PARADE

License: PARADE is open source and free for academic and commercial use under an MIT license and may require explicit attribution when utilized. Please refer to the license for full terms.

Proto is not affiliated with University of California, San Francisco and Autosome.org. This toolkit is open source and builds on the implementations produced by these organizations. Product names, logos, and trademarks are the property of their respective owners.

GitHub GitHub Preprint Preprint Cite Cite Tool Source Tool Source Open as Notebook Open as Notebook Run Locally Run Locally

autosome-ru/parade

View repo

A generative framework for enhanced cell-type specificity in rationally designed mRNAs

Matvei Khoroshkin, Arsenii Zinkevich, … Hani Goodarzi

bioRxiv (2024)

Read preprint

@article{khoroshkin2024parade,
  title={A generative framework for enhanced cell-type specificity in rationally designed mRNAs},
  author={Khoroshkin, Matvei and Zinkevich, Arsenii and Aristova, Elizaveta and Yousefi, Hassan and Lee, Sean B. and Mittmann, Tabea and Manegold, Karoline and Penzar, Dmitry and Raleigh, David R. and Kulakovskiy, Ivan V. and Goodarzi, Hani},
  journal={bioRxiv},
  year={2024},
  publisher={Cold Spring Harbor Laboratory},
  doi={10.1101/2024.12.31.630783}
}

Copy citation

evo-design/proto-tools/proto_tools/tools/sequence_scoring/parade

View source

Open Notebook

Open notebook

Run locally with proto-tools

pip install git+https://github.com/evo-design/proto-tools.git

Toolkit contributors

Function	Description
`run_parade_activity()`	Predict cell-type-specific 5’/3’ UTR activity with the PARADE LegNet model (GPU)	Docs Source
`run_parade_gradient()`	Compute differentiable PARADE UTR-activity losses and gradients for relaxed UTR logits (GPU)	Docs Source
`run_parade_stability()`	Predict 3’ UTR mRNA stability (RNA/gDNA log-ratio) with the PARADE LegNet model (GPU)	Docs Source

Background

PARADE (Khoroshkin et al., 2024) is a generative framework for designing UTRs with tailored cell-type-specific activity. Its predictive core adapts the DREAM-challenge LegNet architecture: an EfficientNet-style convolutional network with squeeze-and-excite blocks that reads a one-hot UTR sequence plus a reading-frame positional channel and, for activity, broadcast cell-condition channels. The activity models are trained per construct type — one for 5’ UTRs and one for 3’ UTRs — and condition on a panel of anonymized cell-line codes (c1, c2, c4, c6, c17, and, for 3’ UTRs, c13), returning a predicted activity mass-center for each. A separate 3’ UTR model predicts mRNA stability as an RNA/gDNA log-ratio. The featurization matches the upstream reference pipeline exactly, so predictions reproduce the published values.

Tools

PARADE UTR Activity (`parade-activity`)

Predicts cell-type-specific activity for one or more 5’ or 3’ UTR sequences, returning one value per requested cell code.

API Reference

Source

Input: ParadeSequenceInput

List[string]

required

UTR sequence(s). A single string is normalized to a one-item list. U is mapped to T and N is allowed; mixed lengths are fine (the tool batches per length group).

Source

Config: ParadeActivityConfig

enum

default:"utr5"

Which UTR model to use — "utr5" (5’ UTR) or "utr3" (3’ UTR). Selects the checkpoint and the cell-code panel. Matching the upstream predictor, the model scores the bare insert (no reporter flanks are added).Available options: utr5, utr3

List[string]

PARADE cell codes to return. Empty means the full panel for construct_type. Requested codes must belong to that panel.

integer

default:"0"

Verbosity level (0=quiet, 1=info, 2=debug, 3=raw subprocess stderr). True is coerced to 1 and False to 0.

string

default:"cuda"

Device used for inference.

integer

default:"3600"

Maximum execution time in seconds. None waits indefinitely.

integer

Random seed. When set, tools run reproducibly up to small GPU float noise (see BaseToolOutput.approx_equal), and the seed participates in cache keys. When None, cacheable seed-sensitive tools skip cache until seeded.

string

default:""

Optional override for the pinned checkpoint — a local .ckpt path or an https link (a schemeless host.tld/path is accepted). Caller overrides run on local devices only (rejected on device="cloud"). Empty uses the pinned per-target checkpoint.

integer

default:"8"

Number of sequences to run per GPU batch.

Source

Output: ParadeActivityOutput

List[ParadeActivityResult]

Per-sequence PARADE predictions. The only stored field, so the output survives the framework’s iterable-cache reconstruction (which preserves only the iterable field).

Show ParadeActivityResult

string

required

UTR sequence that was scored (DNA alphabet).

integer

required

Length of the scored sequence.

enum

required

UTR model that produced these scores. Carried per-result so the provenance survives iterable-cache reconstruction.

