capiti

Tiny protein-function classifier for edge deployment. Given a nucleotide sequence encoding a protein, capiti flags whether the encoded protein is expected to retain the enzymatic function of one of a small reference set.

Weighs ~1 MB on disk, runs inference in tens of milliseconds on a Raspberry Pi. Trained by distilling ProteinMPNN's function-preserving design prior into a small 1D CNN. Bundled console scripts cover both batch validation (capiti, capiti-orf, capiti-translate) and real-time streaming during DNA synthesis (capiti-listen, capiti-watch, capiti-interrupt).

Overview (each ResidualDilatedBlock collapsed to one box):

Inside one ResidualDilatedBlock:

See docs/capiti.summary.txt for the full per-layer size / FLOP table.

Install

pip install capiti

Use

capiti ATGCGTAAAGTGGCC...           # prints TRUE or FALSE (default set ab9)
capiti ATGCGT...  --cutoff 0.8 -v   # TRUE  p_inset=0.995
capiti --fasta seqs.fa              # batch over a FASTA
echo ATGCGT... | capiti --stdin

Reference sets

capiti ships three bundled reference sets, selectable at invocation time via --set NAME (or CAPITI_SET).

set	targets	description
ab9	9	Beta-lactamases relevant to antibiotic resistance plus other soluble enzymes. Default.
E	59	Larger enzyme panel (54 PDB + 5 AlphaFold-only entries).
C	235	Broad enzyme panel sourced from PDB.

capiti ATGCGT... --set ab9
capiti --fasta seqs.fa --set C
CAPITI_SET=E capiti --stdin

Inference-time gate

Capiti pairs the CNN with a SIFTS-backed fixed-position gate by default: if the model picks a target Ti and the query has a mutated residue at any of Ti's catalytic / active-site positions, the in-set score is forced to 0. This catches single-residue active-site knockouts the masked-mean CNN under-weights. Disable with --no-gate.

Exit code is 0 on TRUE, 1 on FALSE, suitable for shell pipelines:

capiti ATGCGT... && echo "in set" || echo "not in set"

Benchmarks

On the held-out test split for each set (gate on, prefix-aug c1 weights as shipped in 0.1.2):

set	targets	AUC	PR-AUC	model size
ab9	9	0.998	0.996	0.94 MB
E	59	0.989	0.991	0.96 MB
C	235	0.985	0.985	1.05 MB

Side-by-side comparison with BLAST and k-mer baselines at docs/benchmark/CE_summary.md. Per-set ROC, PR, per-class plots at docs/benchmark/v3/, docs/benchmark/E_v1/, docs/benchmark/C_v1/.

Streaming inference

capiti-watch reads bases as they're produced by a DNA synthesizer, finds the first ATG, translates each new codon in-process, runs the ONNX model after every codon, and pulses an abort GPIO line when the in-set probability holds above a threshold for several consecutive scorings. The bundled CNN was retrained with prefix-truncation augmentation so it gives calibrated probabilities on partial inputs; no architecture change, just a length-aware training distribution.

capiti-watch                                 # default ab9, fires on real synth
capiti-watch --set any                       # multi-set: load ab9+C+E in parallel
capiti-watch --no-interrupt -v               # dry run (print verdicts, no abort)
capiti-watch --sim-nt "ATG..."               # in-silico, no Pi needed
capiti-watch --sim-fasta queries.fa --set any
capiti-watch --threshold 0.99 --stability 10 # stricter "we're sure" profile

Streaming-coverage at FPR<=5%, per-target TPR>=0.9 (held-out):

set	callable by K=200 AA	by K=400	unreachable
ab9	89%	100%	0
E	47%	83%	10/59
C	66%	89%	25/224

In-silico and hardware demo recipes: docs/streaming_demos.md. Background and training-distribution rationale: docs/streaming_prefix_augmentation.md.

Hardware integration (Raspberry Pi)

script	role
capiti-listen	reconstruct nucleotide sequence from amidite + TET strobe pulses on GPIO (batch capture)
capiti-watch	as above, but in-process scoring + abort pulse on stable trigger (streaming)
capiti-interrupt	one-shot abort pulse on GPIO17 (manual / dry-run testing)

Wiring matches across all three (Pi 4, BCM numbering):

Pi pin	Phys	Direction	Role
GPIO5	29	in	A amidite
GPIO6	31	in	G amidite
GPIO13	33	in	C amidite
GPIO19	35	in	T amidite
GPIO26	37	in	TET activator strobe
GPIO22	15	in	DONE (rising edge stops capture)
GPIO17	11	out	Abort pulse to upstream rig

Each input line wants an external 10 kohm pulldown to GND; common GND between Pi and rig required. Install with pip install 'capiti[pi]' to pull in gpiozero + lgpio. Sim modes (--sim-nt, --sim-fasta) work without the GPIO stack and are useful for in-silico validation before deploying to hardware.

Status

Research-grade. The CLI surface (flags, stdin/FASTA behaviour, exit codes) is stable; bundled models may be retrained and updated between 0.x releases. Not for operational use.

License

MIT.