of their arithmetic mean with

\mathbf{K}_{t}

. Boxes show medians and interquartile ranges; whiskers show the 5th–95th percentiles.

Figure 12: Single-dataset reconstruction and count-replicate agreement. For each observation condition and depth, gray boxes show

\operatorname{PCC}(\mathbf{Y}_{t,d}^{(1)},\mathbf{Y}_{t,d}^{(2)})

and purple boxes show

\operatorname{PCC}(\bm{\mu}_{t,d},\overline{\mathbf{Y}}_{t,d})

from the minimum-validation-loss model. Boxes show medians and interquartile ranges; whiskers show the 5th–95th percentiles. The fitted mean is evaluated on the observed-count scale.

Figure 13: Agreement with the reference-selected shape and mean simulated ribosome profile. For every panel size, connected points show median transcript-level Pearson correlations and vertical intervals show 95% transcript-bootstrap intervals for (A)

\widehat{\mathbf{L}}_{t}

versus

\mathbf{H}_{t}^{(N)}

, (B)

\mathbf{H}_{t}^{(N)}

versus

\overline{\mathbf{q}}_{t}

, and (C)

\widehat{\mathbf{L}}_{t}

versus

\overline{\mathbf{q}}_{t}

. Panels A and C use 1,919, 1,919, and 1,920 eligible validation transcripts matched across all panel sizes at increasing depth. Panel B is depth-independent by construction and uses the deduplicated union of 5,188 validation identities. Intervals are conditional on one training seed and one cumulative bias ordering; they quantify transcript heterogeneity, not variation across independently sampled panels.

Figure 14: Panel-average bias variation and reference-target error. Each point represents a transcript–panel pair. The horizontal axis is the positional standard deviation of the log-geometric-mean multiplier, and the vertical axis is the Pearson correlation error between the reference-selected shape and the mean simulated ribosome profile. Color denotes the number of datasets in the panel. The association is descriptive because transcripts recur across panel sizes.

Figure 15: Shared-profile recovery when all read depths are trained jointly. Each added bias family contributes one dataset at each of the three nominal depths, so the horizontal axis corresponds to

N=3B

datasets. Points show median transcript-level PCC and error bars show 95% transcript-bootstrap intervals over the common 1,920-transcript validation cohort. Equal and depth-ranked reference weights are shown separately. The two schemes assign equal total weight to every bias family and therefore define identical reference-selected shapes

\mathbf{H}_{t}^{(N)}

in this experiment; their comparison isolates only finite-training sensitivity to the reference weighting.

Reference	Mean pair PCC	Pooled PCC	Log- RMSE	Calibration slope	Affected sites within 10%
Equal	$0.9992$	$0.9991$	$0.0169$	$0.9765$	$94.79\%$
Depth-ranked	$0.9994$	$0.9993$	$0.0136$	$0.9965$	$97.83\%$

Appendix

Table 3: Dataset-specific correction recovery for the 30-dataset joint-depth panel. Metrics compare learned and exact gauge-fixed log-corrections. The calibration slope is from the pooled retained positions.

Figure 16: Deterministic observation effects before count sampling. (A) Transcript-level

\operatorname{PCC}(\overline{\mathbf{q}}_{t}\odot\mathbf{b}_{t,f},\overline{\mathbf{q}}_{t})

. (B)

\log_{2}M_{t,f}^{\mathrm{bias}}

, where zero denotes unchanged total expected counts. Points show medians, thick intervals interquartile ranges, and thin capped intervals the 5th–95th percentiles. Both panels use the 19,283-transcript simulator cohort before fitting. Applying the multiplier is not followed by positional-mean normalization.

Figure 17: Learned versus oracle aggregate mass. Points compare learned and oracle log-mass values at representative separate-depth endpoints and in the joint-depth experiment. Dashed lines mark equality; orange lines are fitted calibration associations. Dense point layers are rasterized inside the vector PDF; axes, labels, and fitted lines remain vector graphics. Agreement with the oracle mass does not imply mass one.

Dataset name	GSE	Sample accession(s)	Replicates
akichika_2019	GSE122071	SAMN10359332	1
andreev_2015_rep1	GSE55195	SAMN02646670	1
andreev_2015_rep2	GSE55195	SAMN02646671	1
apostolopoulos_2024_Cas13	GSE232383	SAMN35056827, SAMN35056828	2
apostolopoulos_2024_dCas13	GSE232383	SAMN35056823, SAMN35056824	2
barrington_2023	GSE202900	SAMN28209264, SAMN28209265, SAMN28209266	3

Appendix

Table 4: GEO study and BioSample accessions, with sample-profile counts, for the 114 HEK-derived datasets.

