VIDRAFT.
VIDRAFT / Insights / AI for Science
AI for Science

Which AI model is best at predicting a drug's human intestinal absorption (HIA)?

VIDRAFT's structure-only ensemble topped the independent, blind Polaris HIA leaderboard with an AUROC of 0.990 — its 15th first-place finish across public Polaris leaderboards.

Published 2026-07-31About 2min readby VIDRAFT
Quick answer

On the independent, blind Polaris HIA benchmark, VIDRAFT's 'vidraft-hia-molformer-chemberta-ensemble' scored an AUROC of 0.990, edging out AceProp and TabPFNv2 (both 0.989). It predicts intestinal absorption from a molecule's SMILES structure alone.

What is Human Intestinal Absorption (HIA) in drug discovery?

HIA predicts whether an orally taken drug will pass from the intestine into the bloodstream, using only the molecule's chemical structure. It is one of the earliest screening gates in drug development.

A drug can only work if the body actually absorbs it, so poor intestinal absorption is a common reason oral candidates fail. Computational HIA models take a molecule's SMILES string — a text encoding of its structure — and output a probability that it will be absorbed. Screening this early, before any lab synthesis, filters out molecules likely to fail and saves time and cost. On Polaris, HIA is scored by AUROC, where 1.0 is perfect and 0.5 is random.

Which model ranked #1 on the Polaris HIA benchmark?

VIDRAFT's 'vidraft-hia-molformer-chemberta-ensemble' ranked first with an AUROC of 0.990, ahead of AceProp (0.989) and TabPFNv2 (0.989).

The result was recorded on Polaris's blind HIA leaderboard, where VIDRAFT narrowly beat two strong published baselines. Margins near the top are thin — hundredths of an AUROC point — so leading a field that includes AceProp and TabPFNv2 is meaningful. This marks VIDRAFT's 15th first-place finish across public Polaris leaderboards.

Why does a blind benchmark like Polaris matter?

Polaris is an independent benchmark where test labels are hidden and scoring runs server-side, so a team cannot tune to or inflate its own score. That makes a #1 result externally verifiable rather than self-reported.

Many published model claims are self-scored, which makes them hard to trust and easy to over-fit. Because Polaris keeps the answer key secret and computes scores on its own servers, every submission is graded the same way against unseen molecules. That is why VIDRAFT's 0.990 is described as a blind, external, competitive result — not a self-reported number.

How does VIDRAFT's HIA model work?

It is a structure-only heterogeneous ensemble that combines two chemistry language models, MoLFormer and ChemBERTa, using the molecule's SMILES as the only input. No cell imaging or extra data modalities are used.

MoLFormer and ChemBERTa are transformer models pre-trained on large molecule corpora, and ensembling their complementary views improves robustness. VIDRAFT reports that reinforcing the ChemBERTa component specifically helped on HIA. Honestly scoped, that gain did not transfer to every ADMET task and a single benchmark is not the whole picture — but the HIA win itself is a blind, competitive result, not a self-reported one.

Frequently asked questions

What AUROC did VIDRAFT score on Polaris HIA?
AUROC 0.990, the top score on the blind Polaris HIA leaderboard, ahead of AceProp and TabPFNv2 at 0.989.
Does the model use cell imaging or only the molecular structure?
Only the molecular structure — the SMILES string. It is a structure-only ensemble with no imaging or extra modalities.
Does leading HIA mean VIDRAFT is best at every ADMET task?
No. HIA is one blind benchmark; the ChemBERTa reinforcement helped HIA specifically and did not improve every ADMET task. The result is competitive and externally verified, but a single benchmark is not the whole picture.

Sources

Related

On-device AI
Can you run a 35B AI model with no GPU?
Model reasoning
What is the best Korean LLM on the GPQA Diamond science benchmark?
AI safety
Does AI know what it does not know? What is FINAL-Bench?
↖ Home — vidraft.net

This article is based on VIDRAFT public, measured data and external sources. Performance figures are measurements under the stated conditions and may vary by environment.