Model reference — AI Medical · Drug discovery

AI models for Molecular property prediction

Reported performance clusters at ROC-AUC 0.70–0.85 across the standard ADMET task collection, with the larger representation models at or near state of the art on around ten of those benchmarks.

The therapeutic reasoning models are a different proposition: one reports beating or matching specialist models on 64 of 66 tasks in a therapeutics benchmark suite, and it answers natural-language queries as well as returning numbers. It costs correspondingly more per molecule.

Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.

Molecular property prediction service AI Medical services Pricing

Input type — Molecular structures

Models in this group take SMILES, molecular graphs or 3D coordinates. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.

TxGemma-2B / TxGemma-9B / TxGemma-27B

Vendor: Google DeepMind

What it does: answers therapeutic questions in plain language as well as predicting properties, and beat or matched specialist models on 64 of 66 tasks in the standard therapeutics suite. Three sizes, and the 27B costs accordingly.

Beats or matches specialist models on 64 of 66 Therapeutics Data Commons tasks (developer-reported).

Requirement	Minimum	Medium	High
GPU type	A100 80 GB	A100 80 GB ×2	H100 80 GB ×4
VRAM	80 GB	160 GB	320 GB
vCPUs	24	48	96
RAM	128 GB	256 GB	512 GB
Server	1× A100 SXM 80 GB	2× A100 SXM 80 GB	4× H100 SXM 80 GB
Rate (molecules/hour)	≈ 900	≈ 3,200	≈ 8,100

MolFormer-XL / MoLFormer

Vendor: IBM Research

What it does: is at or near state of the art on ten property benchmarks and produces representations your own regressors sit on top of.

State of the art or near it on 10 MoleculeNet benchmarks (developer-reported).

Requirement	Minimum	Medium	High
GPU type	RTX 3090	L40S 48 GB	A100 80 GB
VRAM	24 GB	48 GB	80 GB
vCPUs	8	16	32
RAM	32 GB	64 GB	128 GB
Server	1× RTX 3090 24 GB	1× L40S 48 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 8,000	≈ 28,000	≈ 72,000

ChemBERTa-2 / ChemBERTa

Vendor: DeepChem and HuggingMolecules community

What it does: screens at very high throughput on modest hardware, at ROC-AUC 0.70–0.85 across ADMET tasks — usually the best cost per molecule for a first filter.

Competitive on MoleculeNet ADMET tasks, ROC-AUC 0.70–0.85 by task (independent).

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	L40S 48 GB
VRAM	12 GB	24 GB	48 GB
vCPUs	4	8	16
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× L40S 48 GB
Rate (molecules/hour)	≈ 60,000	≈ 210,000	≈ 540,000

Uni-Mol

Vendor: DP Technology

What it does: uses 3D coordinates rather than strings, which is what conformation-dependent properties and docking-adjacent work require.

Task-dependent across molecular property, conformation and docking benchmarks.

Requirement	Minimum	Medium	High
GPU type	RTX 3090	L40S 48 GB	A100 80 GB
VRAM	24 GB	48 GB	80 GB
vCPUs	8	16	32
RAM	32 GB	64 GB	128 GB
Server	1× RTX 3090 24 GB	1× L40S 48 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 8,000	≈ 28,000	≈ 72,000

MegaMolBART

Vendor: NVIDIA

What it does: generates molecules as well as embedding them, for library expansion rather than scoring alone.

Task-dependent; evaluated on molecular representation and generation tasks.

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	A100 80 GB
VRAM	12 GB	24 GB	80 GB
vCPUs	6	12	24
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 25,000	≈ 87,500	≈ 225,000

MoleculeSTM

Vendor: MoleculeSTM authors

What it does: links molecules to text, so a library can be searched — and molecules edited — by description.

Task-dependent across molecule-text retrieval and molecular editing benchmarks.

Requirement	Minimum	Medium	High
GPU type	RTX 3090	L40S 48 GB	A100 80 GB
VRAM	24 GB	48 GB	80 GB
vCPUs	8	16	32
RAM	32 GB	64 GB	128 GB
Server	1× RTX 3090 24 GB	1× L40S 48 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 8,000	≈ 28,000	≈ 72,000

GIT-Mol

Vendor: GIT-Mol authors

What it does: takes graphs, images and text together for captioning, question answering and property prediction from one model.

Task-dependent across molecular captioning, QA and property tasks.

Requirement	Minimum	Medium	High
GPU type	RTX 3090	L40S 48 GB	A100 80 GB
VRAM	24 GB	48 GB	80 GB
vCPUs	8	16	32
RAM	32 GB	64 GB	128 GB
Server	1× RTX 3090 24 GB	1× L40S 48 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 8,000	≈ 28,000	≈ 72,000

nach0

Vendor: nach0 authors

What it does: works across chemistry and natural language in one model, for tasks that mix the two.

Task-dependent across chemistry-language tasks.

Requirement	Minimum	Medium	High
GPU type	RTX 3090	L40S 48 GB	A100 80 GB
VRAM	24 GB	48 GB	80 GB
vCPUs	8	16	32
RAM	32 GB	64 GB	128 GB
Server	1× RTX 3090 24 GB	1× L40S 48 GB	1× A100 SXM 80 GB
Rate (molecules/hour)	≈ 8,000	≈ 28,000	≈ 72,000

BioBERT / PubMedBERT DDI-extraction fine-tunes

Vendor: Academic community (DDI-BioBERT lineage)

What it does: extracts drug-drug interaction statements from literature and labels at micro-F1 0.79–0.83 — cheap, well characterised pharmacovigilance text mining.

Micro-F1 0.79–0.83 on DDIExtraction-2013 (independent).

Requirement	Minimum	Medium	High
GPU type	RTX 3060	RTX 4090	L40S 48 GB
VRAM	12 GB	24 GB	48 GB
vCPUs	4	8	16
RAM	16 GB	32 GB	64 GB
Server	1× RTX 3060 12 GB	1× RTX 4090 24 GB	1× L40S 48 GB
Rate (molecules/hour)	≈ 60,000	≈ 210,000	≈ 540,000

Choosing between them

For high-throughput screening, a small SMILES encoder plus your own head is both fastest and usually most accurate on your assays. For exploratory work across many task types, the reasoning models earn their cost. 3D-aware models are worth it only where conformation matters to the property.

Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.

Molecular property prediction service AI Medical services Pricing

From benchmark to production

Send a representative sample, your expected volume and your latency target for molecular property prediction. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.