Figure 18: Harmonized processing of HEK-derived Ribo-seq datasets. Adapter trimming, read-length filtering, contaminant depletion, alignment, and P-site assignment produce codon-resolution profiles. Dataset-specific library settings are retained; 11 datasets follow the optional UMI extraction and deduplication path.

Figure 19: QC ranking of all 114 HEK-derived datasets. Each stacked bar shows the ten equally weighted component ranks that form the composite QC score. Shorter bars indicate lower composite QC scores.

Component	Implemented raw quantity	Direction	Median [ $Q_{1}$ – $Q_{3}$ ]
Periodicity	Fraction of CDS P-sites in frame 0	Higher	0.62 [0.56–0.72]
CDS enrichment	Fraction of assigned P-sites falling in the CDS	Higher	0.94 [0.93–0.95]
Depth	Base-10 logarithm of the CDS P-site count plus one	Higher	6.81 [6.25–7.23]
Transcript support	Number of transcripts with CDS TPM $>1$	Higher	11,896 [11,487–12,213]
RPF-length center	Absolute deviation from 30 nt of the median sample-level mean RPF length (nt)	Lower	1.78 [0.89–2.87]
RPF-length spread	Median sample-level RPF-length interquartile range (nt)	Lower	1.5 [1.0–3.0]

Appendix

Table 5: Ten equally weighted components of the composite QC score. Direction indicates whether higher or lower raw values yield a better component rank. Values are cohort medians, with first and third quartiles in brackets.

Figure 20: Reliability weights across transcripts and datasets. (A–D) Weights as a function of positive-codon coverage and log-transformed mean count for four example datasets, using normalization statistics calculated from training transcripts. (E) Distribution of weights across 114 datasets, with each dataset contributing equally. The envelope shows the pointwise 10th–90th percentile range of dataset-specific densities; the pale band marks the interquartile range of the combined distribution. The vertical line marks a weight of one.

Panel	Datasets	Study groups	Training transcripts
1	29	20	13,261
2	29	21	13,613
3	28	22	13,595
4	28	22	13,340

Appendix

Table 6: Four panels with independent experimental sources. All models share the same 1,593 validation and 1,593 test transcripts.

Figure 21: Reference weighting and agreement across panels. (A) Distribution of global QC ranks within each panel; lower ranks indicate more favourable measured quality. (B–H) Within-transcript Pearson correlations for the six panel pairs under uniform weights and weights favouring better- or worse-scoring datasets at

p=1,3,5

. The 28 fits comprise four panels evaluated under seven reference-weighting policies, all using the same 714 test transcripts. Boxes show medians and interquartile ranges; whiskers show the 5th–95th percentiles.

Reference weights	Mean PCC ↑	Mean RMSE ↓
Uniform	$0.8713\ [0.8695,0.8731]$	$0.2649\ [0.2638,0.2660]$
Better-scoring, $p=1$	$0.8533\ [0.8513,0.8553]$	$0.3111\ [0.3093,0.3131]$
Worse-scoring, $p=1$	$0.8492\ [0.8472,0.8512]$	$0.2987\ [0.2973,0.3001]$
Better-scoring, $p=3$	$0.7463\ [0.7435,0.7490]$	$0.5622\ [0.5564,0.5682]$
Worse-scoring, $p=3$	$0.7453\ [0.7424,0.7483]$	$0.5295\ [0.5261,0.5328]$
Better-scoring, $p=5$	$0.7007\ [0.6977,0.7036]$	$0.6765\ [0.6701,0.6831]$

Appendix

Table 7: Agreement between shared profiles across panels under different reference-weighting policies. Metrics are averaged over the six panel pairs within each transcript, then across transcripts. Brackets give 95% percentile intervals from 5,000 paired transcript-bootstrap resamples.

Figure 22: Distribution of reference weights as datasets are added. (A) Effective number of reference datasets divided by panel size,

N_{\mathrm{eff}}/N

. (B) Reference-weighted mean global QC rank, with lower ranks indicating more favourable measured quality. Solid lines show best-first collections; dashed lines show worst-first collections. These quantities describe the selected datasets and weights, rather than prediction performance.

Figure 23: Shared-profile stability as datasets are added. (A) Best-first collections. (B) Worst-first collections. Each curve reports mean transcripts-level Pearson correlation with the corresponding

N=2

model, using the same selection direction and reference weighting. Bands show 95% transcript-bootstrap intervals over the same 714 test transcripts. Every curve starts at one because its first point compares a model with itself.