ParadeActivityMetrics

required

Predicted activity keyed by cell code.

Metrics (one set per results item)

Metric	Type	Range	Availability
`c1`	float	unbounded	when requested
`c2`	float	unbounded	when requested
`c4`	float	unbounded	when requested
`c6`	float	unbounded	when requested
`c17`	float	unbounded	when requested
`c13`	float	unbounded	when requested

Applications

Use this tool to rank UTR designs by predicted activity, screen candidate UTRs for a target cell line, or quantify the activity differential between cell types for cell-type-specific mRNA design.

Usage Tips

Pick the construct type. Set construct_type to utr5 or utr3; it selects the matching checkpoint and cell-code panel.
Cell codes are panel-specific. c13 exists only for utr3. Leave cell_types empty to return the full panel for the construct type.
Match the training length. Upstream trained the 5’ UTR model on ~50-nt inserts and the 3’ UTR model on ~240-nt (roughly 200–300 nt) inserts; the model accepts any length (adaptive pooling) but predictions are only meaningful near the training regime.
Mixed lengths batch together. Different-length sequences in one call are batched per length group; RNA input (U) is accepted and mapped to T.

PARADE mRNA Stability (`parade-stability`)

Predicts 3’ UTR mRNA stability as an RNA/gDNA log-ratio for one or more sequences; higher is more stable.

API Reference

Source

Input: ParadeSequenceInput

List[string]

required

UTR sequence(s). A single string is normalized to a one-item list. U is mapped to T and N is allowed; mixed lengths are fine (the tool batches per length group).

Source

Config: ParadeCheckpointConfig

string

default:""

Optional override for the pinned upstream checkpoint — a local .ckpt path or an https link (a schemeless host.tld/path is accepted and normalized to https://). A caller override runs on local devices only (rejected on device="cloud", since a checkpoint is an executable pickle). Empty uses the pinned per-target checkpoint, verified against its built-in checksum.

integer

default:"8"

Number of sequences to run per GPU batch.

integer

default:"0"

Verbosity level (0=quiet, 1=info, 2=debug, 3=raw subprocess stderr). True is coerced to 1 and False to 0.

string

default:"cuda"

Device used for inference.

integer

default:"3600"

Maximum execution time in seconds. None waits indefinitely.

integer

Source

Output: ParadeStabilityOutput

List[ParadeStabilityResult]

Per-sequence PARADE stability predictions.

Show ParadeStabilityResult

string

required

3’ UTR sequence that was scored (DNA alphabet).

integer

required

Length of the scored sequence.

ParadeStabilityMetrics

required

Stability metrics; the log_ratio metric (RNA/gDNA log-ratio, higher means more stable) is also exposed via the log_ratio convenience property.

Metrics (one set per results item)

Metric	Type	Range	Availability
`log_ratio`	float	unbounded	always

Applications

Use this tool to rank 3’ UTR designs by predicted mRNA stability or to pair stability with cell-type-specific activity when selecting UTRs for downstream validation.

Usage Tips

Stability has no cell conditioning. The model returns a single log-ratio per sequence.
Use the training length. Upstream trained this stability model on 186-nt sequences (its seqsize); score near that length. Mixed lengths in one call are batched per length group.
Higher is more stable. The log_ratio output is directly comparable across candidates.

PARADE UTR Activity Gradient (`parade-gradient`)

Computes a weighted differentiable UTR-activity objective and, by default, returns the gradient with respect to batched relaxed UTR logits.

API Reference

Source

Input: ParadeGradientInput

List[array]

required

Batched relaxed UTR logits with shape (B, L, 4) in A,C,G,T order. Use B=1 for a single design candidate.

number

default:"1.0"

Softmax temperature used to relax logits.

Source

Config: ParadeGradientConfig

enum

default:"utr5"

UTR model to use — "utr5" or "utr3".Available options: utr5, utr3

List[ParadeGradientLossTerm]

Per-cell objective terms summed into one scalar loss.

Show ParadeGradientLossTerm

enum

default:"c2"

PARADE cell code to optimize; must be in the configured construct_type panel.Available options: c1, c2, c4, c6, c17, c13

enum

default:"max"

"max" minimizes 1 - sigmoid((raw - center) / scale); "min" minimizes sigmoid((raw - center) / scale).Available options: max, min

number

default:"1.0"

Non-negative scalar applied before terms are summed.

number

default:"2.0"

Raw activity value where the sigmoid is 0.5.

number

default:"1.0"

Positive scale for the raw activity transform.

string

default:""

number

default:"1.0"

Blend hard argmax one-hot (0) to softmax probabilities (1).

number

default:"0.0"

Straight-through hard-forward coefficient.

boolean

default:"True"

Run backward pass and return gradient.

integer

default:"0"

Verbosity level (0=quiet, 1=info, 2=debug, 3=raw subprocess stderr). True is coerced to 1 and False to 0.

string

default:"cuda"

Device used for inference and backpropagation.

integer

default:"3600"

Maximum execution time in seconds. None waits indefinitely.

integer

Source

Output: ParadeGradientOutput

List[ParadeGradientSampleMetrics]

Per-sample metric containers with scalar loss and raw per-cell activity.

