NaviKinase 1.0

A from-scratch, decoder-only protein language model for the phosphotransferase superfamily (EC 2.7.-: protein kinases plus sugar/lipid/nucleotide kinases), trained entirely locally on Apple Silicon via MLX — no cloud compute, no fine-tuning of an existing model.

Model details

Training data

• Scope: UniProtKB entries with EC 2.7.- (phosphotransferase superfamily), length 30-2000 residues, non-fragment.
• Deduplicated at the UniRef90 cluster level (one representative sequence per cluster) to prevent overfitting to near-identical bacterial strain variants; train/val/test split at the cluster level to keep held-out evaluation meaningful.
• 2,450,504 training sequences, 1,320,176,103 training tokens.
• Each example is packed as <BOS> <EC-tag> residues <EOS>, where the EC-tag conditions generation on one of 14 EC 2.7.- sub-subclasses (e.g. <EC_2.7.11> for protein-serine/threonine kinases) or <EC_UNKNOWN> for sequences without a recognized sub-subclass (~3.3% of the corpus).

Training

• AdamW (betas 0.9/0.95, weight decay 0.1), cosine LR schedule with warmup, peak LR 3e-4.
• Effective batch size 64 sequences/step (batch 4 × grad-accum 16), 1024-token context.
• Ran the full 40,289-step ceiling (2 epochs) to completion — no early stopping triggered (patience=5 eval intervals, val loss kept improving every eval).
• Trained unattended over several days on a MacBook, with automatic checkpointing, battery-aware pause/resume, and crash recovery via launchd.

Evaluation

The test-set number is the one to cite — it was never touched during training or checkpoint selection (validation was used for the latter, so it carries a small optimistic bias).

Motif presence (qualitative, not a hard benchmark)

50 unconditioned samples were scanned for canonical kinase catalytic motifs (glycine-rich P-loop, catalytic HRD, activation-loop DFG). These are protein-kinase-specific motifs, so hit rates are naturally lower on unconditioned samples than on samples conditioned on a kinase EC tag (e.g. --ec-class 2.7.11):

Usage

This model speaks amino acid sequences (FASTA-style), not English — there is no chat template or instruction-following. Three ways to use it:

Generate from scratch, or continue a real sequence:

pip install mlx
python generate.py --n 10 --max-len 400
python generate.py --n 5 --ec-class 2.7.11 --temperature 0.8       # protein-Ser/Thr kinases
python generate.py --n 3 --prefix MKTAYIAKQRQISFVKSHFSRQ            # continue a real sequence

Score how well a real sequence fits what the model learned (teacher-forced loss/perplexity over your sequence — a plausibility signal, not a classifier or structural prediction):

python score.py --sequence MKTAYIAKQRQISFVKSHFSRQ...
python score.py --fasta my_protein.fasta --ec-class 2.7.11
python score.py --sequence MKTAY... --per-position                  # flag atypical stretches
python score.py --sequence MKTAY... --classify                      # EXPERIMENTAL, see script docstring

Plain-English front end (pattern matching over the tools above — not a language model; NaviKinase itself has zero English understanding. See chat.py's docstring for exactly which request phrasings it recognizes, and what it does when a request doesn't match one). In the interactive REPL, it asks a clarifying question rather than guessing whenever the answer would actually change the output — an ambiguous "kinase" with no subtype specified when generating from scratch, or multiple UniProt name-search matches — but proceeds directly when a real sequence (pasted or fetched) already anchors the result, since asking there would just be friction. In single-shot --query mode there's no one to ask, so it reports what it needs instead of guessing:

python chat.py
python chat.py --query "how plausible is this sequence: MKTAYIAKQRQISFVKSHFSRQ"
python chat.py --query "generate a serine/threonine kinase"
python chat.py --query "fetch protein P00519 and continue it"        # live UniProt lookup
python chat.py --query "search pubmed for kinase inhibitor resistance"
python chat.py --query "research protein P00519"                     # multi-step, see below

