PepINVENT

PepINVENT webserver overview

PepINVENT generates non-natural amino acid fragments and optimizes peptide chemistry with reinforcement learning. ProteinIQ runs the published MolecularAI project at commit e976dd1 with its pretrained generative model and supplied predictive model.

Inputs use CHUCKLES, a sequence of amino acid SMILES fragments separated by |. Each ? marks a position to generate. Ring labels and stereochemistry are part of the input. Natural amino acid sequences and FASTA are not converted to CHUCKLES automatically.

Pricing

PepINVENT uses runtime billing at 27 credits per minute, rounded up to a whole credit. The minimum credit reservation is 27 credits; the reservation is a spending limit, not the final charge. The playground shows the rate and reservation before submission, and unused reserved credits are refunded after settlement.

These examples assume a completed run with enough reserved credits. Neither run mode changes the rate. Actual runtime depends on the number of inputs, samples, learning steps and batch size. A job has a one-hour execution limit, with a 3,500-second limit for peptide computation.

Inputs

The configuration's run_type, when supplied, must match the Run mode selector. For reinforcement learning, supply one masked peptide in input_sequence inside the JSON configuration. The sampling CSV is inactive in this mode.

Input files can be uploaded or selected from saved Files. A batch can contain up to 10 configuration inputs, subject to the plan's batch limit. Each configuration starts a separate job. Sampling processes the rows of its CSV within that job.

Each input file and the combined staged submission are limited to 50 MiB.

The pretrained model, sampling CSV and output locations are managed automatically, including model_path, input_sequences_path, results_output and RL logging paths. The predictive objective uses the supplied predictive_model.pckl and feature_scalar.pckl; arbitrary model files are not accepted. Its model_path and scalar_path may be omitted or use those exact filenames. Submitted input bytes and the effective configuration are included in the downloads.

Settings

All other scientific parameters belong in the Native run configuration JSON. Omitted parameters keep the published program's defaults.

Sampling configuration

The published program uses multinomial sampling. It provides no temperature or random-seed configuration setting. The sampling example uses 10 samples to provide a small starting run; omitting num_samples retains the native default of 1,000.

Reinforcement configuration

Learning settings

These parameters belong inside learning_configuration. An empty object {} keeps all five defaults.

Scoring configuration

These fields belong inside scoring_function.

Each object in scoring_components has the following fields.

substructure_match is a separate multiplier on the aggregate when exactly one such component is configured; its weight is not part of the weighted mean. Both aggregations need at least one other objective with a nonzero total weight.

Score transformations

An optional transformation object inside specific_parameters selects transformation_type. Omitting it or using {} keeps no_transformation. Substructure matching keeps its native match score and does not apply the configured transformation.

Sigmoid descriptions assume low < high and positive slope coefficients. Supplied parameter values determine the transformation's shape.

Custom interpolation parameters belong in the same transformation object.

The published configuration examples show complete objective and transformation objects.

Diversity filter

These parameters belong inside diversity_filter. Omitting this object keeps the defaults below.

CamSol scoring is absent from this published implementation. Its CRBP example cannot run unchanged and is rejected. The hosted modes use pretrained models; supervised pretraining and unavailable internal training data are not included.

Outputs

Tables preserve native column and row order. Sampling groups repeated identical inputs under the same native input key. Neither mode adds a best-candidate ranking. RL does not export an optimized model checkpoint.

Understanding results

In sampling, Generated_smi_* values are filler fragments, not reconstructed complete peptides. NLLs retains the native list of negative log likelihoods as text.

In RL, results.csv contains valid peptides entered into diversity memory at the configured score threshold, with each converted SMILES stored once. It does not list every sample from every step. CHUCKLES contains the filled peptide in the program's chemical representation and SMILES contains its RDKit-converted representation.

raw_* columns retain component values before transformation. Component columns contain the scores used by the aggregation; substructure matching retains its native match score. total_score is the score recorded when the peptide enters diversity memory. The predictive model's probability is not a physical permeability coefficient. Computational scores and generated chemistry require independent scientific and experimental validation.