LiteFold

Announcing Rosalind, the most versatile AI Co-Scientist for computational biology and therapeutics research. Giving every biologist their own frontier research lab. Make every experiment count. It's live. Links in the comments.

We would love to have people from pure sciences! Biology is becoming a sub-domain of applied sciences and computing. Meet cracked folkes behind @try_litefold and talk about speed-running drug discovery <3! luma.com/zndszwdp

@benln litefold.ai x.com/try_litefold/s…

I did some tinkering with @cory_jay44 @anindyadeeps at @localhostIND • Amazed at @try_litefold can do right now!

Introducing PeptAI A fleet of autonomous agents for peptide drug discovery. PeptAI removes the handoffs between discovery, synthesis, and wet-lab validation, the parts that slow most pipelines down. How it works: > Runs a 9-gate pipeline (8 computational gates and 1 wet-lab gate) 24/7 > Publishes every gate decision openly and in real time via @Molecule_sci Labs > Automatically pays from its own wallet for wet-lab experiments (SPR validation via @adaptyvbio) when gate milestones are reached > Learns from wet-lab results and refines candidates across multiple runs The fleet: → Agent-01: GLP-1R, metabolic health: 35 candidates advancing → Agent-02: KISS1R, fertility: 2 of 10 advanced to G9 → Agent-03: OX2R, ADHD: lead candidate ready for wet lab → Agent-04: community-selected target: queued The fleet scales with budget. Each agent runs the same 9-gate pipeline against its receptor, 24/7. There are plenty of AI drug discovery pipelines. Most still have humans deciding which candidates advance. PeptAI's agent makes those calls itself, gate by gate. Candidates have to clear 8 computational gates to earn a wet-lab recommendation. Depending on the gate, the candidate is redesigned or discarded. Every decision is published on-chain. PeptAI runs on a stack of open scientific tools and protocols: > BIOS @bioaidevs: knowledge layer for literature search and novelty checks. When an agent needs to know what's been discovered or what the literature says about a target, it calls BIOS. > Litefold @try_litefold: molecular dynamics for conformational stability across G5 > Adaptyv Bio @adaptyvbio: wet-lab synthesis and SPR validation at G9 > Molecule Labs @Molecule_sci: on-chain publication of every gate decision in real time > @base: x402 infrastructure for machine-to-machine payments What you can do with PeptAI: • Follow along here for gate decisions, candidates, and updates as the pipeline runs • Use the open methodology to benchmark your own pipeline against the gates PeptAI is experimental and actively evolving. Gates, tools, and thresholds change as it learns. Discover PeptAI now: app.bio.xyz/agents/peptai

We just designed a covalent inhibitor for drug-resistant lung cancer (EGFR T790M) entirely from computation. Just the binding pocket and LiteFold's de novo engine. Docking score: -9+ kcal/mol. 100ns MD simulation: ligand never leaves the reactive zone. Three hydrogen bonds held the entire trajectory. Warhead stayed within 4A of Cys797 through every frame. At 35ns the ligand shifted and we thought it was drifting. It wasn't. It was induced fit. The complex just found a better equilibrium and sat there for the remaining 65ns. Read more about it below!

Upnext on small molecules!!

Keytruda is the best-selling cancer drug in history. 14B a year. 20 cancer types. We redesigned it. Full in-silico case study with Rosalind at @try_litefold.

We are releasing one of our first case studies with LiteFold's AI Co-Scientist Rosalind. The task: computationally redesign the CDR loops of Pembrolizumab (Keytruda) while holding the antibody framework fixed. Input: PDB 5GGS, the crystal structure of the Pembrolizumab Fab in complex with PD-1. Rosalind parsed the chains, identified the binding interface, and ran a diffusion-based redesign of the heavy chain CDRs. Every variant was co-folded against PD-1 with Boltz-2 and ranked. We observed that, complex iPDE of 0.612 Å on the top design vs 2.42 Å on the wild-type baseline, with pLDDT 0.96 and Protein ipTM 0.911. To be precise, iPDE is a model confidence score, not a binding measurement. What this says is that the structure model is highly confident in the predicted geometry of the designed interface, which makes this variant a high-priority candidate for experimental follow-up. We also ran composable MD (ten 10 ns segments, 100 ns aggregate) showed stable RMSD, expected loop-vs-core RMSF, and steady Rg. Link in the comments

Dropping a blog today

A few years ago, designing an antibody on the computer was extremely difficult. Today, there are several open-source tools which allow anyone to design antibodies from home. Out today: A step-by-step guide to antibody design. By @btnaughton.

Thanks to @NikoMcCarty and the Asimov team for helping me get this out! Hopefully an interesting view into where we are with AI antibody design today. Featuring @boltz_bio BindCraft @ai4s_protenix @AriaxBio @modal @tamarindbio @diffuse_bio @try_litefold @chaidiscovery @nablabio

Onwards and Upwards

4/ The pipeline isn’t a roadmap. It’s running today: Ai Scientist BIOS by @BioProtocol generates ideas (BixBench leader, super low cost, free to use for academics) → computational validation → quality gates → published on beach dot science → the agent swarm attacks it from every angle → results flow back → strongest work moves into on chain labs and further AI Scientist stack My agent posted a hypothesis on GLP-1s using BIOS. Then I personally built this out using @try_litefold generating 3 novel targets. 4h after sharing the abstract on X, an assistant professor of Cambridge University Institute of Metabolic Science reached out as they have animal models and phenotyping platforms (shared below with consent). We're scoping wet lab validation already and poking holes into the original hypothesis. Specifically, whether biasing toward beta-arrestin recruitment over Gs coupling is good

I started to read and understand biology by doing plenty of chatgpt deepresearch. One thing I always missed was visuals. Since biology needs to be visual to be understandable. Today I just realized, I can just do that using Rosalind by @try_litefold. This is a really well written introductory blog on how CryoEM data is being studied. Link of the document in the description.

You can also generate such beautiful blogs in just one or two prompts. Try this out at: app.litefold.ai