Dataset	Organism / system	Assembly / Annotation / CDS model	RPF lengths (nt)	Profiles
stein_2021	C. elegans N2	WBcel235; Ensembl 115, longest CDS/gene	20–35	2
zhang_2016	E. coli K-12 MG1655	ASM584v2; one CDS/locus tag	20–40	1
stein_2021_yeast	S. cerevisiae BY4741	R64-1-1; longest-CDS model	20–35	2
iwasaki_2014	Human HEK293T	GRCh38; GENCODE v46 MANE Select	26–34	4

Appendix

Table 8: Ribo-seq benchmark datasets across four organisms. Each dataset was reprocessed through the same core workflow using an organism-specific genome and CDS annotation.

Architecture / setting	C. elegans	E. coli	Human	S. cerevisiae	Checkpoint criterion
iXnos native	3,710 (28)	1,010 (24)	2,296 (22)	2,538 (27)	minimum validation MSE
$\hookrightarrow$ matched–U	10,412 (30)	3,307 (30)	11,989 (14)	4,043 (29)	minimum validation MSE
$\hookrightarrow$ matched–W	10,412 (30)	3,307 (16)	11,989 (14)	4,043 (29)	minimum validation MSE
RiboExp native	407 (170)	409 (43)	413 (68)	418 (76)	maximum validation Pearson
$\hookrightarrow$ matched–U	10,412 (7)	3,307 (24)	11,989 (25)	4,043 (47)	maximum validation Pearson
$\hookrightarrow$ matched–W	10,412 (18)	3,307 (21)	11,989 (29)	4,043 (48)	maximum validation Pearson

Appendix

Table 9: Training cohorts and validation-selected checkpoints by architecture and setting. Each organism cell gives the number of training transcripts followed by the selected epoch in parentheses. Native uses the model-specific preprocessing; matched–U and matched–W use the full RiboUnmix cohort and and matched baseline target preparation without or with the reliability weights from Equation 43 . * Human Riboformer matched–U had undefined validation Pearson at every epoch, so it fell back to epoch 1.

	C. elegans			E. coli
Architecture / setting	Pearson	Spearman	RMSE	Pearson	Spearman	RMSE
iXnos native	0.286 $\pm$ 0.020	0.242 $\pm$ 0.020	1.944 $\pm$ 0.271	0.372 $\pm$ 0.030	0.367 $\pm$ 0.027	3.844 $\pm$ 0.698
$\hookrightarrow$ matched–U	0.241 $\pm$ 0.019	0.211 $\pm$ 0.015	2.403 $\pm$ 0.367	0.344 $\pm$ 0.024	0.350 $\pm$ 0.017	4.062 $\pm$ 0.640
$\hookrightarrow$ matched–W	0.271 $\pm$ 0.016	0.237 $\pm$ 0.014	2.048 $\pm$ 0.301	0.350 $\pm$ 0.028	0.353 $\pm$ 0.022	3.882 $\pm$ 0.617
RiboExp native	0.269 $\pm$ 0.022	0.239 $\pm$ 0.017	2.219 $\pm$ 0.371	0.364 $\pm$ 0.026	0.359 $\pm$ 0.021	5.176 $\pm$ 1.401
$\hookrightarrow$ matched–U	0.204 $\pm$ 0.013	0.165 $\pm$ 0.012	2.364 $\pm$ 0.347	0.340 $\pm$ 0.024	0.323 $\pm$ 0.022	3.927 $\pm$ 0.588

Appendix

Table 10: Benchmark performance across four organisms and training settings. Entries report median transcript-level metrics

\pm

half the width of their 95% percentile bootstrap with 10,000 transcript resamples. All metrics compare predicted and observed profiles at positions with positive observed counts after excluding the first and last five codons of each CDS. For normalized RMSE, target and prediction are first divided by the target mean over the complete trimmed window, including zero-count positions For the baselines, native denotes model-specific preprocessing and training, whereas matched–U and matched–W use a common training cohort and target preparation with uniform and reliability-based transcript weights, respectively. Dashes indicate undefined correlations caused by constant test predictions for human Riboformer matched–U. The best point estimate in each organism–metric column is bold and underlined (maximum for correlations; minimum for RMSE). Bold without underlining marks another point estimate that falls inside the best entry’s reported

\widehat{m}_{\mathrm{best}}\pm\Delta_{95,\mathrm{best}}

span. This visual rule is descriptive and is not a paired significance test. See Section I.4 for details.

