alphafoldonline

AI for Science institute for protein intelligence. Fold a sequence into 3D structure in your browser with Protein Explorer
0x792e12ba0c77c66fc6f3f5bdfbe9623148a

[alphafoldonline @alphafoldonline]

The 400-residue live cap is a real constraint today. The planned answer isn't raising it blindly — it's segmented folding, breaking a longer chain into pieces a live pipeline can actually handle well. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Not everyone wants to paste sequences into a browser — some want to call it from their own pipeline. Programmatic API access for researchers is a planned direction, not a live feature yet. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Every fold you run today is yours to download and forget. The direction we're building toward: a real structure library, so predictions become a growing, searchable resource instead of a one-off result. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

One sequence at a time is the starting point, not the ceiling. Folding a whole batch of sequences in one pass — for real research workloads, not just one-off curiosity — is on the roadmap. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Predicted and simulated are where software can take you alone. Assayed and reproduced need a wet lab. The honest long-term goal is climbing toward that last tier — computational results that get physically verified, not just computed. https://t.co/QK7GZVOj0G

[alphafoldonline @alphafoldonline]

The plan is to release first-party datasets progressively, as real research artifacts actually exist — not a data-dump promise with no substance behind it. If it's not real yet, it's not released yet. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

A single 100-residue protein has roughly 20^100 possible sequences — a number bigger than the atoms in the observable universe. Nature has sampled a sliver of that space. Computational folding is how you explore more of it than a lab ever could alone. https://t.co/abEQeDgDQW

[alphafoldonline @alphafoldonline]

Why we draw a hard line between 'real predicted structure' and 'illustrative visualization': in science communication, blurring the two is how false confidence spreads. We'd rather be boring and accurate than exciting and wrong. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Why cap at 400 residues? Because folding live, in your browser, in seconds, is a real compute and latency budget — not an arbitrary limit. We'd rather state the real constraint than pretend there isn't one. https://t.co/abEQeDgDQW"

[alphafoldonline @alphafoldonline]

The goal was never 'a folding demo.' It's making a piece of real computational biology tooling as easy to open as a search bar — no install, no lab, no gatekeeping between a question and a structure. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

The five-stop color ramp on every structure isn't decoration — warm orange marks low-confidence, flexible regions; deep blue marks high-confidence, well-resolved backbone. Look at the color, and you're reading the model's own uncertainty. https://t.co/abEQeDgDQW

[alphafoldonline @alphafoldonline]

For years, predicting a protein's structure meant searching for evolutionary relatives first — a slow step no tool skipped. Language-model-based folding is a genuine recent shift in the field, and Explorer runs on that shift, not around it. https://t.co/abEQeDgDQW

[alphafoldonline @alphafoldonline]

New compute, new models, new science — all real directions we're building toward, and all explicitly labeled as roadmap, not shipped. We'd rather say 'not yet' than let a research direction quietly get mistaken for a live feature. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

One model's answer is a hypothesis, not a verdict. Part of where this is headed: running a structure through more than one model and comparing, instead of taking a single prediction as ground truth. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Wrapping someone else's model is the first step, not the destination. The direction we're building toward is our own protein-science models — trained and tuned on the specific problems we care about, not just borrowed weights. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

ESMFold isn't the ceiling, it's the starting engine. The architecture is built to swap in newer open-source folding models — and models we develop ourselves — as they're ready, instead of freezing around one model's generation forever. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

Compute behind Explorer is scaling, not fixed. The direction is more headroom for heavier workloads — longer sequences, batch folding, higher-throughput research use. That capacity build-out is in progress; it's a roadmap item, not a promised date. https://t.co/QK7GZVOQQe

[alphafoldonline @alphafoldonline]

The model doesn't just hand you a shape and call it done — it tells you, residue by residue, where it's confident and where it isn't. Model doubt made visible is the difference between an AI demo and an AI research tool. https://t.co/abEQeDgDQW

[alphafoldonline @alphafoldonline]

ESMFold treats a protein sequence the way a language model treats a sentence — learning the statistical patterns of biology from a vast corpus of sequences, then reading structure out of what it learned. That's a genuinely modern AI system at work, not a lookup table.

[alphafoldonline @alphafoldonline]

Before we'd trust a pipeline on a sequence nobody's seen, we run it against sequences everyone has — Trp-cage, ubiquitin, structures with decades of deposited ground truth behind them. Validate against the known before you trust the unknown. https://t.co/abEQeDhbGu