LIVIA — About
Metric definitions, color schemes, and references
Metric definitions, color schemes, and references
LIVIA has no server for your data. Every file you load (structures, PAE and confidence files, lis.py CSVs, bundles) is read and analyzed in your browser and is never uploaded. To add names, domains and reported interactions, some pages look things up at public services. This is everything they send:
| What is sent | To | When |
|---|---|---|
| A checksum of each chain’s sequence (LIVIA’s own index receives only its first three characters) | LIVIA’s sequence index (flyark.github.io) and UniProt | Prediction Analysis and cLIP, to name each chain |
| A chain’s amino-acid sequence | EBI BLAST (ebi.ac.uk) | Prediction Analysis only, when you click Identify by BLAST beside a chain of 25 or more residues that no checksum or name lookup identifies (a construct or a fragment, for example); never automatically |
| UniProt accessions | UniProt, InterPro and AlphaFold DB | Once a chain is named, or for an ID you enter: gene names, domains (UniProt, Pfam, TED), pLDDT and AlphaMissense |
| NCBI Gene IDs | MIST (flyrnai.org) and BioGRID, through LIVIA’s proxy | Reported Interactions, for the named chains |
| Gene names or IDs from your data | MyGene.info and UniProt | cLIP and PPI Network, when names are converted to gene symbols |
| The address of a prediction you ask for | AlphaFold DB; FlyPredictome (flyrnai.org), through LIVIA’s proxies | AFDB Dimer, AFDB Monomer Subdomain, FlyPredictome and Ortholog Interactome |
| A page visit | GoatCounter | Every page: anonymous visit counts, no cookies |
Nothing else leaves your browser: not your structures, not PAE or confidence values, not scores, and not the figures or scripts you download. LIVIA’s proxies (a Cloudflare worker, and a Google Apps Script for ortholog pages) only forward the requests above; the worker caches the public answers it fetches. The worker forwards only to an allow-list of hosts (flyrnai.org, AlphaFold DB, and OSF for prediction bundles that OSF serves without cross-origin headers) and adds the BioGRID access key itself, so the key is never in page code. To have a host added, open an issue on the LIVIA repository; to run your own copy of the proxy, see tools/cloudflare-worker. The pages also load their code libraries (Mol*, Pyodide) from public CDNs. Where a page says “upload” (a FASTA, a structure, a list), it means loading the file into your browser; the file is not sent anywhere.
.cxc script and structure file, place them in the same folder, then open the script in ChimeraX or PyMOLCDK2, TP53)famdb_details_ortho URL directly in the Paste URL tabP69905), a model entity ID (e.g. AF-0000000066214167), or a full AlphaFold DB URLP31749) — autocomplete suggestions appear as you typecLIP takes the output of a screen — one bait against many partners — and clusters the partners by which residues of the bait they contact. Partners that bind the same surface fall into the same cluster, so a large screen resolves into a handful of interaction modes. Everything runs in the browser from lis.py output; no structure files are required.
lis.py output (a CSV or a ZIP of CSVs). Add a FASTA in the same panel for exact sequences (optional) — or click a bundled examplelis.py CSV, the paired FASTA, a self-contained report HTML, or SVG/PNG figureslis.py CSV or a ZIP of CSVs. Drop the lis.py output and a FASTA together; files are sorted by content automaticallylis.py CSV, paired FASTA, self-contained report HTML, and SVG/PNG figuresPPI Network turns an all-by-all screen into an interaction graph and finds communities of densely connected proteins. Community detection runs entirely in the browser using Pyodide and python-igraph — no server, no installation.
lis.py output (a CSV or a ZIP of CSVs); include a FASTA to carry sequences to cLIP (optional) — or try the examplelis.py CSV or a ZIP of CSVs; a FASTA can be included (dropped alongside, or inside the ZIP) to carry sequences to cLIPAll tools include adjustable LIR display settings that affect the visualization script and 3D viewer:
Click Apply after changing values to update the visualization script and 3D viewer.
The analysis tools include a sequence viewer that displays the actual amino acid sequence (one-letter codes) for each chain or domain. Residues are color-coded based on their interaction status:
Residue numbering is shown at intervals. The viewer scrolls horizontally for long sequences. In cLIP the sequence is grouped by ten and stays searchable (⌘F / Ctrl-F), and when no FASTA is loaded it falls back to the UniProt canonical sequence.
