ChordexBio UPLB Agrinnovation 2026 Back to home
Research Deck · ChordexBio at UPLB Agrinnovation Marketplace 2026
DNA Encoding · Machine Learning · Bioinformatics

Encoding DNA for Machine Learning

Marvin De los Santos — ChordexBio TIPs-VF UPLB Agrinnovation Marketplace 2026
T
TIPs-VFDNA encoding for ML · ChordexBio

Encoding DNA for Machine Learning

TIPs-VF: Translator-Interpreter Pre-seeding for Variable-length Fragments — an augmented DNA sequence encoding framework for ML applications in bioinformatics, genomics, and beyond.

Presenter
Marvin De los Santos — ChordexBio
Event
Agrinnovation 2026 · UPLB
Status
biorXiv preprint · 2025
TIPs-VF
Augmented DNA encoding framework
Variable-length
Fragments handled natively
2 publications
TIPs-VF + Covary bioRxiv/preprint
Substrate
Underlying Covary engine
TIPs-VF · Agrinnovation 2026
1 / 6
T
TIPs-VFThe problem

DNA sequences are data. But not all encodings are ML-ready.

DNA and protein sequences are, at heart, data — ordered strings over a small alphabet. For machine learning to work on them, the encoding must preserve biological signal while being computationally tractable. That is a harder problem than it looks.

Challenge 1 What most encodings get wrong
  • Fixed-length windows chop variable-length sequences at arbitrary boundaries — losing signal that lives across fragment boundaries
  • k-mer counts lose positional context — the same k-mer composition means different things in different arrangements
  • One-hot encodings are sparse, high-dimensional, and carry no notion of biological proximity
  • Frequency-based approaches lose translation-frame and directional information
  • Many encodings are designed for one downstream task — not a reusable substrate
Cost to ML Why it matters

A bad encoding forces the model to waste capacity re-learning structure the encoding threw away. A good encoding surfaces the structure the biology already contains — so the model can spend its capacity on the question, not on reconstructing the signal.

The TIPs-VF premise. The encoding is not a preprocessing step to minimize — it is the substrate the model learns from. The encoding design is, in that sense, half the model.
TIPs-VF · Agrinnovation 2026
2 / 6
T
TIPs-VFWhat the name means

Translator-Interpreter Pre-seeding for Variable-length Fragments.

The name is compressed but not accidental. Each word is a design claim about how the encoding is built.

Translator Translating biology into a computable form

The encoding translates sequence — the raw language of DNA — into a representation a machine-learning model can actually train on. It is a deliberate mapping, not a throwaway step.

Interpreter A representation the model can read

The encoded form is designed to be interpretable by downstream models — not an opaque hash. The representation carries the structure a model needs to make use of.

Pre-seeding Pre-conditioning the representation

The encoding is pre-seeded with biological priors — translation frame awareness, directional reading, codon-level context — so the model starts with signal rather than having to re-derive it from scratch.

Variable-length Fragments Built for fragments of any length

Whole genomes, gene regions, amplicons, reads, fragments — TIPs-VF handles variable-length inputs natively rather than forcing everything into a fixed window. This is what makes it a usable substrate across use cases.

A name as a spec. TIPs-VF is not just a label. It is a compact statement of the encoding assumptions: translation-aware, interpretable, pre-seeded with biology, and fragment-length-agnostic.
TIPs-VF · Agrinnovation 2026
3 / 6
T
TIPs-VFThe framework

How TIPs-VF encodes a sequence.

TIPs-VF represents genetic sequences by relative proximity, directional alignment, and translation awareness — in a way that is agnostic to where one fragment ends and another begins.

Relative proximity Positions in context

Rather than treating each token independently, TIPs-VF encodes how sequence positions relate to one another — the relative arrangement of bases carries biological meaning that position-independent encodings lose.

Directional alignment Reading direction matters

Sequence is directional — 5' to 3', sense to antisense, forward to reverse complement. TIPs-VF preserves directional information rather than collapsing it into a bag of tokens.

Translation awareness Codon-bound context

In coding regions, the triplet reading frame carries meaning that single-base encodings miss. TIPs-VF incorporates codon-bound context for biologically meaningful sequence comparison, working in both coding and non-coding sequence.

