We revisit a 2021 PSO-weighted ensemble for plant-protein subcellular localization: rebuild its voting idea with locally executable predictors, test it on a larger multilabel dataset, and expose every classifier answer before weighting.
A protein's location inside a cell is closely connected to what it does. Experimental localization is valuable but can be slow and costly, so computational predictors are often used to prioritize hypotheses. Those predictors are built from different evidence and can disagree. The research question is therefore not merely “which classifier wins?” but “can their complementary evidence be combined transparently and reliably?”
The 2021 system collected six web-classifier outputs through Selenium and BeautifulSoup RPA, stored them in SQL, and used PSO to learn a classifier-by-localization weight matrix. In later operation, remote queues and changing web interfaces could make one request take more than five minutes.
This lab replaces remote RPA with isolated local Docker predictors. Raw results are cached, aligned to a shared localization vocabulary, and displayed before coefficients are applied. The aim is lower operational latency, reproducibility, and an audit trail—not to claim ownership of the borrowed base models.
We rerun the idea on 5,116 Viridiplantae proteins with both multilabel evaluation and a primary-location experiment faithful to the 2021 decision rule. Adding ngLOC slightly improves the paper-compatible TVPSO result, but harms the multilabel voter, showing why every variant must be validated rather than assumed better.
What this page contributes. It is a reproducible continuation and systems study: local execution, explicit score normalization, a larger partitioned dataset, visible evidence, two evaluation formulations, confusion matrices, and preserved model artifacts. It is not a new claim that PSO, ngLOC, LOCALIZER, WoLF PSORT, Light Attention, or SherLoc2 were invented here.
The present lab stands on two earlier ensemble efforts. Credit is listed explicitly because the software architecture, PSO schedule, and biological framing came from collaborative research rather than from a single implementation session.
Weighted Ensemble for Plant Protein Subcellular Localization Using Particle Swarm Optimization, ECTI-CON 2021.
This work established the plant-classifier RPA workflow, localization-specific PSO weights, result ranking, and the presentation structure reproduced here.
A Consensus Method for Optimizing Multiple Classifier Results for Predicting the Subcellular Localization of Bacterial Proteins, BioMed Research International 2019.
PSO-LocBact documented the time-varying PSO parameters completed in this reproduction: 25 particles, 1,000 objective calls, inertia .9→.4, c1 2.5→.5, and c2 .5→2.5.
Base-predictor creators and citations are credited separately near the end of this page. Their published models supply evidence to the ensemble; their outputs are not presented as models created by the PlantLoc team.
A protein is not necessarily confined to one compartment. It may be transported, retained only temporarily, modified in one location and act in another, or genuinely perform different functions in several compartments. This spatial movement is part of the biology—not merely noise in an annotation table.
Development, stress, signalling and post-translational modification can redirect a plant protein. A single top-1 label can hide this conditional or sequential behaviour.
Multi-location predictions help researchers select organelles, markers and time points for validation instead of treating every non-primary observation as an error.
A primary-location score can be high while missing valid secondary compartments. We therefore report both a 2021-compatible single-label experiment and exact, micro, macro and Hamming multilabel metrics.
Practical consequence. The production panel applies one threshold per localization and can return more than one answer. The pie chart is a distribution of ensemble support; it is not evidence that a protein physically spends that percentage of time in each compartment.
The original method is preserved as a separate paper-compatible experiment; the operational system and the broader multilabel study are deliberately modernized.
| Aspect | 2021 PlantLoc | Current lab |
|---|---|---|
| Classifier execution | Six remote web predictors submitted by RPA | Four local production predictors; ngLOC is tested as a fifth research variant |
| Delivery architecture | Frontend, backend, SQL queue and web crawling | Isolated Docker services, local adapters, result cache and rate limiting |
| Experimental data | Dataset S: 753 proteins | 5,116 experimental-evidence Viridiplantae proteins from DeepLoc partitions |
| Localization formulation | 12 primary-location classes | 10 shared classes with multilabel targets; ngLOC cytoskeleton is retained raw but excluded from the shared vote |
| Optimization | Time-varying PSO maximizing primary-location accuracy | The same TVPSO decision rule is reproduced, plus a separate thresholded multilabel PSO voter |
| Evaluation | 10-fold evaluation on primary labels | Locked tests: 831 single-label proteins and 963 multilabel proteins, with class-level confusion matrices |
| Transparency | Ranked score table and pie chart | Raw answer → coefficient → weighted contribution → threshold/output, weights and confusion matrices |
| Headline result | 92.3% on the original 753-protein setting | 90.5% primary accuracy for five-predictor TVPSO; 75.0% exact multilabel accuracy for production |
The percentages are not directly comparable: the datasets, label spaces, base predictors and evaluation questions differ. The current page shows both formulations so that a systems improvement is not mistaken for a like-for-like accuracy claim.