Show ParadeGradientSampleMetrics

List[Dict[string, any]]

Per-objective-term metadata, including direction, weight, sigmoid transform, and weighted score.

string

Name of the metric that best summarizes the result overall (e.g. "avg_plddt" for AlphaFold2). Used by downstream UI and reporting to pick a headline value.

string

Concrete Metrics subclass tag; enables typed reconstruction after a serialization round-trip.

array

Gradient tensor matching input UTR logits, or None when compute_gradient=False.

number

required

Sum of per-sample weighted scalar objective values. Per-sample values are available in sample_metrics.

Dict[string, any]

Auxiliary metadata bundle from the standalone worker, including raw scores, objective-term metadata, and relaxation parameters.

List[string]

required

DNA column ordering for logits and gradient.

Metrics

Metric	Type	Range	Availability
`loss`	float	≥ 0.0	always
`c1`	float	unbounded	when requested
`c2`	float	unbounded	when requested
`c4`	float	unbounded	when requested
`c6`	float	unbounded	when requested
`c17`	float	unbounded	when requested
`c13`	float	unbounded	when requested

Applications

Use this tool inside gradient-based UTR design loops (e.g. Fast SeqProp) to maximize activity in an on-target cell line while minimizing it in off-target cell lines. It is designed for optimizer calls rather than final biological validation.

Usage Tips

Logits are batched. Pass logits with shape B x L x 4 in A,C,G,T order; use B=1 for a single candidate.
Terms target cell codes. Each loss term names a cell_type, a direction (max/min), and a weight; all codes must be in the construct_type panel.
Soft/hard mixing controls relaxation. soft=1.0, hard=0.0 is fully soft; increasing hard uses a straight-through hard-forward estimator.

Toolkit Notes

These apply to every PARADE tool in this toolkit (parade-activity, parade-stability, parade-gradient).

Runs on GPU or CPU. The tools load a small PyTorch LegNet checkpoint; a GPU speeds up large batches but is not required.
Weights are provisioned automatically. By default, the standalone worker downloads the published checkpoint from the pinned autosome-ru/parade commit into the managed model cache and verifies its MD5 checksum.
Predictions are faithful to the reference. The vendored PARADE model/data modules are the verbatim upstream bodies (with only a provenance/Ruff header added per file), so the published checkpoints load and score exactly as they do upstream.

Example notebook: See the full working example for a copy-paste-ready walkthrough.

Infrastructure Guides

The following guides cover how to run tools efficiently and at scale.

Tool Persistence

Keep a tool’s model warm across calls instead of reloading it every invocation.

Device Management

How GPUs are allocated to tools and how to target specific devices.

Parallel Execution

Fan a batch of inputs out across multiple GPUs.

Additional Information

References

Khoroshkin, M. et al. A generative framework for enhanced cell-type specificity in rationally designed mRNAs. bioRxiv (2024). DOI: 10.1101/2024.12.31.630783
PARADE repository: autosome-ru/parade
LegNet architecture: Penzar, D. et al. LegNet: a best-in-class deep learning model for short DNA regulatory regions. Bioinformatics 39 (2023). DOI: 10.1093/bioinformatics/btad457

Tools

Guides

Tool Catalog

Concepts

Implementing a Tool

Background

Tools

PARADE UTR Activity (`parade-activity`)

API Reference

Applications

Usage Tips

PARADE mRNA Stability (`parade-stability`)

API Reference

Applications

Usage Tips

PARADE UTR Activity Gradient (`parade-gradient`)

API Reference

Applications

Usage Tips

Toolkit Notes

Infrastructure Guides

Tool Persistence

Device Management

Parallel Execution

Additional Information

​Background

​Tools

​PARADE UTR Activity (parade-activity)

​API Reference

​Applications

​Usage Tips

​PARADE mRNA Stability (parade-stability)

​API Reference

​Applications

​Usage Tips

​PARADE UTR Activity Gradient (parade-gradient)

​API Reference

​Applications

​Usage Tips

​Toolkit Notes

​Infrastructure Guides

Tool Persistence

Device Management

Parallel Execution

​Additional Information

Background

Tools

PARADE UTR Activity (`parade-activity`)

API Reference

Applications

Usage Tips

PARADE mRNA Stability (`parade-stability`)

API Reference

Applications

Usage Tips

PARADE UTR Activity Gradient (`parade-gradient`)

API Reference

Applications

Usage Tips

Toolkit Notes

Infrastructure Guides

Additional Information