Variable-length No arbitrary chopping

TIPs-VF is designed to handle variable-length fragments without imposing a fixed window that destroys boundary-spanning signal. This is what lets the same encoding substrate serve phylogenomics, pathogen surveillance, and sequence-based analysis at scale.

Agnostic
The encoding does not assume any one downstream task — it is a transferable substrate meant to serve whatever the model is trained to do.
Borderless
No hard per-organism or per-task boundaries — the representation is meant to generalize, not overfit to one clade or one question.
TIPs-VF · Agrinnovation 2026
4 / 6
T
TIPs-VFValidation · 4 use cases

Validated across classification, identification, relationship, and prediction.

TIPs-VF was published as the encoding foundation underlying Covary. It has been validated across four use cases — each compared against gold-standard alignment-based pipelines.

01

Classification

Taxonomic classification using Staphylococcus taiwanensis 16S rRNA, n=101.

Result: Covary (powered by TIPs-VF) generated similar phylogenetic clustering as ETE3, delineating the outgroup from subsequences.TIPs-VF · ETE3 · FastTree 2
02

Identification

Sequence identification of Staphylococcus spp. and SARS-CoV-2 unknowns, n=94.

Result: Both Covary (powered by TIPs-VF) and ETE3 identified and placed unknown sequences in correct clades.TIPs-VF · ETE3 · MAFFT · IQ-TREE
03

Relationship

Mutational relationship context in TP53, GDC-TCGA HNC, n=220.

Result: Covary provided better mutational relationship context, clustering wildtypes distinctly from variants — which ETE3 failed to demonstrate.TIPs-VF · ETE3 · MAFFT · IQ-TREE
04

Prediction

Genome integrity inference from TP53 mutational markers, n=220.

Result: Covary inferred mutational distances and provided predictive genome-integrity outcomes based on TP53 markers.TIPs-VF
The pattern. On classification and identification, TIPs-VF-based Covary matches gold-standard alignment pipelines — "similar to." On relationship context, it outperforms — "better context than." On prediction, the encoding brings signal the conventional pipeline did not surface. That is the profile of a good encoding: recover what alignment recovers, and surface what alignment leaves buried.
TIPs-VF · Agrinnovation 2026
5 / 6
T
TIPs-VFWhat we're seeking

What we are seeking at Agrinnovation 2026.

TIPs-VF is the encoding substrate behind Covary's alignment-free whole-genome intelligence. It is published, validated, and deployed as part of a working framework. At Agrinnovation we are here to find partners who want to put this substrate to work on problems in agricultural genomics, bioinformatics, and the life sciences.

Encoding

Encoding frameworks for ag-genomics

Partnerships to extend TIPs-VF-style encoding into agricultural sequence problems — crop pathogens, germplasm, metagenomes, and livestock disease agents.

Substrate

ML substrate for sequence tasks

Co-development of machine-learning applications that use the TIPs-VF representation as a reusable substrate — not a one-off pipeline, but a substrate that serves multiple tasks.

Validation

Benchmarks and use cases

More validation — more organisms, more tasks, more comparisons against gold-standard pipelines — to strengthen the evidence base for the encoding approach.

Translation-aware

Codon-context for coding sequence

Extending translation-aware encoding into agricultural coding-sequence problems — where codon context, reading frame, and directional alignment genuinely matter.

Cross-domain

Beyond phylogenetics

TIPs-VF's substrate is not limited to phylogenetics — the same encoding underpins oncogenomics, epidemiology, forensics, and precision medicine in Covary's own roadmap.

Honest note

A published preprint

TIPs-VF is a preprint, not a textbook. The evidence is real and growing, but the framework is young — and we are here at Agrinnovation to find the partnerships that help it mature.

TIPs-VF is the encoding substrate behind Covary's alignment-free whole-genome intelligence. We are at Agrinnovation 2026, UPLB, 19–20 October 2026. Questions and collaboration welcome.
Marvin De los Santos · ChordexBio
Agrinnovation 2026 · UPLB
TIPs-VF · Agrinnovation 2026
6 / 6
TIPs-VF research deck · 6 slides