All tools use consistent, fixed colormaps for maps and matrices. These are not affected by the color presets.
| Map | Colormap | Scale |
|---|---|---|
| PAE map | blue – white – red (bwr) | 0 Å (confident) → 30 Å (uncertain) |
| LIS map | matplotlib Blues | 0 → 1 (higher = stronger interaction) |
| cLIS map | matplotlib Greens | 0 → 1 (higher = stronger contact-filtered interaction) |
| Score matrix (iLIS) | matplotlib Oranges | Lower-left triangle |
| Score matrix (cLIR / ipTM) | white → green (or Purples for ipTM) | Upper-right triangle |
The visualization script (.cxc for ChimeraX, .pml for PyMOL) uses chain colors that you can customize via presets. The default is Teal/Coral gradient.
The Monomer tool provides additional coloring presets beyond the default domain-colored view:
Color presets and custom colors update the visualization script (ChimeraX/PyMOL) automatically when you select a preset. The PAE/LIS/cLIS maps use fixed colorscales and are not affected by color choices.
| Metric | Full Name | Definition |
|---|---|---|
| iLIS | integrated LIS | √(LIS × cLIS) — geometric mean of LIS and cLIS; balanced score combining confident domain and confident contact |
| LIS | Local Interaction Score | Average of inversely scaled PAE (0–1, higher is better) within LIA |
| cLIS | contact-filtered LIS | Average of inversely scaled PAE within cLIA (contact-filtered) |
| LIR | Local Interaction Residues | Residues in the confident interaction region (PAE ≤ 12 Å) |
| cLIR | contact-filtered LIR | Residues in direct physical contact (PAE ≤ 12 Å & Cβ ≤ 8 Å) |
| LIA | Local Interaction Area | Confident interaction area (PAE ≤ 12 Å) |
| cLIA | contact-filtered LIA | Confident interaction area within contact distance (PAE ≤ 12 Å & Cβ ≤ 8 Å) |
| iLIA | integrated LIA | √(LIA × cLIA) — geometric mean of confident interface area and contact-supported interface area; reflects interface size weighted by contact density |
| iLISA | integrated LIS-area | iLIS × iLIA — composite score combining interface confidence (iLIS) with interface size (iLIA); useful for ranking large vs. small interfaces of comparable confidence |
| pDockQ | predicted DockQ | The original formula (Bryant et al., 2022): interface pLDDT and the number of contacting residue pairs (Cβ ≤ 8 Å), with no PAE |
| LIpDockQ | Local Interaction-derived pDockQ | The pDockQ formula computed only on cLIR pairs, so contacts the PAE does not support no longer count (Kim & Perrimon, 2026) |
| pDockQ2 | predicted DockQ v2 | The original formula (Zhu et al., 2023): each contact of the Cα-based interface weighted by its PAE |
| LIpDockQ2 | Local Interaction-derived pDockQ2 | The pDockQ2 formula computed only on cLIR pairs, so a small confident interface is not diluted by a large, poorly placed contact surface (Kim & Perrimon, 2026) |
Default cutoffs: PAE ≤ 12 Å (confident interaction) and Cβ ≤ 8 Å (direct contact). These can be adjusted before processing in the Universal and Dimer tools.
PAE transformation: For each inter-chain residue pair (i, j), the PAE value is converted to a confidence score: confidence = 1 − (PAE / cutoff) if PAE ≤ cutoff, otherwise 0. This linearly maps PAE = 0 Å to confidence = 1.0 (highest) and PAE = cutoff to 0. The PAE cutoff of 12 Å was determined by ROC analysis to maximize AUC (Kim et al., 2024).
LIS calculation: LIA (Local Interaction Area) is the count of inter-chain residue pairs with PAE ≤ cutoff. LIS is the mean confidence score across all LIA pairs. cLIA further restricts to pairs also in physical contact (Cβ ≤ 8 Å), and cLIS is the mean confidence within cLIA. The inter-chain PAE block is symmetrized by averaging (A→B) and (B→A) directions. iLIS = √(LIS × cLIS), combining interface confidence with direct contact evidence into a single robust metric (Kim et al., 2026). Two derived size-weighted metrics are also reported for completeness: iLIA = √(LIA × cLIA) summarizes interface size, and iLISA = iLIS × iLIA combines interface confidence with size. These secondary metrics may help differentiate predictions with similar iLIS but different interface sizes.