Multi-step research pipeline (research.py) — "research protein X" chains fetch → score → EXPERIMENTAL classify → motif scan → PubMed search into one synthesized report, autonomously (you don't specify each step). The step sequence is fixed, not dynamically planned — this is the one useful compound goal for this model's narrow domain. Real example (human ABL1, P00519): correctly fetched, scored as typical (perplexity 1.64), the EXPERIMENTAL classifier's top pick (EC 2.7.10, tyrosine kinases) matches ABL1's real annotation, and the motif scan found all three canonical kinase motifs — a genuine, coherent result, though the classify step remains an unvalidated proxy same as everywhere else in this repo.

Live lookups (lookup.py) — unlike everything else here, these hit real, public APIs (UniProt, NCBI PubMed) and need internet access:

• fetch_uniprot_by_accession(accession) — real protein sequence by exact accession, unambiguous.
• fetch_uniprot_candidates(query, n) — best-effort free-text name/keyword search, returns up to n real candidates (accession, name, organism, sequence) for disambiguation — verified against the live API to not reliably rank the intended protein first for ambiguous multi-word queries, which is why callers should show candidates rather than assume the top hit.
• fetch_pubmed_abstracts(query, n) — real citations/abstracts for a keyword search. Informational only — displayed as plain text for you to read; NaviKinase cannot process or reason about this text, it is never passed into the model.

from model import KinaseGPT, ModelConfig
from tokenizer import BOS_ID, EOS_ID, ec_field_to_tag_id, decode
import mlx.core as mx

cfg = ModelConfig.from_json("config.json")
model = KinaseGPT(cfg)
model.load_weights("model.safetensors")
model.eval()

Real 3D structure prediction (structure.py / fold.py) — a genuinely different model, not NaviKinase. NaviKinase was trained only on 1D sequence text and has no representation for 3D coordinates at all; fold.py calls ESMFold (Meta's public structure-prediction API, api.esmatlas.com) on a sequence — one NaviKinase generated, or a real one fetched from UniProt — and returns what that service predicts. Needs internet access. Confirmed by live testing: this free service is real but genuinely flaky (a request can 504 and succeed on retry seconds later, independent of sequence length), so fold_sequence() retries with backoff by default — this is required for reliable use, not optional robustness:

python fold.py --sequence MKTAYIAKQRQISFVKSHFSRQ... --out structure.pdb
python fold.py --fasta my_protein.fasta --out structure.pdb
python chat.py --query "fold this sequence: MKTAYIAKQRQISFVKSHFSRQ..."

Output includes a mean per-atom confidence score (0-1 scale, ESMFold's reuse of the PDB B-factor column for a pLDDT-style estimate — higher is more confident). The resulting PDB structure data is Meta Platforms, Inc., licensed CC-BY-4.0 (see the REMARK lines in any returned PDB file) — separate from and unrelated to this repo's own CC BY-NC 4.0 license on the NaviKinase model itself; only the model weights/code here are non-commercial-restricted, not ESMFold's output.

Files

• model.safetensors — model weights (fp32)
• model.py — standalone MLX architecture definition
• tokenizer.py — frozen tokenizer/vocab for this specific checkpoint
• config.json — architecture hyperparameters
• generate.py — generation / sequence-continuation script
• score.py — plausibility scoring for a real sequence
• motifs.py — regex catalytic-motif scan (P-loop, HRD, DFG)
• chat.py — plain-English pattern-matching front end over all the tools here
• nlu.py — shared request-parsing/clarification helpers (used by chat.py and research.py)
• research.py — multi-step autonomous "research protein X" pipeline
• lookup.py — live UniProt sequence fetch + PubMed citation search (needs internet)
• structure.py — real 3D structure prediction via ESMFold (needs internet; a different model, not NaviKinase)
• fold.py — standalone CLI for structure.py

Important: use the bundled tokenizer.py, not a shared/updated one

This repo ships its own frozen copy of the tokenizer. If you've seen a later, broader version of this project (e.g. a follow-up model covering the full EC 2.- superfamily), do not use that tokenizer with this checkpoint — its EC-tag vocabulary was extended and <EC_UNKNOWN>'s token ID shifted as a result. The tokenizer.py in this repo is the exact, correct vocabulary this model was trained with (<EC_UNKNOWN> = token ID 44).