The live panel retains the four-predictor multilabel voter because adding ngLOC reduced held-out exact accuracy. A separate five-predictor experiment reproduces the 2021 decision rule: time-varying PSO learns one coefficient per classifier and localization, then selects the maximum weighted score as the primary localization.
Production multilabel voter: 3,045 fit / 1,108 validation / 963 test. Paper-compatible TVPSO: 3,348 single-label development / 831 single-label test.
This is live local inference—not a demonstration value. Paste one protein sequence and run the four-predictor voter.
First-time inference can be slow. Identical sequences use the local cache. Public requests are rate-limited.
The evidence table will show raw score and native Yes/No → coefficient → weighted numerator → vote/threshold → final output.
| Model | Macro F1 | Micro F1 | Weighted F1 | Exact accuracy | Hamming loss |
|---|---|---|---|---|---|
| Light Attention | .766 | .839 | .831 | .764 | .034 |
| Production PSO · four predictors | .764 | .833 | .829 | .750 | .038 |
| PSO + ngLOC · five predictors | .674 | .827 | .808 | .734 | .037 |
| ngLOC native top-1 | .500 | .608 | .582 | .562 | .084 |
| Previous PSO · three predictors | .743 | .833 | .833 | .771 | .037 |
| SherLoc2 sequence-only | .327 | .415 | .372 | .307 | .112 |
Adding ngLOC does not improve the multilabel voter. Production therefore stays on the four-predictor model. Exact accuracy requires all ten labels to match.
| Model | Primary accuracy | Macro F1 | Weighted F1 |
|---|---|---|---|
| TVPSO · five predictors including ngLOC | .905 | .840 | .905 |
| TVPSO · four predictors | .901 | .822 | .900 |
| Light Attention | .901 | .820 | .900 |
| ngLOC 1.0 plant | .651 | .537 | .637 |
| WoLF PSORT | .596 | .350 | .573 |
| LOCALIZER | .514 | .181 | .419 |
| SherLoc2 sequence-only | .434 | .357 | .445 |
TVPSO uses 25 particles, 1,000 objective evaluations, inertia .9→.4, c1 2.5→.5, c2 .5→2.5, weights constrained to [0,1], and seed 42. Only exactly-one-label proteins are used because DeepLoc does not supply a separate primary label for multilabel rows.
Per-label F1, presented like Table II of the original paper.
| ngLOC 1.0 | Installed · plant | Official GPL archive compiled with the plant switch; 7,835 accepted training sequences and 5,116 experimental predictions completed locally. |
| SherLoc2 | Installed · sequence-only | Official models run locally without unavailable EpiLoc and optional GO/InterProScan features. |
| Plant-mPLoc | Web only | No standalone model release. It is not included in any accuracy claim. |
ngLOC also predicts cytoskeleton. That raw score is retained, then excluded and the ten shared localization scores are renormalized before voting.
Rows are experimental labels; columns are predicted labels. Darker cells contain more proteins.
Light Attention was developed around DeepLoc. Partition 4 was not used to optimize this voter, but it is not guaranteed unseen by the pretrained base model. This is a PSO development and systems experiment, not an independent Light Attention benchmark.
These are the methods whose executable outputs were actually used.
Weighted Ensemble for Plant Protein Subcellular Localization Using Particle Swarm Optimization, ECTI-CON 2021, pp. 777–780.
The 2026 experiment changes the predictor execution path, dataset partitions, multilabel evaluation and learned coefficients; it does not present the base predictors as our own models.