Additional confidence metrics: ipSAE (interaction prediction Score from Aligned Errors; Dunbrack, 2025) uses a PTM-like scoring function on PAE values at contact residues. actifpTM (actual interface pTM; Varga et al., 2025).
pDockQ family: pDockQ (Bryant, Pozzati & Elofsson, 2022) and pDockQ2 (Zhu et al., 2023) are reproduced here exactly as published: pDockQ is a sigmoid on mean interface pLDDT × log(interface contact count), over a purely geometric interface (Cβ–Cβ ≤ 8 Å, Cα for glycine — no PAE at all); pDockQ2 is the same sigmoid form, refit to mean pLDDT × mean PAE-decay over a Cα–Cα ≤ 8 Å interface. Neither formula gates which residue pairs count as "interface" by PAE — a pair the model happened to place close together, without genuine confidence in their relative orientation, still counts toward pDockQ, and is only down-weighted (not excluded) in pDockQ2. LIpDockQ and LIpDockQ2 apply the identical published formulas to this file's own cLIR interface (PAE ≤ cutoff and physical contact) instead of the original pure-distance one, so a geometrically-close but PAE-unconfident pair is excluded from the interface entirely, rather than merely down-weighted. (Note: despite the "LI" prefix, LIpDockQ/LIpDockQ2 are built on the contact-filtered cLIR interface, not the PAE-only LIR — the LI/c naming convention used elsewhere on this page does not apply here.)
pDockQ availability: pDockQ2/LIpDockQ2 need the raw numeric PAE matrix, so they are not computable from a lightweight bundle, a FlyPredictome/Ortholog Interactome partner-search result, or a cLIP CSV that doesn't already carry them — those views show pDockQ/LIpDockQ only, computed from whichever structure (and, for FlyPredictome/Ortholog, interaction residues) is actually available. On the Dimer tool and the FlyPredictome/Ortholog pair-detail views, all four are computed live once a specific pair is selected and its structure is fetched.
Contact map: Built from Cβ atom distances (Cα for glycine). Two residues are in contact if their Cβ–Cβ distance ≤ 8 Å. For nucleic acids, phosphorus (P) atoms are used with a 4 Å distance adjustment.
Symmetrization: Results for chain pairs (A→B) and (B→A) are averaged. LIR/cLIR residue sets are the union across all models.
Suggested threshold: iLIS ≥ 0.223 was established at 10% FPR using large-scale Y2H reference datasets in yeast (Yu et al., 2008), fly (Tang et al., 2023), and human (Braun et al., 2009). This threshold was determined using AlphaFold-Multimer predictions (via ColabFold). See Kim et al., 2026, for details.
A lightweight bundle shrinks a full prediction — which for a higher-order complex can be tens to hundreds of MB, dominated by the numeric PAE — down to a ~1–3 MB package that LIVIA renders with no loss of any figure: the 3D viewer, chord diagram, contact map, score matrix, and sequence viewer all work exactly like a full load. This works because LIVIA only ever uses the numeric PAE at one point — computing the LIS metrics and interaction residues — and those are already in the scores CSV; the bundle keeps that CSV and trades the numeric PAE for a static PAE image.
In-browser, one click: after loading any prediction, a ↓ Lightweight bundle button appears next to Download CSV at the top of the results — no install, no Python. Adjustable PAE color scale, image size, and font scale before export.
Command line, for many predictions at once: use make_lightweight_bundle.py from the python/ directory of this repository.
Features: keeps every ranked model’s scores in one lis.csv, auto-generates the PAE image, self-validates against the bundle format, and --delete-original only reclaims disk space once a bundle has already passed every check.
For the exact bundle format — useful if you’re hosting your own catalog or building other tooling around it — see the lightweight bundle authoring guide.
Two ways this looks in practice: the FlyPredictome higher-order complex catalog is a standalone, sortable catalog page built entirely from lightweight bundles, while the FlyPredictome Network Explorer links the same bundles contextually — a “3D structure in LIVIA” button per cluster, surfaced from inside an existing interactome browser rather than a separate catalog.
For large-scale batch analysis of many predictions, use lis.py from the AFM-LIS repository. It supports all the same platforms (AlphaFold3, ColabFold, Boltz, Chai-1, OpenFold3, Protenix-v2, ESMFold2), auto-detects the prediction format, and outputs CSV.
Features: .gz/.xz decompression, incremental CSV output (safe to interrupt and resume), progress bar with ETA.
For batch generation of ChimeraX visualization scripts from the lis.py CSV, use lis_to_cxc.py (also in AFM-LIS). It writes one .cxc per fold × rank with LIR (light) and cLIR (dark) per-chain coloring, an iLIS/cLIS label panel, and an optional LIR-only partial PDB for Foldseek-multimer search.
Features: per-chain LIR/cLIR coloring with custom palettes, iLIS-based rank reassignment for AF3/Boltz/Chai/OpenFold3/Protenix, gap-fill + isolated-residue pruning, ChimeraX figure-save tip embedded.
AFM-LIS
Structure Prediction
Confidence Metrics
Y2H Reference Datasets (iLIS Benchmarks)
Visualization
Databases
This Tool
Cite this tool
If you use LIVIA in your research, please cite:
If you use the LIS or iLIS metric, please also cite: