gitwtfhub

wtf is scireasoner?

spectrai-initiative/scireasoner — explained in plain English

Analysis updated 2026-05-18

20Audience · researcherComplexity · 5/5Setup · hard

TL;DR

A scientific AI model that reasons over the real structure of proteins, molecules, DNA, RNA, and materials, not just text.

Mindmap

mindmap
  root((SciReasoner))
    What it does
      Structure aware reasoning
      Multi domain science model
      Prediction and design
    Tech stack
      Qwen3-14B base
      Foldseek 3Di
      ConfSeq
      SLICES tokens
    Use cases
      Protein function prediction
      Materials property prediction
      Chemical reaction planning
    Audience
      Science researchers
    Limits
      Weights not released yet
      No demo yet
      Research artifact

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Predict properties of molecules, proteins, or materials from their structure.

REASON 2

Reason over protein structure to predict function or binding behavior.

REASON 3

Plan chemical reactions or retrosynthesis routes.

REASON 4

Predict crystal and materials properties for chemistry or physics research.

What's in the stack?

Qwen3-14BFoldseekConfSeqSLICES

How it stacks up

spectrai-initiative/scireasonera-shojaei/constructdrawingaiaaronz345/athena-personal-academic-page
Stars202020
LanguagePythonJavaScript
Setup difficultyhardmoderatemoderate
Complexity5/54/52/5
Audienceresearcherdeveloperresearcher

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1day+

Model weights, demo, and evaluation code are all listed as coming soon, so there is nothing runnable yet.

Wtf does this do

SciReasoner is a scientific AI foundation model built to reason about the structure of proteins, molecules, DNA and RNA, and materials like crystals, rather than just treating them as plain text. Instead of only reading a chemical formula or a protein sequence as a string of characters, it converts these structures into specialized tokens that capture real three dimensional and structural information, then reasons over those tokens to answer questions, make predictions, classify things, and help design new molecules or proteins. The model is built starting from an existing large language model called Qwen3-14B, then adapted with vocabularies specific to science, so it can connect details like amino acid residues, molecular fragments, crystal symmetry, and chemical bonding patterns to real world properties scientists care about. According to the README, it reaches state of the art results on the large majority of tasks it was tested on, and its performance on many specialized tasks matches or beats models built specifically for one narrow scientific domain, based on comparisons published in the accompanying research paper. The model covers a wide scope: chemistry tasks like predicting how well a drug candidate might dissolve or planning chemical reactions, materials science tasks like predicting properties of crystals and alloys, protein tasks like predicting protein stability and function, and genomics tasks like predicting enhancer activity or CRISPR targeting efficiency. Each of these areas has its own table of specific inputs and tasks listed in the README, giving a sense of just how broad the intended use is. As of this README, the actual model weights and a public demo are both marked as coming soon, and evaluation scripts are described as arriving with a future public code release, so this repository currently functions mainly as documentation and a companion to the published research paper rather than something you can immediately download and run. It is aimed squarely at scientific researchers working in chemistry, biology, and materials science, not general software developers.

Yoink these prompts

Prompt 1
Explain how SciReasoner turns protein and molecule structures into evidence tokens instead of plain text.
Prompt 2
Summarize what tasks SciReasoner covers across chemistry, materials, protein, and genomics domains.
Prompt 3
How does SciReasoner compare to specialist models like ESM2 or MolCLR on classification benchmarks?
Prompt 4
Walk me through what Qwen3-14B contributes to SciReasoner's architecture.
Prompt 5
What is still missing from this repo before I could actually run SciReasoner myself?

Frequently asked questions

wtf is scireasoner?

A scientific AI model that reasons over the real structure of proteins, molecules, DNA, RNA, and materials, not just text.

How hard is scireasoner to set up?

Setup difficulty is rated hard, with roughly 1day+ to a first successful run.

Who is scireasoner for?

Mainly researcher.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.