Explore similar work

Jun 15, 2026q-bio.QM

MultiMolecule: a modular ecosystem for biomolecular sequence-model workflows

Biomolecular sequence models are increasingly reused outside the studies in which they were introduced, but public checkpoints rarely preserve the execution context needed to inspect source-defined behavior, adapt models to new assays, compare models under shared task definitions or deploy biological predictions. MultiMolecule is an open-source Python ecosystem that turns heterogeneous RNA, DNA and protein sequence-model releases into complete, source-checked model-family implementations with shared loading, workflow and prediction interfaces. The Resource state reported here includes 53 complete model-family implementations with 112 standardized model checkpoints, together with 16 curated dataset resources released through 39 public dataset repositories and 10 user-facing prediction pipelines. Standardized components are linked to source provenance, conversion or preparation code, source-reference checks, Extended Data summaries and public documentation, allowing users to inspect what was standardized, what behavior was checked and how each component enters training, evaluation, inference or deployment. By shifting reuse from repository-specific checkpoints to executable implementations connected to standardized checkpoints, curated datasets, Runner workflows and biological prediction pipelines, MultiMolecule provides common infrastructure for preserving source-defined model behavior, adapting models to new assays, enabling controlled evaluation and deploying biomolecular predictions.

Zhiyuan Chen

1*DanLing Team.

Bioinformatics Intermediate Checkpoints Modular Framework Agentic Workflow Design Open-Source

Ask PDF

May 7, 2026q-bio.GN

OmicsLM: A Multimodal Large Language Model for Multi-Sample Omics Reasoning

Interpreting transcriptomic data is one of the most common analytical tasks in modern biology. Yet most current models either consume expression profiles without producing natural-language biological explanations, or reason in language without direct access to quantitative omics measurements. We introduce OmicsLM, a multimodal LLM that connects quantitative omics profiles with natural-language biological tasks. OmicsLM represents each transcriptomic profile as a compact continuous representation within the LLM context. This interface preserves quantitative expression signal while allowing natural-language instructions, explicit gene mentions, and multiple interleaved biological samples to be processed together in one model context. We train OmicsLM on more than 5.5 million instruction-following examples spanning over 70 task types, combining continuous transcriptomic inputs, experimental data rendered through diverse language templates, and free-text biological knowledge and question-answering data. This mixture covers cell type annotation, perturbation prediction, clinical prediction, pathway reasoning, and open-ended biological question answering. Existing benchmarks evaluate either profile-level prediction or text-only biological QA, leaving language-guided, multi-sample reasoning over real expression profiles unmeasured. To close this gap, we introduce GEO-OmicsQA, a benchmark for multi-sample biological question answering built from real Gene Expression Omnibus (GEO) studies. We demonstrate that OmicsLM can use expression profiles directly and perform comparably to specialized omics models on profile-level tasks, while outperforming both omics-specialized models and general LLMs on language-guided biological reasoning over expression data.

Maciej Sypetkowski, Joanna Krawczyk, Łukasz Smoliński +4

Ingenix.AI, Warsaw, Poland

Multi-Omic Integration Transcriptomics Biomedical Text Multimodal Large Language Models

Ask PDF

May 12, 2026q-bio.BM

Multimodal Alignment and Preference Optimization for Zero-Shot Conditional RNA Generation

The design of RNA molecules that interact with specific proteins is a critical challenge in experimental and computational biology. Despite recent progress in natural language modeling and deep learning-based protein design, there remains significant room to improve the frequency of successful interactions and the authenticity of generated sequences for functional applications. In this work, we frame conditional RNA sequence generation as a multi-stage alignment problem, introducing Moirain: a suite of models optimized via multimodal supervised fine-tuning (SFT) and Direct Preference Optimization (DPO). Our approach begins with large-scale pretraining on diverse RNA corpora to capture the fundamental grammars of sequence plausibility. To achieve target-specific generation, we employ a multimodal SFT architecture that conditions RNA synthesis on protein structural and sequential features. Finally, we leverage DPO to refine the model using synthetic interaction data: taking advantage of DPO's unique ability to navigate non-aligned preference spaces, we improve functional fitness without collapsing the learned natural distribution. Extensive evaluation of the Moirain series (Moirain-Base, -Multi, and -DPO) demonstrates that our framework consistently produces novel, diverse, and biologically plausible RNA sequences with superior binding affinities compared to existing baselines.

Roman Klypa, Alberto Bietti, Sergei Grudinin

Univ. Grenoble Alpes, CNRS, Grenoble INP, LJK · Grenoble, France · Center for Computational Mathematics, Flatiron Institute +1

Messenger Rna Protein Design Multimodal Alignment

Ask PDF