Responsible use / dual-use disclosure

This model generates protein sequences — real biomolecules, not text. Anyone using generated output for wet-lab work bears full responsibility for independently reviewing it before synthesis.

No built-in screening: this repo has no content moderation, no toxin/hazard blocklist, and no automated biosecurity screening on generated or scored sequences. It has not been reviewed by a biosecurity expert. Every fresh generation from generate.py, chat.py, or the webapp carries an automatic non-synthesis notice for this reason — that notice is the extent of the safeguard, not a substitute for real review.

Why the realistic risk is limited (not eliminated): the model's entire vocabulary and training scope is EC 2.7.- phosphotransferases (kinases) — one of the most common, well-characterized enzyme families in biology, present in nearly all organisms and not associated with toxicity or pathogenicity. Its vocabulary is closed and EC-tag-conditioned, not free-form English prompting, so there is no mechanism to prompt it into generating a different, higher-risk protein class. That narrowness is a real, structural constraint — but it was a consequence of data-availability and model-size decisions (see Training data above), not an engineered safety feature, so don't treat scope alone as a safety guarantee.

Before synthesizing any generated sequence:

• Treat it as an unvalidated hypothesis, not a confirmed real protein — see Evaluation above for exactly what has and hasn't been measured.
• Route synthesis through a DNA synthesis provider that performs standard biosecurity screening (most reputable providers already do this as a matter of course).
• Independently review function and safety yourself; nothing in this repo screens for toxicity, pathogenicity, or hazard.

Limitations

• Scope is EC 2.7.- specifically — it has no exposure to other transferase classes, non-enzymatic proteins, or non-protein biomolecules.
• Character-level, single-sequence model: no structural, multiple-sequence- alignment, or homology information is used, only raw sequence statistics.
• ~3.3% of training data has no recognized EC sub-subclass and was trained under a shared <EC_UNKNOWN> tag.
• Motif-presence checks are a cheap directional signal, not a validated functional or structural correctness benchmark.
• chat.py is keyword/regex pattern matching, not a language model — it only recognizes the request styles listed in its docstring, and says so explicitly rather than guessing when a request doesn't match. It does not give NaviKinase any actual English understanding.
• score.py --classify is an improvised proxy (comparing loss under each EC tag) — the model was never trained as a classifier, so treat its output as a rough signal, not a validated prediction.
• lookup.py's UniProt name/keyword search (as opposed to exact accession lookup) is best-effort free-text ranking, not guaranteed to return the specific protein you meant for ambiguous or multi-word queries — always check the accession it reports. fetch_pubmed_abstracts results are shown to you verbatim and are never interpreted by the model. Both require internet access, unlike the rest of this release, which is fully offline.
• research.py's pipeline runs a fixed sequence of steps (fetch, score, classify, motif scan, literature search) — it is not a dynamic planner that decides what to do based on reasoning about your request. The classify step inherits the same "experimental, unvalidated" caveat as score.py --classify above.
• structure.py/fold.py call ESMFold, a separate third-party model — not NaviKinase, and not covered by this repo's evaluation numbers above. It's a free public service with no uptime guarantee and no length limit enforced here; large sequences may be slow or fail independent of NaviKinase itself. Requires internet access.

License

CC BY-NC 4.0 (Attribution-NonCommercial). Free to use, share, and adapt for non-commercial purposes with attribution. Commercial use — including any use intended to generate revenue — requires prior written permission from the author. See LICENSE for full terms, or contact jaydevkbhatt100@gmail.com for commercial licensing inquiries.